Essential works on AI and human capability

What should you read to understand what AI is doing to human capability? This is the map: hard data, experimental evidence, serious interpretation and the intellectual foundations most current writing rediscovers without attribution.

Question this page answers

Which books and papers on AI and human capability actually matter?

All 616 questions this research covers

This is a map of the field, not a reading list and not a defence. 135 works on what increasingly capable AI does to human capability, judgement, learning, expertise, work and agency, each classified by the role it plays, each with a one-line finding, an honest note on how strong the evidence actually is, and a short reading of why it matters. Some of it supports the SuperSkills argument. Some of it complicates the argument. At least one entry contradicts it directly, and is marked as doing so.

Inclusion does not mean endorsement. This list contains evidence and arguments that support, qualify and challenge the SuperSkills thesis. The purpose is to have read the field, not to assemble a hundred sources that agree with one person.

How this is classified

Every entry carries one of five classifications. They describe the role a work plays, which is a different question from how strong its evidence is, and both are stated separately because collapsing them is how bad citations happen.

Core (33). Sources anyone entering this field should know. Roughly twenty-five of them.
Strong evidence (44). Peer-reviewed work, large empirical datasets, or institutional analysis credible enough to build on.
Important perspective (21). A serious argument that shapes the debate. Not empirical proof, and not treated as such here.
Foundation (34). Older work on cognition, expertise, automation, judgement and skill acquisition. Most of this predates AI and explains it better than most writing about AI.
Watch (3). A living programme or dataset whose conclusions will change. Track it rather than cite it once.

Two axes, and what the second is for. Every work carries a class, which is how much weight it holds, and a source family, which is what kind of thing it is. They are independent on purpose. A government evidence assessment and a consultancy survey can both be strong and should be read with completely different reflexes, because one has no product to sell and the other has a sales motion attached to its conclusion. Filter by source above to see the field one family at a time.

How far the full reading goes. 12 of the 135 works carry the complete treatment: what they establish, what they do not establish, where they stand against the argument on this site, and which of the questions on the map they help answer. Those are the major reports, the academic research, enterprise evidence, institutional evidence, labour-market evidence, policy and regulation, ai-provider evidence sources executives cite at each other. Books and foundation papers are deliberately excluded, because asking what evidence an argument uses is the wrong question and answering it anyway produces filler. Of the works read in full, 4 support the argument here, 7 complicate it and 1 contradicts it directly.

Every external link has been fetched and confirmed. Where a statistic is quoted, its base is given: the BCG deskilling finding rests on seventy executives, and saying so is more useful than the headline. Where a work has been superseded, the earlier edition is kept as a longitudinal comparator rather than replaced, because what the field predicted and what happened is itself evidence.

Browse#

Filter by classification or by theme. Everything is on this page, so a search within it will find anything the filters do not.

Classification

Theme

Showing all 135 works

CoreWorking paper · 2026#

What Work Does Generative AI Do?#

Bick, A., Blandin, A., Deming, D. and Schumacher, T. · Federal Reserve Bank of St. Louis working paper, August 2026

Finding: A nationally representative survey linking generative AI use to detailed occupations and tasks. Adoption is broad but shallow: at least one in five workers use it in 80 per cent of occupations and 40 per cent of tasks, while within most of those tasks fewer than half of workers use it at all. Exposure measures correlate with adoption but explain only about half the variation between workers doing similar work. Some occupations use it mainly for high-expertise work and others for low-expertise work. The authors also find that platform chat-log data OVERCLASSIFIES basic, generic tasks relative to what survey respondents report.
Evidence strength: Nationally representative survey rather than one platform's traffic, with David Deming among the authors. A working paper rather than a peer-reviewed article, US only, and self-reported use.
Establishes: That adoption is broad but shallow, with at least one in five workers using generative AI across 80 per cent of occupations while within most tasks fewer than half do. That exposure explains only about half the variation in adoption between workers doing similar work. And that platform chat-log data overclassifies basic generic tasks against what survey respondents report.
Does not establish: Any effect on output, capability or employment. It measures who is using the tools for what, from self-report, at one point in time. A working paper, not peer reviewed, and US only.
Where it stands: Supports the argument Supports the distinction this estate insists on, that capability, exposure, adoption, usage and impact are five separate quantities routinely collapsed into one. Its chat-log caution is also the reason the two provider entries in this canon are read alongside it rather than on their own.
The SuperSkills read: The paper that lets this estate keep its central distinction straight: capability, exposure, adoption, usage and impact are five different quantities, and this measures the gaps between them directly. Its finding that chat logs overclassify generic tasks is a methodological caution that applies to the OpenAI and Anthropic entries above, and is the reason both are read here alongside it rather than on their own.
Questions it helps answer: How do you measure AI adoption properly?, How do expert AI users work differently from beginners?, Can capability loss from AI actually be measured?

jobsexpertiseskills questions

Questions answered on this page

What should I read to understand AI and human capability?

Start with the twenty-five core entries. They span hard data (Stanford HAI's AI Index, the WEF Future of Jobs series, PwC's AI Jobs Barometer), experimental evidence (Bastani on learning, Dell'Acqua on the jagged frontier, Vaccaro's meta-analysis on human-AI teams), serious interpretation (Mollick, Vallor, Beane, Carr, Agrawal on the economics of prediction) and intellectual foundations (Bainbridge 1983 on the ironies of automation, Polanyi 1966 on tacit knowledge, Dreyfus on novice-to-expert progression).

What is the most important overlooked source on AI and deskilling?

Lisanne Bainbridge's 1983 paper Ironies of Automation, in Automatica. She wrote that the more advanced a control system is, the more crucial the human operator's contribution may become, and that by taking away the easy parts of a task, automation makes the difficult parts harder. That is the modern deskilling argument, made forty-three years ago about process control, and most current writing on AI and skill restates it without knowing it exists.

Does the essential works only include sources that agree with Rahim Hirji?

No, and it says so prominently: inclusion does not mean endorsement. This list includes Macnamara and Maitra's re-examination, which complicates the deliberate-practice argument this research relies on, and Acemoglu and Johnson's Power and Progress, which challenges the assumption that protecting capability is sufficient for benefits to follow. The purpose is to have read the field rather than to assemble sources that agree.

How is this list different from the evidence base?

The evidence base answers what the research shows: empirical work only, graded by method, with what each study does not support. This list answers what to read to understand the field: it includes books, arguments, foundational theory and living datasets that are not evidence at all, classified by the role they play. They cross-link, and shared entries link to their graded version.

CoreProvider telemetry analysis · 2026#

Labor market impacts of AI: A new measure and early evidence

Massenkoff, M. and McCrory, P., Anthropic · Anthropic, 5 March 2026

Finding: Introduces OBSERVED exposure, combining Eloundou et al.'s theoretical task capability with actual Claude usage, weighting automated over augmentative and work-related over personal use. Actual coverage is far below theoretical: Claude covers 33 per cent of Computer and Math tasks where theory allows 94. Computer programmers are most exposed at 75 per cent coverage; 30 per cent of workers have zero. Every 10 percentage points of coverage corresponds to 0.6 points lower BLS projected growth to 2034, a correlation absent from the theoretical measure alone. No systematic rise in unemployment for exposed workers since late 2022. Hiring of 22 to 25 year olds into exposed occupations is down about 14 per cent, which the authors call barely significant.
Evidence strength: COMMERCIAL INTEREST: published by a model provider, using its own product's traffic, and a finding of no unemployment effect is a comfortable one for that publisher. Against that: the paper opens by cataloguing how badly previous exposure work has predicted, states its young-worker result is barely significant, publishes its appendix and issued a public correction on 8 March 2026 when a figure's labels were reversed. US only, one provider's traffic, and the Eloundou metric is anchored to early-2023 capability.
Establishes: That theoretical exposure and observed exposure are different quantities and can be measured apart. Actual coverage sits far below what capability allows: 33 per cent of Computer and Math tasks against a theoretical 94. And that observed exposure tracks independently produced BLS growth projections where the theoretical measure alone does not.
Does not establish: That AI is not displacing anyone. It finds no systematic unemployment effect, which is a different claim, and the authors say so. Unemployment also misses workers who leave the labour force or never enter it, which is precisely the population the young-worker result concerns. One provider's traffic, US only, and the 14 per cent hiring result is barely significant on the authors' own account.
Where it stands: Complicates the argument The slowdown in hiring 22 to 25 year olds into exposed occupations is this estate's missing rungs argument arriving in national survey data, which supports it. What complicates the argument is the rest of the paper: no unemployment effect, and a careful opening section on how badly this kind of forecasting has performed before. It is a warning against confident capability claims as much as evidence for one.
The SuperSkills read: The separation of theoretical from observed exposure is the single most useful methodological move in this literature, because it makes the gap between what AI could do and what it is doing into a measurable quantity rather than a rhetorical one. The young-worker hiring result is the missing rungs argument appearing in national survey data.
Questions it helps answer: Will AI replace entry-level jobs?, How much of the workforce is actually exposed to AI?, Will AI replace my job?, Which jobs are safest from AI?, Who ends up worse off as AI spreads?

jobsearly careersskills questions

CoreInstitutional survey · 2026#

Agents, human agency, and the opportunity for every organization

Microsoft · 2026 Work Trend Index Annual Report, 5 May 2026

Finding: Trillions of Microsoft 365 signals plus 20,000 AI-using workers across ten markets, organised around human agency, judgement and the delegation of work to agents.
Evidence strength: COMMERCIAL INTEREST: published by a company that sells the thing being measured. Institutional survey plus product telemetry. Large and current; the vendor sells the tools whose adoption it measures, and telemetry counts usage rather than value.
Establishes: That a specific organisational model, human-agent teams with people directing agents, is being actively promoted to executives and adopted by some, and what its proponents believe it looks like.
Does not establish: That the model works, or that the organisations adopting it retain the capability to operate without it. It is a vision document with survey data attached, published by a company selling the agents the vision requires.
Where it stands: Contradicts the argument The clearest direct opposition in the canon, and included for that reason. It argues that the constraint on organisations is capacity, solvable by adding digital labour. This estate argues the constraint is judgement, which cannot be added and can be lost. If Microsoft is right, the capability question is a transitional worry. Readers should see the strongest version of that case.
The SuperSkills read: The closest major corporate research programme to this thesis, and worth watching for that reason. When the company with the most usage data starts organising its annual report around human agency and judgement allocation, the question has moved from the margins to the centre. Read it as a signal about the field as much as evidence about the world.
Questions it helps answer: What does an AI-native operating model look like?, What happens when an employee manages more agents than people?

agencyorganisation designjudgement ai agents and human judgement · how should leaders respond to ai

CorePolicy brief · 2026#

AI and skills: What we know so far

OECD · OECD policy brief, 5 June 2026

Finding: Fewer than one percent of workers will need advanced AI-specific skills such as programming or model development, while most require digital, analytical and interpretive capability.
Evidence strength: Institutional analysis. The headline figure derives from Green and Lamby (2023), built largely on job-posting data, so it describes demand as advertised.
Establishes: A careful institutional account of what is currently known about AI and skill requirements across countries, and an explicit statement of the boundaries of that knowledge.
Does not establish: What happens to an existing capability when AI performs the activity through which that capability was previously maintained. The paper is about skill REQUIREMENTS and skill SUPPLY, and that question sits outside its frame rather than being answered badly by it.
Where it stands: Complicates the argument The most important comparator this estate has, and the cleanest demonstration of where the estate's question actually lives. The OECD asks what skills will be needed. This estate asks what happens to the skills people already have. Neither answer substitutes for the other.
The SuperSkills read: The most quotable single finding for anyone being told to retrain as an AI specialist. It is the policy-grade version of the argument that tool fluency is table stakes rather than a career, and it comes from a body with no product to sell.
Questions it helps answer: How does this research compare with the OECD's work on skills?, Which skills become more valuable as AI improves?

skillsjobs why learn to prompt is weak career advice · human skills in the age of ai

CoreLabour-market analysis · 2026#

2026 Global AI Jobs Barometer: Two futures for jobs in an AI era

PwC · PwC, 15 June 2026

Finding: Analysing 2.4 million US entry-level jobs, entry-level roles most exposed to AI are seven times more likely to require traditionally senior human-intensive skills such as leadership and creativity; those roles grew 35 percent since 2019 while other entry-level roles shrank 10 percent.
Evidence strength: Large-scale job-advertisement analysis across more than a billion adverts. Observed demand rather than stated preference, though adverts describe what employers ask for rather than what work requires. The entry-level finding is US-only.
The SuperSkills read: The strongest external evidence for synthetic seniority anywhere. Entry-level roles are not disappearing so much as being seniorised, which means the junior work through which people used to become senior is being removed while the expectation of senior judgement arrives on day one. That is the missing rungs problem, measured.
Questions it helps answer: Which skills become more valuable as AI improves?

early careersjobsskills will ai replace entry level jobs · synthetic seniority · missing rungs

CoreInstitutional report · 2026#

The 2026 AI Index Report

Stanford HAI · Stanford Institute for Human-Centered AI, 9th edition

Finding: The most comprehensive annual account of AI capability, investment, adoption and economic effect, now with a labour chapter that engages directly with learning effects and early-career workers.
Evidence strength: Compiled review. Aggregates many third-party sources; strong on breadth, not original research.
Establishes: The broadest verified statistical base in the field, across capability, economy, education, medicine, governance and public attitudes. On education, that student AI use has run well ahead of institutional policy.
Does not establish: Causation anywhere. It is a compiled review that aggregates third-party sources, so its numbers inherit the methods and the weaknesses of whoever produced them, and a figure being in the Index is not independent verification of that figure.
Where it stands: Supports the argument Supports on the education gap specifically, and is the place to check a number before repeating it, which is a service this estate relies on.
The SuperSkills read: The master factual reference for what AI can actually do, and the right place to check a capability claim before repeating it. Its limitation is also its usefulness: it measures the machine side of the equation, the half this research does not cover. Use its public data rather than only its PDF.
Questions it helps answer: How do you assess students when AI can do the assignment?, What do universities owe students on this?, Can capability loss from AI actually be measured?

technology capabilityjobs ai workforce strategy

CoreGovernment evidence assessment · 2026#

Assessment of AI capabilities and the impact on the UK labour market

UK Department for Science, Innovation and Technology · DSIT, 28 January 2026

Finding: Splits the UK workforce into 35 per cent high exposure with high complementarity, 32 per cent high exposure with low complementarity and 33 per cent low exposure, drawing on IMF estimates that around 70 per cent of UK workers are in occupations containing tasks AI could perform or enhance. States that ex ante measures overstate the adoption actually observed.
Evidence strength: Government synthesis with no product to sell, and unusually explicit about the gap between predicted and revealed exposure. It is a synthesis of other people's estimates rather than new measurement.
Establishes: A defensible UK exposure split, 35 per cent high exposure with high complementarity, 32 per cent high exposure with low complementarity, 33 per cent low exposure. And, in its own text, that ex ante exposure measures overstate the adoption actually observed.
Does not establish: Anything about capability. Exposure describes the task composition of a job, not what happens to the person doing it, and the assessment does not claim otherwise. Nor does high exposure imply displacement: the complementarity split is the whole point and is usually dropped when the figure is quoted.
Where it stands: Complicates the argument Complicates by being properly uncertain where this estate is willing to argue. It declines to draw the capability conclusion the estate draws, and is right to, because its evidence does not reach that far.
The SuperSkills read: The UK anchor this estate lacked, and the rarest quality in this literature: a document that separates what is known from what remains uncertain and says so in its own text.
Questions it helps answer: How much of the workforce is actually exposed to AI?, Will AI replace my job?, How should I analyse my own job for AI exposure?

jobsskills questions

CoreJournal article · 2025#

Expertise

Autor, D. and Thompson, N. · NBER Working Paper 33941; Journal of the European Economic Association, 23(4), 1203-1271

Finding: Automation that removed the LESS expert tasks raised wages and reduced employment. Automation that removed the EXPERT tasks lowered wages and increased employment.
Evidence strength: Peer-reviewed study. Anything measured about generative AI. The data ends in 2018, so this is a lens, not a forecast.
The SuperSkills read: Gives the career question a testable form. Not whether AI will take your tasks, but whether removing them raises or lowers the expertise of what remains, which determines whether your role appreciates or commoditises.

jobsexpertiseskills staying valuable in the age of ai · will ai replace my job · which tasks do workers not want automated · what becomes more valuable as ai gets cheaper graded entry

CoreJournal article · 2025#

Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics

Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakci, O. and Mariman, R. · Proceedings of the National Academy of Sciences, 122(26)

Finding: Grades rose 48 percent with unrestricted access and 127 percent with the tutor while the tool was present. With access removed, the unrestricted group scored 17 percent LOWER than students who never had it. The guardrailed tutor largely removed the harm.
Evidence strength: Peer-reviewed study. What a guardrailed interface should look like for professional work. It was school mathematics over a bounded period.
The SuperSkills read: The cleanest demonstration that the design of the tool, not the presence of AI, decides whether a person is taught or carried. The same model produced both the best and the worst learning outcome in one experiment.

learningeducationdeskilling how humans learn with ai · chro guide to ai · ai and human judgement graded entry

CoreJournal article · 2025#

The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects from a Survey of Knowledge Workers

Lee, H.-P. et al. · Microsoft Research and Carnegie Mellon, CHI 2025

Finding: Higher confidence in the tool was associated with less critical thinking, and the thinking that remains shifts from producing to verifying, from solving to integrating.
Evidence strength: Peer-reviewed study. Causation. People who think differently may use AI differently, and a survey cannot separate the two.
The SuperSkills read: The first large study of what AI does to professional thinking in real work. Its finding that thinking migrates to verification rather than disappearing is more precise, and more useful, than the deskilling headlines it generated.

critical thinkingjudgement ai and critical thinking · ai and human judgement · is screen time the same argument as ai use graded entry

CoreInstitutional survey · 2025#

The Future of Jobs Report 2025

World Economic Forum · World Economic Forum, Geneva, January 2025

Finding: Analytical thinking is the most valued core skill, and skills gaps are named the single biggest barrier to business transformation.
Evidence strength: Institutional survey. What employers actually do. Stated skill preference and hiring behaviour diverge routinely.
The SuperSkills read: The dataset everyone quotes. Useful for what employers say they want, which is a real signal, and not evidence of what they do.

skillsjobs human skills in the age of ai · ai workforce strategy · ai and human judgement graded entry

CoreWorking paper · 2024#

The Simple Macroeconomics of AI

Acemoglu, D. · NBER Working Paper 32487; published in Economic Policy, 40(121), 2025, 13-58

Finding: Estimates that AI's effect on total factor productivity over ten years will be modest, roughly an order of magnitude smaller than the most quoted forecasts.
Evidence strength: Peer-reviewed in its published form. Assumption-driven macro estimate; Acemoglu is explicit that the numbers depend on which tasks prove genuinely automatable.
The SuperSkills read: The necessary corrective to every consultancy trillion-dollar figure, including several elsewhere in this Canon. It also complicates the urgency this research assumes: if the economic effect is small and slow, the capability effect may be too. Worth reading immediately after any report promising transformation.
Questions it helps answer: What return should we expect from our AI investment?, Who captures the productivity gains from AI?

jobsorganisation design ai workforce strategy · how should leaders respond to ai

CoreWorking paper · 2024#

Applying AI to Rebuild Middle Class Jobs

Autor, D. · NBER Working Paper 32140, February 2024

Finding: Argues that AI, unlike previous automation, can extend expert decision-making to a wider set of workers, potentially rebuilding middle-skill work rather than hollowing it out further.
Evidence strength: Working paper, not peer-reviewed. An argument built on economic reasoning and historical analogy rather than measurement of generative AI itself.
The SuperSkills read: The strongest rigorous optimistic case in the field, and a genuine challenge to this research rather than a soft counterpoint. If Autor is right, AI restores the middle by putting expert judgement within reach of more people. If the capability argument on this site is right, that only holds where the expertise is genuinely acquired rather than borrowed from the machine. Both cannot be fully true, and that tension is worth sitting with.
Questions it helps answer: Which jobs are safest from AI?, Will AI make experts more or less valuable?

jobsexpertiseskills staying valuable in the age of ai · will ai replace my job

CoreBook · 2024#

The Skill Code: How to Save Human Ability in an Age of Intelligent Machines

Beane, M. · Harper Business

Finding: Draws on field research into how novices actually learn from experts, and shows that intelligent machines are removing the apprenticeship through which skill has always been transmitted.
Evidence strength: Important perspective, built on the author's own ethnographic field research including robotic surgery.
The SuperSkills read: The closest book to the missing-rungs argument, and essential reading for anyone interested in it. Beane arrived at the same problem from a different direction, through direct observation of surgical residents losing hands-on time to robotic systems, the same mechanism this research describes in knowledge work.

expertiselearningearly careers missing rungs · the missed reps of ai seniority · how humans learn with ai

CoreJournal article · 2024#

GPTs are GPTs: Labor market impact potential of LLMs

Eloundou, T., Manning, S., Mishkin, P. and Rock, D. · Science, 384(6702), 1306-1308

Finding: Around 80 percent of US workers could have at least 10 percent of tasks affected; about 19 percent could see at least half affected.
Evidence strength: Peer-reviewed study. That any job will be lost. This is exposure, not displacement, and the authors say so explicitly. It is the most misquoted number in the field.
The SuperSkills read: The most misquoted number in the field. It measures task exposure, not displacement, and the authors say so explicitly. Included partly so that the distinction is easy to point at.

jobsskills will ai replace my job graded entry

CoreBook · 2024#

Co-Intelligence: Living and Working with AI

Mollick, E. · Portfolio

Finding: Sets out practical principles for working with AI, including always inviting it to the table and treating it as a person while remembering it is not one.
Evidence strength: Important perspective, grounded in the author's own experiments rather than controlled research.
The SuperSkills read: The best mainstream book on practical human-AI collaboration, and the most useful single recommendation for someone starting out. Where this research differs is emphasis: Mollick is strongest on how to get value from the tool today, and lighter on what habitual use does to capability over years.

human-AI collaborationskills using ai without dependency

CoreJournal article · 2024#

When combinations of humans and AI are useful: a systematic review and meta-analysis

Vaccaro, M., Almaatouq, A. and Malone, T. · Nature Human Behaviour, 8, 2293-2303

Finding: Human-AI combinations performed significantly WORSE on average than the better of human or AI alone (Hedges' g = -0.23). Losses concentrated in decision-making; gains in content creation. Pairing gained where humans beat the AI and lost where the AI beat humans.
Evidence strength: Peer-reviewed study. That human-AI teams are useless. The benchmark is an oracle-selected best performer, which you rarely know in advance. Also predates current frontier models.
The SuperSkills read: The most under-absorbed finding in the field. It is the empirical case against treating a human in the loop as a control, and it should change how oversight is designed rather than merely how it is described.

human-AI collaborationjudgement human ai decision making · how should leaders respond to ai · ai and human judgement graded entry

CoreBook · 2024#

The AI Mirror: How to Reclaim Our Humanity in an Age of Machine Thinking

Vallor, S. · Oxford University Press

Finding: Argues that AI systems are mirrors reflecting our past rather than minds, and that the danger is our shrinking conception of what humans are.
Evidence strength: Important perspective. Philosophy, not evidence.
The SuperSkills read: The closest philosophical work to the question underneath this research. Vallor's concern is not that machines will surpass us but that we will define ourselves downwards to match them, which is the same worry as capability debt stated at the level of the species rather than the organisation.

ethicsagency what stays human

CoreJournal article · 2024#

AI can help people feel heard, but an AI label diminishes this impact

Yin, Y., Jia, N. and Wakslak, C. J. · PNAS, 121(14), e2319112121

Finding: AI-generated replies made recipients feel MORE heard than replies from untrained humans, and labelling the reply as AI removed the advantage.
Evidence strength: Peer-reviewed study. That the label effect is stable. Norms around disclosed AI assistance are moving, and nobody has measured this over time.
The SuperSkills read: The most clarifying study in the debate about what stays human. The words were not the thing being valued; the identity of whoever wrote them was.

agencyethics outsourced recognition · what stays human graded entry

CoreJournal article · 2023#

Generative AI at Work

Brynjolfsson, E., Li, D. and Raymond, L. · Quarterly Journal of Economics, 140(2), 889-942. DOI 10.1093/qje/qjae044. Advance Access 4 February 2025. Earlier version NBER Working Paper 31161

Finding: Resolutions per hour rose 15 per cent on average, 15.2 per cent in the preferred specification with agent and tenure fixed effects. Less skilled and less experienced workers gained a 30 per cent increase in issues resolved per hour, rising to 36 per cent for the lowest skill quintile, while the most skilled saw no significant productivity change and small declines in conversation quality and customer satisfaction. Customer sentiment improved by half a standard deviation and requests to speak to a manager fell about 25 per cent. During unplanned outages, agents with longer AI exposure still handled chats faster than their pre-AI baseline, but only those who had adhered closely to the suggestions.
Evidence strength: Peer-reviewed study. Whether those novices became experts. It measures output over months in one firm, one occupation and a stable product environment, and the authors say so. Wages, labour demand and hiring composition were not observed. The outage estimates are the authors' own noisiest, because outages are rare and may not be comparable chats.
The SuperSkills read: Establishes that AI raises the floor far more than the ceiling. Read alongside the missing-rungs argument, it poses the question the study itself does not answer: whether the novices it lifted ever became experts.

jobsexpertiseearly careers how humans learn with ai · staying valuable in the age of ai · ai and human judgement · how will ai change customer service · what becomes more valuable as ai gets cheaper graded entry

CoreWorking paper · 2023#

Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality

Dell'Acqua, F. et al. · Harvard Business School and BCG working paper

Finding: Inside the frontier, AI-assisted consultants were dramatically better and faster. Outside it, they performed worse than consultants with no AI at all.
Evidence strength: Working paper, not peer-reviewed. Where the frontier runs in your domain. That is local and must be learned.
The SuperSkills read: The jagged frontier is the single most useful mental model for anyone using AI at work. It also shows the cost of not knowing where the frontier runs, which is the capability that does not commoditise.

human-AI collaborationjudgementexpertise human ai decision making · why learn to prompt is weak career advice · ai and human judgement graded entry

CoreCompetency framework · 2019#

OECD Learning Compass 2030

OECD · OECD Future of Education and Skills 2030 project

Finding: A framework placing student agency, and the capacity to set goals, reflect and act responsibly, at the centre of education, alongside knowledge, skills, attitudes and values.
Evidence strength: Official framework developed through multi-country expert consultation. Evolving rather than a single dated report; normative, not empirical.
The SuperSkills read: The most important framework on this list for anyone arguing about human capability, because it put agency at the centre years before AI made the argument urgent. Its anticipation-action-reflection cycle is a considerably more rigorous account of learning to learn than most current writing on the subject.

frameworkseducationagencylearning how humans learn with ai · human skills in the age of ai

CoreBook · 2018#

Prediction Machines: The Simple Economics of Artificial Intelligence

Agrawal, A., Gans, J. and Goldfarb, A. · Harvard Business Review Press, updated edition 2022

Finding: AI reduces the cost of prediction. When prediction becomes cheap, the value of its complements rises, and the most important complement is human judgement.
Evidence strength: Important perspective, built on economic reasoning rather than empirical measurement.
The SuperSkills read: The cleanest economic articulation of why judgement becomes more valuable rather than less. It is the argument this research makes, arrived at from price theory rather than from capability. It is the version to give a sceptical executive.

judgementorganisation design ai and human judgement · staying valuable in the age of ai

CoreResearch synthesis · 2018#

How People Learn II: Learners, Contexts, and Cultures

National Academies of Sciences, Engineering, and Medicine · National Academies Press

Finding: A consensus synthesis of learning science, covering how context, culture, motivation and prior knowledge shape what is actually learned.
Evidence strength: Consensus report by an expert committee, the strongest form of evidence synthesis available in education research.
The SuperSkills read: The reference to reach for before accepting any claim about AI and learning. Its central lesson, that learning is situated and depends on far more than information transfer, is why an AI tutor that answers well is not the same as one that teaches.
Questions it helps answer: How do adults learn with AI?, What is desirable difficulty?

learningeducation how humans learn with ai

CoreBook · 2014#

The Glass Cage: Automation and Us

Carr, N. · W. W. Norton

Finding: Argues that automation degrades the skills and attention of the people it assists, drawing on aviation, medicine and architecture.
Evidence strength: Important perspective. Journalistic synthesis of research rather than original evidence.
The SuperSkills read: The historical warning, written before generative AI and largely vindicated by it. Carr's value is that he made the argument when it was contrarian and about autopilots rather than chatbots, and the current wave of AI-and-thinking writing so often reads as a rediscovery of it.

deskillingjudgement ai and human judgement · capability debt

CoreBook · 2001#

Managing the Unexpected

Weick, K. E. and Sutcliffe, K. M. · First edition Jossey-Bass, San Francisco, 2001. Third edition, Sustained Performance in a Complex World, Wiley, 2015

Finding: High-reliability organisations sustain performance under uncertainty through collective mindfulness: preoccupation with failure, reluctance to simplify, sensitivity to operations, commitment to resilience, and deference to expertise rather than to rank.
Evidence strength: Foundational organisational theory grounded in studies of aircraft carriers, nuclear plants and wildland firefighting.
The SuperSkills read: The best answer available to what an organisation must remain able to do when the system fails, and to how much slack resilience requires. Deference to expertise is also the direct organisational counter to supervision that has become approval.

organisation designgovernancejudgement who supervises work they cannot do · what professions can learn from aviation

CoreBook · 1991#

Situated Learning: Legitimate Peripheral Participation

Lave, J. and Wenger, E. · Cambridge University Press, Cambridge

Finding: Learning is a social process of moving from the periphery of a community of practice towards full participation. Newcomers acquire competence by doing real but peripheral work alongside practitioners, not by instruction beforehand.
Evidence strength: Ethnographic and theoretical, built from apprenticeship studies including tailors and midwives. Not a controlled test.
The SuperSkills read: The mechanism the junior-work argument depends on, stated forty years before it was needed. If competence is acquired by doing legitimate peripheral work, then removing that work removes the route to competence rather than merely the task.

learningexpertiseearly careers

CoreBook · 1988#

In the Age of the Smart Machine: The Future of Work and Power

Zuboff, S. · Basic Books, New York

Finding: From five years of fieldwork in computerising workplaces, Zuboff distinguishes automating, where technology replaces human judgement, from informating, where it generates information that deepens the worker's understanding. The same system can do either, and which one happens is a management choice rather than a property of the technology.
Evidence strength: Long-form ethnographic fieldwork across multiple sites. Foundational and pre-digital in its examples.
The SuperSkills read: The canonical source for the distinction this estate calls drift versus design, made in 1988 and better specified. Zuboff establishes that the outcome is chosen rather than inherited, which is the claim the whole research programme rests on, and she watched organisations choose wrongly for a decade before anyone said AI.

organisation designskillstechnology capability design versus drift · ai workforce strategy

CoreBook · 1984#

Normal Accidents: Living with High-Risk Technologies

Perrow, C. · Basic Books, New York. Updated edition Princeton University Press, 1999

Finding: In systems that are both interactively complex and tightly coupled, serious accidents are a structural property rather than a failure of care. Adding warnings and safeguards increases complexity and can make the system less safe.
Evidence strength: Sociological analysis of accident cases. The framework is contested by high-reliability theorists, Weick among them, who argue such systems can be managed safely.
The SuperSkills read: The argument against solving an automation risk by adding another automated check. It also supplies the vocabulary for asking how tightly coupled an organisation has become to a system it does not control.

organisation designgovernancehuman-AI collaboration

CoreJournal article · 1983#

Ironies of Automation

Bainbridge, L. · Automatica, 19(6), 775-779

Finding: In the author's words, the more advanced a control system is, the more crucial the contribution of the human operator may become; and by taking away the easy parts of the task, automation can make the difficult parts harder.
Evidence strength: Foundational theoretical paper, extensively validated in aviation and process control over four decades.
The SuperSkills read: The single most important source on this list, and the least cited in current debate. Bainbridge described the missing-rungs and missed-reps problem in 1983, in process control: automation removes the routine work through which operators stayed practised, then depends on those same operators for the hard cases. Everything written about AI and deskilling since is a restatement of this paper, usually without knowing it.
Questions it helps answer: What is automation complacency?, Who supervises work they cannot do themselves?, How do you redesign a job around AI?

deskillingexpertisejudgementorganisation design capability debt · the missed reps of ai seniority · missing rungs

CoreBook · 1976#

Computer Power and Human Reason: From Judgment to Calculation

Weizenbaum, J. · W. H. Freeman and Company, San Francisco

Finding: Weizenbaum, who built ELIZA and watched people confide in it, argues that the question of what computers can do is separate from what they ought to be asked to do, and that judgements requiring compassion, wisdom or interpersonal responsibility should not be delegated however capable the system becomes.
Evidence strength: Argument by the person best placed to make it. No empirical component.
The SuperSkills read: The oldest and still the sharpest statement of the position this research keeps arriving at. Written fifty years ago by an AI researcher, it separates capability from appropriateness in a way most current writing does not, and its title names the exact substitution at issue: judgement replaced by calculation.

judgementethicsagency what stays human · what is judgement

CoreBook · 1974#

Labor and Monopoly Capital: The Degradation of Work in the Twentieth Century

Braverman, H. · Monthly Review Press, New York

Finding: Scientific management progressively deskills labour by separating conception from execution: the planning moves to a specialist layer and the worker is left with the residue. Deskilling is presented as a structural tendency of the way work is organised rather than an accident of technology.
Evidence strength: Historical and theoretical. Heavily debated for fifty years, including by labour-process scholars who reject its determinism.
The SuperSkills read: The origin of deskilling as a concept, and the argument this research has been citing in prose without listing. Braverman also supplies the challenge worth answering: if deskilling is structural rather than technological, then AI is the current instrument of an older process, and the remedy is organisational rather than technical.

deskillingskillsorganisation design what is deskilling · missing rungs

CoreBook · 1966#

The Tacit Dimension

Polanyi, M. · University of Chicago Press, current edition 2009 with a foreword by Amartya Sen

Finding: We can know more than we can tell. A large part of expert knowledge cannot be made explicit, and is acquired through practice rather than instruction.
Evidence strength: Foundational philosophy, not empirical work.
The SuperSkills read: The reason expertise cannot be written into a prompt. If the most valuable part of what an expert knows cannot be articulated, then a system trained on articulated knowledge is missing the part that matters, and a novice who never practises never acquires it.

expertiselearning what stays human · how humans learn with ai

Strong evidenceInstitutional analysis · 2026#

Happy Labor Day? How geopolitics, immigration and AI will reshape work

Allianz Research · Allianz Trade, 30 April 2026

Finding: Models the combined effect of AI, demographics and migration on labour supply and task composition.
Evidence strength: Institutional modelling. Measured effects. Like all exposure modelling, it is an estimate of what could be affected.
The SuperSkills read: A cross-country modelling view from outside the consultancy sector, which is worth having for that reason alone.

jobs will ai replace my job graded entry

Strong evidenceExecutive survey · 2026#

When Everyone Uses AI, Companies Risk Losing Critical Skills

BCG Henderson Institute · Boston Consulting Group, 17 June 2026

Finding: Half of the executives surveyed report already observing deskilling in their organisations, and more than sixty percent expect it to become a material threat within three to five years.
Evidence strength: Executive survey, and the base is small: seventy C-suite and senior leaders. Directionally interesting, statistically light.
The SuperSkills read: The most direct external corroboration of capability debt available, and it must be cited honestly. Seventy executives is a signal that senior people are noticing something, not a measurement of how widespread it is. Anyone quoting the fifty percent without the base of seventy is overstating it.
Questions it helps answer: Which professions face the greatest deskilling risk?, What is a capability audit?

deskillingjudgementleadership capability debt · ai workforce strategy

Strong evidenceInstitutional report · 2026#

Changing landscape of skills in the age of AI

ILO, ETF, Cedefop, Eurofound, European Commission and UNESCO · Joint publication, 13 August 2026

Finding: A six-agency synthesis on how AI is reshaping skill demand, with higher-order cognitive skills, socioemotional skills, adaptability and human agency as organising themes.
Evidence strength: Institutional synthesis. Six agencies, so unusually broad, but a synthesis rather than new primary data.
Establishes: Task-level occupational exposure across global employment, including low and middle income countries, with differences by gender and income group made explicit rather than averaged away.
Does not establish: Outcomes. An exposure index is a statement about what tasks a job contains against what models can do, and contains no claim about what employers will choose.
Where it stands: Supports the argument Supports the estate's insistence on the task rather than the job as the unit of analysis, and corrects its heavy Anglo-American centre of gravity.
The SuperSkills read: Notable less for new evidence than for the language. When six intergovernmental bodies converge on human agency and higher-order cognition as the organising frame, the territory this research occupies has become the mainstream policy question. Cite it as ILO and others, not ILO alone.
Questions it helps answer: What is happening outside the US and UK?, Which jobs are safest from AI?

skillsjobsagency human skills in the age of ai

Strong evidenceSurvey · 2026#

What 81,000 people told us about the economics of AI

Massenkoff, M. and Huang, S. (Anthropic) · Anthropic, 22 April 2026

Finding: A survey of 80,508 respondents on productivity, displacement anxiety and the economics of AI use.
Evidence strength: COMMERCIAL INTEREST: a model provider surveying its own users. Very large sample, self-selected towards AI users. Strong on scale, weak on representativeness.
The SuperSkills read: Useful mainly for early-career anxiety and perceived displacement, which are real economic forces regardless of whether the displacement itself materialises. Read alongside the Danish administrative data, which finds no measured earnings effect at all.

jobsearly careers will ai replace my job

Strong evidenceInstitutional analysis · 2026#

Agents, robots, and us: How AI reshapes work and skills in Europe

McKinsey Global Institute · McKinsey Global Institute, May 2026

Finding: Models how agentic AI and robotics together reshape task composition and skill demand across Europe.
Evidence strength: Institutional modelling. Outcomes. Task exposure modelling has consistently over-predicted the pace of realised change.
The SuperSkills read: The most current institutional modelling of the agentic shift. Task-exposure modelling has consistently over-predicted pace.

jobsorganisation design ai agents and human judgement · ai workforce strategy graded entry

Strong evidencePolicy analysis · 2026#

Skills in the AI age

OECD · OECD Artificial Intelligence Papers No. 60, July 2026

Finding: A policy synthesis on how AI changes skill requirements across economies, rather than a forecast of job losses.
Evidence strength: Institutional analysis, methodologically documented, not peer-reviewed.
Establishes: Deeper institutional treatment of skills demand, training provision, shortages and the rising weight of social and emotional skills.
Does not establish: Maintenance. It describes what should be taught and to whom, at population scale, and does not address whether a formed skill survives the removal of the practice that formed it.
Where it stands: Complicates the argument Same boundary as the June paper, drawn deeper. Its social and emotional skills finding runs alongside Stanford's shift towards interpersonal competencies, which is two institutions arriving at a similar place by different routes.
The SuperSkills read: The serious policy treatment of skills, as distinct from the consultancy treatment. Useful precisely because it resists the headline framing that dominates this subject.
Questions it helps answer: How does this research compare with the OECD's work on skills?, Can you regain a skill you have lost?

skillseducation human skills in the age of ai

Strong evidenceProvider telemetry analysis · 2026#

Work at the Frontier: How AI is Expanding What People Do at Work

OpenAI Economic Research · OpenAI, 27 July 2026

Finding: Analysis of more than 800,000 messages from US ChatGPT users. 16.8 per cent of work-related messages, and 43.5 per cent of occupation-specific messages, concern tasks associated with a different occupation. Names the pattern task crossover. Customer experience workers 77 per cent, designers 75, HR 69, legal 56, marketers 53. Design imports heavily (35.2 per cent of designers' messages are outside work) while exporting almost nothing (1.7 per cent); engineering is the reverse. Crossover is higher in small workspaces, 18.9 per cent at 2 to 5 seats against 16.3 per cent above 100.
Evidence strength: Large-sample behavioural data of a kind nobody outside a model provider can obtain, and the first real measurement of task crossover. COMMERCIAL INTEREST: OpenAI benefits from the finding that its product expands what workers can do. It is usage data, so it shows what people ASKED for, not whether the answer was any good or whether the user could tell.
Establishes: That work is crossing occupational boundaries in AI use at scale and at measurable rates, and that the direction differs by occupation: design imports tasks and barely exports any, engineering does the reverse.
Does not establish: That any of this crossover work is being done well. It is a record of what people ASKED a model to do, with no measure of output quality and no way to tell whether the user could evaluate the answer. It also cannot establish that roles are durably changing, because a request is not a responsibility.
Where it stands: Complicates the argument It supports the claim that job boundaries are dissolving faster than job descriptions, which this estate argues. It complicates the estate's framing by showing the movement as expansion rather than loss: people are gaining reach at the same time as they lose the ability to check the work they have gained reach over. Both are happening and the estate has mostly written about one.
The SuperSkills read: The most important new source for this estate's argument, and not for the reason OpenAI publishes it. If a marketer is troubleshooting a website and a designer is doing finance, the crossover is into work they were never trained to evaluate. That is the verification problem arriving as measured behaviour rather than as an argument, from the provider's own data.
Questions it helps answer: Does AI blur the boundaries between professions?, How do you redesign a job around AI?, How should small organisations approach AI?, Will AI make experts more or less valuable?

jobsexpertiseorganisation design questions

Strong evidenceInstitutional survey · 2026#

AI Readiness: Building the Bridge from Higher Education to Work

Pearson and AWS · Pearson and Amazon Web Services, 13 April 2026

Finding: More than 2,700 responses from learners, higher-education leaders and employers across six countries, framing readiness as technical fluency combined with problem solving, communication, collaboration, adaptability and human-AI judgement.
Evidence strength: Institutional survey, multi-country, commissioned by organisations that sell the remedy.
The SuperSkills read: The most directly relevant source on the education-to-work transition, which is where the missing rungs problem begins. Its definition of readiness is notable: judgement and adaptability sit alongside technical fluency rather than behind it.

educationearly careersskills will ai replace entry level jobs · how humans learn with ai

Strong evidenceInstitutional survey · 2026#

The AI Workforce Pulse: The Adaptability Imperative

Pearson and Cognizant · Pearson and Cognizant, June 2026

Finding: A survey of 750 HR leaders across the US, UK and India, fielded March to April 2026 among directors and above at organisations of 1,000 or more employees.
Evidence strength: Institutional survey, decent sample of senior HR decision-makers, commissioned by vendors of the remedy.
The SuperSkills read: Useful for what HR leadership currently believes about middle managers, early careers and the durability of human skills. Belief is not evidence, but in workforce strategy what leaders believe determines what gets funded.

leadershipskillsorganisation design chro guide to ai

Strong evidenceInstitutional report · 2026#

The Human Advantage: Stronger Brains in the Age of AI

World Economic Forum with the McKinsey Health Institute · World Economic Forum, 15 January 2026

Finding: Frames human cognitive capability, or brain capital, as an economic asset to be actively maintained rather than assumed.
Evidence strength: Institutional report, launched at Davos, co-produced with a consultancy health institute. Framing and synthesis rather than new primary evidence.
The SuperSkills read: The clearest sign that the argument on this site has reached the mainstream institutional agenda. When the WEF publishes on maintaining human cognitive capability rather than on reskilling for new tools, the question has shifted from what people should learn to whether they retain the capacity to learn at all.
Questions it helps answer: Are human skills actually becoming scarcer?

agencyskillslearning human skills in the age of ai · capability debt

Strong evidenceField experiment · 2025#

The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers

Cui, Z. K., Demirer, M., Jaffe, S., Musolff, L., Peng, S. and Salz, T. · Management Science, 2025

Finding: Across three randomised experiments and 4,867 developers at Microsoft, Accenture and a Fortune 100 company, completed tasks rose 26.08 percent, with larger gains among less experienced developers.
Evidence strength: Randomised controlled trials at real firms, which is rare and valuable. The headline estimate carries a standard error of 10.3 percent, so the true effect is wide.
The SuperSkills read: The strongest evidence that AI gains are real in high-skilled work, and it repeats the pattern found in customer support: the least experienced gain most. Quote the 26 percent with its standard error, or not at all.
Questions it helps answer: Does AI help experienced or inexperienced workers more?

jobsexpertisehuman-AI collaboration staying valuable in the age of ai

Strong evidenceInstitutional survey · 2025#

2025 Global Human Capital Trends

Deloitte · Deloitte Insights

Finding: Frames the worker-organisation relationship as a set of unresolved tensions rather than a set of solved problems.
Evidence strength: Institutional survey. Causal claims. It is a sentiment survey by a firm that sells the remedies it recommends, and should be read with that in view.
The SuperSkills read: Large-scale practitioner sentiment, published by a firm that sells the remedies it recommends. Both facts matter.

organisation designleadership ai workforce strategy · chro guide to ai graded entry

Strong evidenceField experiment · 2025#

Shifting Work Patterns with Generative AI

Dillon, E., Jaffe, S., Immorlica, N. and Stanton, C. · Microsoft Research; NBER Working Paper 33795

Finding: A six-month randomised field experiment across roughly 6,000 knowledge workers found users spent about three fewer hours per week on email, with no significant change in time spent in meetings.
Evidence strength: Randomised, six months, cross-industry, measuring behaviour rather than self-report. Among the best-designed studies in this literature.
The SuperSkills read: Important because it measures what changed rather than what people said changed, and because the answer is modest and specific. Time moved out of email and did not move out of meetings, which is a useful corrective to claims of wholesale transformation.

organisation designjobs ai workforce strategy

Strong evidenceJournal article · 2025#

AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking

Gerlich, M. · Societies, 15(1), 6

Finding: A negative correlation between frequent AI use and critical-thinking scores, mediated by cognitive offloading, strongest among the youngest users.
Evidence strength: Peer-reviewed study. Causation, and it carries a published correction (Societies 2025, 15(9), 252) which anyone citing it should read alongside.
The SuperSkills read: Widely cited, and worth citing carefully. It shows correlation with a plausible mechanism, and it carries a published correction that most people quoting it have not read.

critical thinkingdeskilling ai and critical thinking · ai and human judgement · is screen time the same argument as ai use graded entry

Strong evidenceWorking paper · 2025#

Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI

Humlum, A. and Vestergaard, E. · NBER Working Paper 33777, revised March 2026

Finding: Precise null effects on earnings and hours two years after ChatGPT, ruling out effects larger than 2 percent, alongside substantial task reorganisation and new tasks in AI oversight and integration.
Evidence strength: Working paper, not peer-reviewed. That the same holds elsewhere. Denmark is high-trust, high-wage and heavily unionised, and two years is early.
The SuperSkills read: The best available answer to what has actually happened rather than what might. Its real lesson is that work reorganises long before earnings move, so pay data is the slowest signal a leader can watch.

jobsorganisation design ai workforce strategy · will ai replace my job · ai and human judgement graded entry

Strong evidenceWorking paper · 2025#

Generative AI and Jobs: A Refined Global Index of Occupational Exposure

International Labour Organization (with NASK) · ILO Working Paper 140, 20 May 2025

Finding: A refined global index of occupational exposure to generative AI, updating the ILO's 2023 analysis with a more internationally balanced view of augmentation versus automation.
Evidence strength: Institutional working paper with published methodology. Exposure modelling, so it estimates what could be affected rather than what will be.
The SuperSkills read: The best correction to the US-centric bias running through most of this literature. Exposure looks very different in economies with different occupational structures, and the ILO's finding that augmentation dominates automation in most contexts is the more defensible reading of the same underlying data.
Questions it helps answer: What is happening outside the US and UK?, Will AI replace tasks or whole jobs?

jobsskills will ai replace my job

Strong evidenceInstitutional survey · 2025#

The State of AI: Global Survey

McKinsey and Company, QuantumBlack · McKinsey, 5 November 2025

Finding: 1,993 respondents across 105 countries. 88 per cent of organisations use AI in at least one function, up from 78 per cent. Nearly two thirds have not begun scaling AI across the enterprise. 39 per cent report enterprise-level EBIT impact attributable to AI, and most of those put the figure below 5 per cent.
Evidence strength: Large multi-country sample and a consistent annual series. COMMERCIAL INTEREST: McKinsey sells AI transformation work, and a scaling gap is a market. Self-reported by executives about their own organisations, and EBIT attribution to any single cause is an opinion rather than an accounting fact.
Establishes: That adoption and effect have come apart. 88 per cent of organisations use AI somewhere; 39 per cent report any enterprise-level EBIT impact, and most of those put it below 5 per cent, with nearly two thirds not yet scaling at all.
Does not establish: That AI does not produce value. Executives self-reporting EBIT attribution is an opinion about causation, not an accounting fact, and absence of measured effect two years in is what the Danish study would also predict for a technology whose early gains sit in task reorganisation.
Where it stands: Supports the argument The clearest available counterweight to the rhetoric of ubiquitous transformation, and it comes from a firm that sells transformation.
The SuperSkills read: The most useful available counterweight to the rhetoric of ubiquitous transformation, and it comes from a firm with every reason to say the opposite. Adoption at 88 per cent alongside enterprise EBIT impact at 39 per cent, mostly under 5 per cent, is the gap between having the tools and changing anything.
Questions it helps answer: Where is AI actually creating economic value in our organisation?, What return should we expect from our AI investment?, How do you measure AI adoption properly?, How much should we be investing in AI?

organisation designleadershipjobs questions

Strong evidenceInstitutional survey · 2025#

2025 Work Trend Index Annual Report: The Year the Frontier Firm is Born

Microsoft and LinkedIn · Microsoft WorkLab, April 2025

Finding: Describes the emergence of firms organised around human-agent teams and a shift towards workers managing AI agents.
Evidence strength: Institutional survey. Independence. Microsoft sells the tools whose adoption it is measuring, and telemetry measures usage rather than value.
The SuperSkills read: Kept as the longitudinal predecessor to the 2026 edition, which moved much closer to the human-agency question.

organisation designagency ai agents and human judgement · ai workforce strategy graded entry

Strong evidenceInstitutional analysis · 2025#

OECD AI Capability Indicators: Technical Report

OECD · OECD Publishing, Paris, November 2025

Finding: Builds nine indicator scales that rate AI capability against human ability levels, so machine and human performance can be compared on the same dimensions.
Evidence strength: Institutional methodology report with expert elicitation. New and not yet widely validated; the scales are judgement-based rather than benchmark-derived.
The SuperSkills read: The most serious attempt anywhere to answer what AI can do relative to a person, on comparable scales, rather than through benchmark scores that mean nothing to a manager. If it matures, it becomes the instrument this whole field has lacked.
Questions it helps answer: How do you assess capability rather than output?

technology capabilityskills what stays human · human ai decision making

Strong evidenceInstitutional analysis · 2025#

PwC Global AI Jobs Barometer 2025

PwC · PwC, 2025

Finding: Reports wage premiums for AI skills and shifting skill requirements in AI-exposed occupations.
Evidence strength: Institutional modelling. Causation, and job advertisements describe what employers ask for rather than what the work requires.
The SuperSkills read: Kept as the longitudinal predecessor to the 2026 edition, whose entry-level finding is the more striking one.

jobsearly careers will ai replace entry level jobs graded entry

Strong evidenceGovernment survey · 2025#

AI Labour Market Study and Survey

UK Department for Science, Innovation and Technology · DSIT

Finding: UK employer demand for AI skills, the shortages employers report, and how stated AI skill requirements are changing in vacancies.
Evidence strength: Employer-side UK data with a government sampling frame. Employers describing their own demand, which is a statement of intent as much as of fact, and vacancy text is a noisy instrument for skill requirements.
The SuperSkills read: The demand side of the UK picture, which pairs with the DSIT exposure assessment on the supply side.
Questions it helps answer: Which skills become more valuable as AI improves?

skillsjobs questions

Strong evidenceInstitutional report · 2025#

Digital Progress and Trends Report 2025: Strengthening AI Foundations

World Bank · World Bank

Finding: A data account of the global AI divide across four foundations: connectivity, compute, context and competency. High-income countries hold 87 per cent of notable AI models, 86 per cent of AI start-ups and 91 per cent of venture funding while holding 17 per cent of the world's people, and 77 per cent of co-location data centre capacity against under 0.1 per cent in low-income countries. Generative AI vacancies rose ninefold from 2021 to 2024, with one in five of them in middle-income countries.
Evidence strength: Institutional compilation with a development mandate and no product to sell. Infrastructure and vacancy data rather than measurement of what happens to workers, and vacancy counts are a noisy proxy for demand.
Establishes: The scale of the global AI divide across infrastructure, capital and skills: high-income countries hold 87 per cent of notable models, 91 per cent of venture funding and 77 per cent of data centre capacity while holding 17 per cent of the world's people.
Does not establish: Anything about what AI does to workers anywhere. It measures foundations, not effects, and its labour material rests on vacancy counts, which are a statement of advertised demand rather than of work performed.
Where it stands: Complicates the argument Complicates by relocating the question. This estate argues about what happens to capability when a tool does the work. For most of the world the prior question is whether the tool, the electricity and the training are available at all, and the answer is largely no. The capability argument is a rich-country argument and this is the evidence that says so.
The SuperSkills read: Correction for the rich-country bias that this canon would otherwise have. It also supplies the distributional point the estate has argued without evidence: the jobs being created are polarised, a small number of highly paid AI roles alongside a large volume of low-paid data work.
Questions it helps answer: What is happening outside the US and UK?, Who ends up worse off as AI spreads?

jobsskillseducation questions

Strong evidenceWorking paper · 2024#

The Rapid Adoption of Generative AI

Bick, A., Blandin, A. and Deming, D. J. · NBER Working Paper 32966

Finding: By late 2024, nearly 40 percent of US adults aged 18-64 used generative AI and 23 percent of employed respondents had used it for work in the previous week, but only 1 to 5 percent of all work hours were assisted.
Evidence strength: Working paper, not peer-reviewed. Quality of use. Self-reported use counts any use at all.
The SuperSkills read: Enormous reach, thin penetration. The gap between how many people use AI and how few hours it touches is the whole of the adoption-versus-redesign argument.

jobsorganisation design ai workforce strategy · why learn to prompt is weak career advice graded entry

Strong evidenceInstitutional analysis · 2024#

Gen-AI: Artificial Intelligence and the Future of Work

Cazzaniga, M. et al. (International Monetary Fund) · IMF Staff Discussion Note 2024/001, 14 January 2024

Finding: Estimates AI exposure across advanced, emerging and low-income economies, and analyses the implications for inequality within and between countries.
Evidence strength: Institutional macro analysis, not peer-reviewed. Built on occupational exposure measures with the usual limits.
The SuperSkills read: The macroeconomic counterweight to company surveys, and the only major source here that takes inequality between countries seriously. Its uncomfortable finding is that advanced economies face more exposure and are better placed to benefit, which widens rather than narrows global gaps.
Questions it helps answer: Will AI replace my job?, How much of the workforce is actually exposed to AI?

jobs will ai replace my job

Strong evidenceBook · 2024#

Human + Machine: Reimagining Work in the Age of AI

Daugherty, P. R. and Wilson, H. J. · Harvard Business Review Press, updated and expanded edition

Finding: Sets out the missing middle, where humans and machines collaborate rather than substitute, and names hybrid roles including trainer, explainer and sustainer.
Evidence strength: Important perspective from consultancy practice, with case material rather than controlled evidence.
The SuperSkills read: The most usable role taxonomy for human-AI work, and it predates the generative wave by six years. Cite the 2024 expanded edition, which adds generative AI; the 2018 original is often quoted for claims it does not make.

human-AI collaborationorganisation design human ai decision making · ai workforce strategy

Strong evidenceJournal article · 2024#

Generative AI enhances individual creativity but reduces the collective diversity of novel content

Doshi, A. R. and Hauser, O. P. · Science Advances, 10(28)

Finding: AI-assisted stories were rated more creative, better written and more enjoyable, with the largest gains for the least creative writers, and were markedly more similar to one another.
Evidence strength: Peer-reviewed study. That this generalises beyond one short creative task with one form of assistance.
The SuperSkills read: Individual creative quality and collective creative range move in opposite directions. A social dilemma in which everyone is individually right and the result is duller.

agencyorganisation design what stays human graded entry

Strong evidenceInstitutional analysis · 2024#

A new future of work: The race to deploy AI and raise skills in Europe and beyond

McKinsey Global Institute · McKinsey Global Institute, May 2024

Finding: Projects large-scale occupational transitions and rising demand for social, emotional and higher cognitive skills.
Evidence strength: Institutional modelling. What will happen. Scenario models are assumption-driven, and McKinsey's prior transition estimates have moved substantially between editions.
The SuperSkills read: A transparent scenario model, useful for direction. Its prior transition estimates have moved substantially between editions.

jobsskills ai workforce strategy · will ai replace my job graded entry

Strong evidenceCompetency framework · 2024#

AI competency framework for teachers

Miao, F. and Cukurova, M. (UNESCO) · UNESCO, Paris

Finding: Defines what teachers need to know and be able to do with AI, across a human-centred mindset, ethics, foundations, pedagogy and professional learning.
Evidence strength: Official UNESCO framework, expert-developed and internationally consulted. Normative guidance rather than evidence.
The SuperSkills read: Notable for leading with a human-centred mindset and ethics rather than tool skills, which is the opposite ordering from most corporate AI-literacy training. A useful model for what capability-first rather than tool-first development looks like in practice.

frameworkseducation how humans learn with ai · chro guide to ai

Strong evidenceJournal article · 2023#

Comparing Physician and Artificial Intelligence Chatbot Responses to Patient Questions Posted to a Public Social Media Forum

Ayers, J. W. et al. · JAMA Internal Medicine, 183(6), 589-596

Finding: Chatbot responses were rated good or very good quality 78.5 percent of the time against 22.1 percent for physicians, and empathetic or very empathetic 45.1 percent against 4.6 percent.
Evidence strength: Peer-reviewed study. That a machine can care for anyone. Doctors answering strangers free of charge between patients are not doing the job they trained for.
The SuperSkills read: The study that should end the claim that empathy is safe from AI. On the observable performance of empathy in text, the machine already wins comfortably.

agencyethics what stays human · outsourced recognition graded entry

Strong evidenceWorking paper · 2023#

Generative AI and Jobs: A Global Analysis of Potential Effects on Job Quantity and Quality

Gmyrek, P., Berg, J. and Bescond, D. (International Labour Organization) · ILO Working Paper 96, 21 August 2023

Finding: Finds that most jobs are more likely to be augmented than automated, with clerical work the most exposed category and a markedly larger effect for women in higher-income countries.
Evidence strength: Institutional working paper with published methodology. Exposure analysis rather than measured outcome; superseded in part by the 2025 refined index.
The SuperSkills read: The first major analysis to conclude that augmentation dominates automation, and still among the most balanced. Its finding on clerical work and gendered exposure is one of the few in this literature that engages seriously with who bears the cost.

jobsskills will ai replace my job

Strong evidenceInstitutional analysis · 2023#

The Skills Imperative 2035: An analysis of the demand for skills in the labour market in 2035 (Working Paper 3)

NFER · National Foundation for Educational Research, with the University of Sheffield, funded by the Nuffield Foundation, May 2023

Finding: Projects rising demand for a set of essential employment skills in the UK to 2035.
Evidence strength: Institutional modelling. Precision. It maps US skill data onto UK occupations, and a revised working paper corrects coding errors in the underlying labour force survey.
The SuperSkills read: A rare UK-specific, independently funded projection. Maps US skill data onto UK occupations, which is its main limitation.

skillseducation human skills in the age of ai graded entry

Strong evidenceGovernance framework · 2023#

Artificial Intelligence Risk Management Framework (AI RMF 1.0)

National Institute of Standards and Technology · NIST AI 100-1, 26 January 2023; Generative AI Profile, NIST AI 600-1, July 2024

Finding: A voluntary framework organised around governing, mapping, measuring and managing AI risk, with human oversight and accountability as explicit functions.
Evidence strength: Official framework, widely adopted as a de facto standard. Voluntary and process-oriented; it describes what to do rather than evidencing what works. Currently under revision.
The SuperSkills read: The practical vocabulary for the accountability argument, and the document a board will already recognise. Its weakness is instructive: it can be satisfied procedurally, by an organisation that documents oversight without exercising it, the failure mode this research is concerned with.
Questions it helps answer: What should a board ask about AI?, When should we stop or reverse an AI deployment?

governanceorganisation designethics human ai decision making · chro guide to ai

Strong evidenceInstitutional analysis · 2023#

OECD Skills Outlook 2023: Skills for a Resilient Green and Digital Transition

OECD · OECD Publishing, Paris, November 2023

Finding: Analyses the skills required for green and digital transitions across member economies.
Evidence strength: Institutional modelling. Anything AI-specific and current. The underlying data collection predates the generative-AI period.
The SuperSkills read: A methodologically transparent cross-national baseline, from data collected before the generative-AI period.

skillseducation human skills in the age of ai graded entry

Strong evidenceInstitutional survey · 2023#

The Future of Jobs Report 2023

World Economic Forum · World Economic Forum, Geneva, April 2023

Finding: Analytical thinking leads, with creative thinking second, and a growing emphasis on self-efficacy skills.
Evidence strength: Institutional survey. Considered judgement on generative AI. It was fielded too early for that.
The SuperSkills read: The first post-ChatGPT edition, fielded five months after launch and therefore too early for considered judgement on it.

skillsjobs human skills in the age of ai graded entry

Strong evidenceJournal article · 2023#

Why Johnny Can't Prompt: How Non-AI Experts Try (and Fail) to Design LLM Prompts

Zamfirescu-Pereira, J. D., Wong, R. Y., Hartmann, B. and Yang, Q. · CHI 2023

Finding: Non-experts approached prompting opportunistically rather than systematically, over-generalised from single successes and failures, and struggled to form an accurate model of the system.
Evidence strength: Peer-reviewed study. That this persists. It used 2023 models, and providers are actively engineering the difficulty away.
The SuperSkills read: The strongest case for teaching prompting, included on a site that argues prompting is weak career advice. It shows the difficulty is real; it does not show the difficulty is durable.

skillshuman-AI collaboration why learn to prompt is weak career advice graded entry

Strong evidenceLabour-market analysis · 2022#

Pearson Skills Outlook: Power Skills

Pearson · Pearson, November 2022

Finding: Analysis of 21.8 million job advertisements across the US, Canada, the UK and Australia found human capabilities dominating employer demand.
Evidence strength: Very large observed-demand dataset. Note the date: this is a 2022 study forecasting demand to 2026, not a 2026 publication.
The SuperSkills read: Valuable as pre-ChatGPT evidence. It establishes that human capabilities were already the scarce thing in the labour market before generative AI arrived, which undercuts the claim that this is a fashion created by the current wave.

skillsjobs human skills in the age of ai

Strong evidenceCompetency framework · 2022#

DigComp 2.2: The Digital Competence Framework for Citizens

Vuorikari, R., Kluzer, S. and Punie, Y. (European Commission JRC) · Publications Office of the European Union, JRC128415

Finding: The European reference framework for digital competence, updated in 2022 with examples covering AI and datafication.
Evidence strength: Official framework, expert-developed and widely adopted in European policy. A normative taxonomy rather than empirical evidence.
The SuperSkills read: One of the taxonomies any human-capability framework is implicitly competing with, and worth knowing for that reason. Its 2022 AI additions are thin on judgement and accountability, and that gap is the one the SuperSkills framework addresses.

frameworksskillseducation human skills in the age of ai

Strong evidenceInstitutional survey · 2021#

Defining the skills citizens will need in the future world of work

McKinsey and Company · McKinsey Public and Social Sector Practice, June 2021

Finding: Identifies distinct elements of talent, the DELTAs, associated with employment, income and job satisfaction.
Evidence strength: Institutional survey. Anything about AI. It was fielded in 2019, before the generative-AI period entirely.
The SuperSkills read: An unusually large individual-level dataset on skills and outcomes, fielded in 2019 and therefore silent on AI.

skills human skills in the age of ai graded entry

Strong evidenceInstitutional survey · 2020#

The Future of Jobs Report 2020

World Economic Forum · World Economic Forum, Geneva, October 2020

Finding: Named critical thinking and problem solving as leading skills, and forecast large-scale reskilling need.
Evidence strength: Institutional survey. A clean read on AI. The 2020 edition is dominated by COVID-era disruption.
The SuperSkills read: The pandemic edition. Read for what the field expected in 2020, and discounted for the disruption dominating it.

skillsjobs human skills in the age of ai graded entry

Strong evidenceJournal article · 2019#

The role of deliberate practice in expert performance: revisiting Ericsson, Krampe and Tesch-Romer (1993)

Macnamara, B. N. and Maitra, M. · Royal Society Open Science, 6, 190327

Finding: Accumulated practice explained considerably less of the difference between performers than the original is usually taken to claim.
Evidence strength: Peer-reviewed study. That practice does not matter. It does; the simple dose-response reading is what fails.
The SuperSkills read: Included deliberately because it complicates the argument this research relies on. Practice matters; the simple dose-response reading of Ericsson does not survive.

expertiselearning how humans learn with ai graded entry

Strong evidenceInstitutional survey · 2018#

The Future of Jobs Report 2018

World Economic Forum · World Economic Forum, Geneva, September 2018

Finding: Set out expected skill demand to 2022, with analytical thinking and innovation, active learning and creativity leading the list.
Evidence strength: Institutional survey. A representative picture of employers. Respondents are drawn from a self-selected membership network.
The SuperSkills read: Kept deliberately as a longitudinal comparator. Its predictions for 2022 are now checkable, which is a rarer thing than another forecast.

skillsjobs human skills in the age of ai graded entry

Strong evidenceCompetency framework · 2016#

EntreComp: The Entrepreneurship Competence Framework

Bacigalupo, M., Kampylis, P., Punie, Y. and Van den Brande, G. (European Commission JRC) · Publications Office of the European Union, JRC101581

Finding: A framework of fifteen competences across ideas and opportunities, resources, and action, describing initiative and value creation as learnable capabilities.
Evidence strength: Official framework, expert-developed. Normative rather than empirical.
The SuperSkills read: The best existing treatment of initiative and opportunity recognition as trainable rather than innate. Useful as prior art for anyone arguing that human capabilities can be developed systematically rather than merely possessed.

frameworksskills human skills in the age of ai

Strong evidenceBook · 2015#

The Future of the Professions: How Technology Will Transform the Work of Human Experts

Susskind, R. and Susskind, D. · Oxford University Press, updated edition

Finding: Argues that the professions as currently organised are a solution to the problem of unevenly distributed expertise, and that technology dissolves the grand bargain that sustained them.
Evidence strength: Important perspective with substantial case material across law, medicine, accountancy, education and consulting.
The SuperSkills read: The most rigorous treatment of what happens to professional expertise specifically, written a decade ago and increasingly accurate. Read it alongside Polanyi: the Susskinds argue much professional work can be decomposed and routinised, and Polanyi explains what resists that.

expertisejobsorganisation design staying valuable in the age of ai · what stays human

Strong evidenceArticle · 1997#

Dynamic Capabilities and Strategic Management

Teece, D. J., Pisano, G. and Shuen, A. · Strategic Management Journal, 18(7), 509-533

Finding: Competitive advantage under rapid technological change rests on the firm's ability to integrate, build and reconfigure internal and external competences, rather than on any fixed stock of resources.
Evidence strength: Theoretical synthesis. The construct has been criticised as difficult to operationalise and measure.
The SuperSkills read: The reason a capability audit cannot be a static inventory. What is being asked is whether the organisation can still reconfigure, which is a different question from what it currently holds.
Questions it helps answer: What is an organisation's capability, as distinct from its people's?, How do you keep expertise in an organisation?

organisation designframeworks

Important perspectiveBook · 2026#

SuperSkills: The Seven Human Skills for the Age of AI

Hirji, R. · Kogan Page

Finding: Argues that increasingly capable AI reaches human judgement before it reaches jobs, and that most organisations drift into adoption rather than deciding in advance which decisions stay human.
Evidence strength: Important perspective, not evidence. A trade book, and the person who compiled this list is its author, which is stated here rather than left for the reader to notice.
The SuperSkills read: Included because a map of the field that leaves out the mapmaker's own position hides where the compiler is standing rather than achieving neutrality. Do not take the argument on the book's authority. The graded sources it rests on are in the evidence base, including the entries that qualify it and the one that contradicts it directly.

judgementskillsdeskilling ai and human judgement · capability debt · evidence

Important perspectiveEncyclical · 2026#

Magnifica Humanitas: on safeguarding the human person in the time of artificial intelligence

Leo XIV · Encyclical Letter, The Holy See, 15 May 2026

Finding: Argues that AI use can weaken creativity and judgement (100), that every technology shapes those who use it and education must teach when not to use it (140), that current approaches can paradoxically de-skill workers (150, quoting Antiqua et Nova), that a slower pace of adoption is not opposition to progress (106), and that moral judgment cannot be reduced to calculation (198).
Evidence strength: A doctrinal and moral argument, not evidence. It measures nothing and tests nothing, and its deskilling claim is quoted from an earlier Vatican note rather than from a study. Read for the position it takes and the audience it reaches.
The SuperSkills read: The deskilling argument, stated in a magisterial document addressed to a global audience on the 135th anniversary of Rerum Novarum. Almost every widely read piece on AI and human capability is American and written for an Anglophone readership; this is neither, and it arrives at the same conclusions as the secular literature by an entirely different route. Paragraph 140, on the ease of obtaining answers extinguishing the desire to ask questions, is the clearest statement of the curiosity argument published anywhere in 2026.

deskillingjudgementethicseducationjobs the best writing on ai · superskill curiosity

Important perspectiveInstitutional report · 2026#

Mind the Learning Gap: How We Can Achieve AI's Full Potential

Pearson · Pearson, January 2026

Finding: Argues that AI productivity depends on learning and augmentation rather than deployment alone, estimating that augmentation could add 4.8 to 6.6 trillion dollars to the US economy by 2034.
Evidence strength: Economic modelling by a learning company arguing for the importance of learning. The direction is well argued; the figure is an estimate with wide bounds.
The SuperSkills read: Included as a serious argument rather than as evidence. Its core proposition, that the return on AI depends on whether people learn alongside it, is the constructive form of the capability-debt argument.

learningorganisation design ai workforce strategy

Important perspectiveSuperSkills position · 2026#

What I argue, and what I do not

Rahim Hirji · thesuperskills.com

Finding: Ten interpretive positions, each separated from its evidence and each stating what it does not claim.
Evidence strength: Argument, not evidence. It is the author's reading of a body of work held elsewhere on the estate, and the page exists so that reading can be told apart from the findings it rests on.
The SuperSkills read: The page to read before deciding how much of the rest of this estate to believe. If a claim on any other page is not traceable to graded evidence, it should be traceable to one of these ten positions or it should not be there.

deskillingjudgementexpertise what i argue

Important perspectiveSuperSkills position · 2026#

Capability debt

Rahim Hirji · thesuperskills.com

Finding: Proposes capability debt: a cost taken on when AI performs the activity through which a capability was previously maintained, invisible on every measure an organisation currently watches, and paid later.
Evidence strength: A coinage and a frame rather than a finding. The component evidence on skill decay, offloading and automation is graded; the synthesis into a debt metaphor is the author's, and the metaphor is not itself tested.
The SuperSkills read: The estate's central proposition, and the one most likely to be wrong in an interesting way. Its testable form is whether organisations that removed developmental work can still do that work unaided later, which nobody has yet measured.

deskillingorganisation designexpertise capability debt

Important perspectiveSuperSkills position · 2026#

Questions About Humans and AI: the map of the territory

Rahim Hirji · thesuperskills.com

Finding: A maintained architecture of the questions people ask about AI and human capability, each marked answered, partial or open, with the open ones left visible and external work mapped onto them.
Evidence strength: Not evidence at all. It is an ontology, and its claim is completeness of COVERAGE rather than correctness of answers. Fewer than half the questions have a dedicated answer, which the page states in its own opening line.
The SuperSkills read: The object the rest of this canon is organised around. Every major report here is mapped to the questions it bears on, which is what turns a reading list into a map.

judgementskillsjobs questions

Important perspectiveWorking paper · 2025#

Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task

Kosmyna, N. et al. · MIT Media Lab preprint, arXiv:2506.08872

Finding: The LLM group showed the weakest brain connectivity and the lowest sense of ownership over their own writing.
Evidence strength: Working paper, not peer-reviewed. Anything settled. 54 participants, a preprint, and reproducibility flagged by its own commentators. Treat claims of proof with suspicion.
The SuperSkills read: Included because it is quoted constantly and rests on 54 participants in a preprint. A useful test of whether someone has read what they are citing.

critical thinking ai and critical thinking · ai and human judgement · is screen time the same argument as ai use graded entry

Important perspectiveInstitutional analysis · 2025#

Guidance on AI and Children, Version 3.0: Recommendations for AI policies and systems that uphold child rights

UNICEF Innocenti · UNICEF Innocenti, December 2025

Finding: Sets out ten requirements and 48 recommendations for AI systems and policy affecting children.
Evidence strength: Institutional modelling. Empirical claims about learning or capability effects. It is a normative guidance document.
The SuperSkills read: A rights-based standard developed with children rather than about them, which almost nothing else in this space does.

educationethics how humans learn with ai graded entry

Important perspectiveBook · 2024#

Nexus: A Brief History of Information Networks from the Stone Age to AI

Harari, Y. N. · Random House, New York

Finding: History is read as a sequence of information networks, with a standing tension between networks that pursue truth and networks that impose order. AI is treated as the first technology that is an information agent rather than an information tool, able to originate and decide rather than only transmit.
Evidence strength: Synthesis for a general readership. Sweeping, and criticised by specialists for compression.
The SuperSkills read: The widest available frame for what changes when the network stops being passive. Included as the most-read articulation of the position rather than as evidence, and useful precisely because audiences arrive already holding it.

technology capabilitygovernancecritical thinking neither ai hype nor doom

Important perspectiveBook · 2023#

Power and Progress: Our Thousand-Year Struggle Over Technology and Prosperity

Acemoglu, D. and Johnson, S. · PublicAffairs

Finding: A thousand years of evidence that technology's gains are distributed according to choices about direction and power, not automatically by the technology itself.
Evidence strength: Important perspective, historically grounded, contested by other economists.
The SuperSkills read: The essential dissent. It challenges the assumption underneath most workforce writing, including some of this research: that if capability is protected, the benefits follow. Acemoglu and Johnson argue that who benefits is decided politically, and that is a claim this site does not answer.

ethicsjobsorganisation design how should leaders respond to ai

Important perspectiveBook · 2021#

Think Again: The Power of Knowing What You Don't Know

Grant, A. · Viking, New York

Finding: Rethinking is a distinct skill, separable from intelligence, and is helped by adopting the posture of a scientist rather than that of a preacher, prosecutor or politician.
Evidence strength: Popular synthesis with original studies woven in. Organisational psychology rather than measurement of AI effects.
The SuperSkills read: The accessible route into the same territory as Galef, and the one most leadership audiences have already read, which makes it useful shared vocabulary rather than new evidence.

critical thinkinglearning what is intellectual humility · how do i get ai to challenge me

Important perspectiveBook · 2019#

Range: Why Generalists Triumph in a Specialized World

Epstein, D. · Riverhead Books, New York

Finding: In wicked domains, where feedback is poor and rules are unstable, breadth, delayed specialisation and analogical thinking outperform early specialisation.
Evidence strength: Popular synthesis, selected to make a case. The deliberate-practice literature it argues against is stronger in kind domains than Epstein sometimes allows.
The SuperSkills read: The necessary counterweight to Ericsson, and the reason this estate should not tell people to specialise harder. It also supplies the distinction between kind and wicked environments that decides where recognitional judgement can be trusted.

skillsexpertiselearning staying valuable in the age of ai · what should i tell my children to study

Important perspectiveBook · 2019#

The Technology Trap: Capital, Labor, and Power in the Age of Automation

Frey, C. B. · Princeton University Press

Finding: A history of automation showing that periods of technological progress have frequently produced decades of falling wages and political backlash before broad gains arrived.
Evidence strength: Important perspective, historically grounded. Frey is also co-author of the widely quoted and widely criticised 2013 estimate that 47 percent of US jobs were at risk.
The SuperSkills read: The long view, and a warning against arguing from eventual outcomes. Even where technology ultimately raised living standards, the transition punished a generation. Read it as the reason the design choices this research argues for cannot be left to work themselves out.

jobsethics how should leaders respond to ai

Important perspectiveBook · 2019#

The Age of Surveillance Capitalism: The Fight for a Human Future at the New Frontier of Power

Zuboff, S. · PublicAffairs, New York

Finding: Dominant technology firms operate an economic logic that treats human experience as free raw material for extraction and behavioural prediction, sold as certainty about what people will do.
Evidence strength: Extensively documented argument. Contested by economists on the framing rather than the facts.
The SuperSkills read: The power dimension this research has under-covered. It supplies the reason the same AI deployment can be read as capability building or as extraction depending on who captures the surplus, which is the question most human-capability writing avoids.

agencygovernanceethics who captures the productivity gains from ai

Important perspectiveBook · 2015#

The World Beyond Your Head: On Becoming an Individual in an Age of Distraction

Crawford, M. B. · Farrar, Straus and Giroux, New York

Finding: Attention is not simply willed. It is cultivated through embodied skill and engagement with environments that resist us, and the attention economy erodes the conditions under which autonomous selfhood forms.
Evidence strength: Philosophical argument drawing on phenomenology and case studies. Not empirical.
The SuperSkills read: The strongest account of why frictionless tools are not straightforwardly good. Where this research argues that removing difficulty removes the practice, Crawford argues it also removes the conditions for agency, which is a stronger and more contestable claim.

agencyskillscritical thinking is attention a trainable skill · what stays human

Important perspectiveBook · 2014#

A More Beautiful Question: The Power of Inquiry to Spark Breakthrough Ideas

Berger, W. · Bloomsbury USA, New York

Finding: Innovation follows a three-stage question sequence, why then what if then how, and organisations systematically suppress it after childhood.
Evidence strength: Journalistic synthesis and interview. Illustrative rather than evidential.
The SuperSkills read: Useful for the shape of the argument and for audiences, and it should not be leaned on for evidence. Rothstein and Santana carry the teachable-method claim.

critical thinkingframeworks what makes a good question

Important perspectiveBook · 2012#

So Good They Can't Ignore You: Why Skills Trump Passion in the Quest for Work You Love

Newport, C. · Business Plus, New York

Finding: Satisfying work follows from building rare and valuable skills, career capital, which can then be traded for autonomy, rather than from identifying a pre-existing passion.
Evidence strength: Argument with case illustration. Not empirical.
The SuperSkills read: The career-advice counterpart to the capability argument, and useful because it reaches the audience most exposed to advice about learning to prompt.

skillsearly careers why learn to prompt is weak career advice · staying valuable in the age of ai

Important perspectiveBook · 2008#

The Craftsman

Sennett, R. · Yale University Press, New Haven

Finding: Craftsmanship, the desire to do a job well for its own sake, is a durable human capacity that extends well beyond manual trades, and it requires sustained practice and engagement with resistant material.
Evidence strength: Historical and sociological essay. Argument rather than measurement.
The SuperSkills read: The case that skill is not only instrumental. It supplies the reason people resist automation of work they are good at, which productivity framings cannot see, and it connects practice to satisfaction rather than only to competence.

skillsexpertiselearning what is productive struggle · what stays human

Important perspectiveBook · 2004#

The Paradox of Choice: Why More Is Less

Schwartz, B. · Ecco, New York

Finding: Past a threshold, more options increase anxiety, paralysis and regret rather than welfare, and the effect is strongest in people seeking the objectively best option rather than a good-enough one.
Evidence strength: Popular synthesis of choice research. Several constituent effects have been contested since.
The SuperSkills read: Relevant because AI expands the option set at almost no cost, which is treated everywhere as an unambiguous gain. Schwartz is the reason to check that assumption.

agencyjudgement should i let ai make personal decisions for me

Important perspectiveBook · 1984#

Technology and the Character of Contemporary Life: A Philosophical Inquiry

Borgmann, A. · University of Chicago Press, Chicago

Finding: Modern technology follows a device paradigm: it delivers a commodity while concealing the machinery, skill and social context that produced it. A fireplace requires wood, tending and company; central heating delivers warmth and requires nothing. What disappears is the engagement, not the outcome.
Evidence strength: Philosophy of technology. Conceptual, and the source text for the device paradigm.
The SuperSkills read: The best available description of what a good AI interface does to work. The output improves, the engagement vanishes, and nothing in the result signals what was removed. Borgmann's focal practices, the things people deliberately keep difficult, are the philosophical form of the argument for protecting unaided work.

technology capabilityagencyskills capability debt · should i let ai summarise everything i read

Important perspectiveBook · 1973#

Tools for Conviviality

Illich, I. · Harper and Row, New York

Finding: Tools and institutions cross a threshold beyond which they stop serving human purposes and begin to deskill and dominate their users. Illich distinguishes convivial tools, which expand what a person can do under their own direction, from manipulative ones, which make people dependent on the system that supplies them.
Evidence strength: Political and philosophical argument. No measurement.
The SuperSkills read: The 1973 version of the augmentation-versus-automation distinction, and a sharper one, because Illich makes dependence the criterion rather than task allocation. A tool that produces a better output while leaving the user less able to act without it has crossed the threshold he describes.

agencyskillstechnology capability using ai without dependency · am i becoming dependent on ai

FoundationBook · 2021#

The Scout Mindset: Why Some People See Things Clearly and Others Don't

Galef, J. · Portfolio, New York

Finding: Accurate belief depends less on intelligence than on a set of trainable habits: wanting to know what is true rather than to defend a preferred conclusion. Galef calls the two orientations scout and soldier.
Evidence strength: Popular synthesis of judgement and forecasting research. Argument and illustration rather than new data.
The SuperSkills read: The clearest statement that critical thinking is an orientation rather than a technique, which matters because a system that will argue either side for free makes the orientation the only thing left doing work.

critical thinkingjudgement what is critical thinking · how do i get ai to challenge me

FoundationJournal article · 2020#

Habitual use of GPS negatively impacts spatial memory during self-guided navigation

Dahmani, L. and Bohbot, V. D. · Scientific Reports, 10, 6310, 14 April 2020. DOI 10.1038/s41598-020-62877-0. Read in full at source 5 September 2026

Finding: Cross-sectionally, greater lifetime GPS experience was associated with lower use of hippocampus-dependent spatial strategies (r = -0.22 on the first probe trial), lower navigation strategy scores (r = -0.20), poorer map drawing (r = -0.22) and fewer landmarks noticed (r = -0.26). In the 13-person follow-up, hours of GPS use since first testing tracked a steeper decline in spatial memory strategy use (r = -0.68) and in map drawing (r = -0.52).
Evidence strength: Peer-reviewed study. Anything about the brain, because no scan was taken. Anything about dementia, Alzheimer's or atrophy: those words appear nowhere in the paper. And little with confidence about the longitudinal effect, which rests on 13 people from an unplanned follow-up. The authors' own words: they caution against any strong conclusions as spurious correlations are possible. Their discussion elsewhere uses notably firmer causal language than that sentence licenses. Nothing here transfers to reasoning: spatial memory is not judgement.
The SuperSkills read: The closest long-run analogue we have for what habitual reliance does to a human capability. Spatial memory is not judgement, and the analogy should carry weight without carrying certainty.

deskillingexpertise does gps damage your brain · what is cognitive offloading · using ai without dependency · ai and human judgement graded entry

FoundationJournal article · 2019#

Algorithm Appreciation: People Prefer Algorithmic to Human Judgment

Logg, J. M., Minson, J. A. and Moore, D. A. · Organizational Behavior and Human Decision Processes, 151, 90-103

Finding: People often weight algorithmic advice MORE heavily than human advice. Domain experts are the notable exception.
Evidence strength: Peer-reviewed study. Which tendency dominates in any given workplace.
The SuperSkills read: Algorithm appreciation, the mirror image of aversion. Together they show miscalibration runs in both directions, with domain experts the notable exception.

judgementhuman-AI collaboration human ai decision making graded entry

FoundationBook · 2018#

Thinking in Bets: Making Smarter Decisions When You Don't Have All the Facts

Duke, A. · Portfolio, New York

Finding: Decisions are bets under uncertainty, and the quality of a decision must be judged separately from the quality of its outcome. Duke calls the conflation resulting.
Evidence strength: Practitioner account drawing on professional poker and decision research.
The SuperSkills read: The cleanest available statement of the decision-versus-outcome distinction, which this estate already argues and had not sourced. It also gives individuals the language for holding a view in probabilities, which is what makes revision cheap.

judgementcritical thinking decision quality · should i let ai make personal decisions for me

FoundationBook · 2016#

Peak: Secrets from the New Science of Expertise

Ericsson, K. A. and Pool, R. · Houghton Mifflin Harcourt, New York

Finding: Expert performance comes from deliberate practice: focused work at the edge of current ability, with feedback and correction, usually under a teacher. Not repetition, and not time served.
Evidence strength: Author's summary of a career of research. The strong version of the claim, that practice explains most of the variance, was weakened by later meta-analysis.
The SuperSkills read: It defines the mechanism precisely enough to ask the AI question properly: which of those difficult, feedback-rich repetitions disappear when the machine makes the first attempt. Read alongside Macnamara and Maitra, who cut the effect size down.

expertiselearningskills what is deliberate practice · how humans learn with ai

FoundationJournal article · 2016#

Cognitive Offloading

Risko, E. F. and Gilbert, S. J. · Trends in Cognitive Sciences, 20(9)

Finding: Defines cognitive offloading as using physical action or an external tool to reduce the mental demand of a task, and shows people offload not only when a task is hard but when they judge it to be hard.
Evidence strength: Peer-reviewed study. Nothing about generative AI specifically; it predates it.
The SuperSkills read: The mechanism underneath everything on this site. Its most useful finding is that we offload when we judge a task to be hard, and that judgement is frequently wrong.

judgementcritical thinking what is cognitive offloading · ai and human judgement graded entry

FoundationBook · 2016#

Technology and the Virtues: A Philosophical Guide to a Future Worth Wanting

Vallor, S. · Oxford University Press, New York

Finding: Emerging technologies call for technomoral virtues, cultivated habits of practical wisdom drawn from Aristotelian, Confucian and Buddhist traditions, rather than fixed rules, because the situations they create cannot be anticipated by regulation.
Evidence strength: Systematic philosophical work. Normative rather than empirical.
The SuperSkills read: The systematic argument behind Vallor's later and better-known AI Mirror. It matters here because it treats practical wisdom as something cultivated by practice and erodible by disuse, which is the ethical form of the capability argument this research makes economically.

ethicsagencyjudgement what stays human · what is judgement

FoundationJournal article · 2015#

Why Are There Still So Many Jobs? The History and Future of Workplace Automation

Autor, D. H. · Journal of Economic Perspectives, 29(3), 3-30. DOI 10.1257/jep.29.3.3

Finding: Automation substitutes for some tasks and complements others, and raises demand for the labour it complements, so forecasts that count only substitution overstate net job loss. Tacit, uncodifiable skills resist automation because, following Polanyi, we know more than we can tell.
Evidence strength: Peer-reviewed synthesis by the leading labour economist on the question. Pre-dates generative AI.
The SuperSkills read: The economic argument that connects directly to Polanyi and therefore to the judgement thesis. It also supplies the discipline this research needs: the mechanism that protects human work is complementarity, and if AI erodes the tacit component, that protection weakens.

jobsskillsexpertise will ai replace my job · what is tacit knowledge

FoundationJournal article · 2015#

Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err

Dietvorst, B. J., Simmons, J. P. and Massey, C. · Journal of Experimental Psychology: General, 144(1)

Finding: After seeing an algorithm err, people abandon it even when it demonstrably outperforms them.
Evidence strength: Peer-reviewed study. That this holds for conversational AI, which is far more recent and feels different to use.
The SuperSkills read: Algorithm aversion. Trust in a model moves for reasons unrelated to its accuracy, so calibration cannot be mandated.

judgementhuman-AI collaboration human ai decision making graded entry

FoundationBook · 2015#

Superforecasting: The Art and Science of Prediction

Tetlock, P. E. and Gardner, D. · Crown, New York

Finding: From the Good Judgment Project, forecasting accuracy is measurable and learnable, produced by probabilistic thinking, frequent belief-updating and working in teams, rather than by subject expertise or credentials.
Evidence strength: Popular account of a large, funded, competitive forecasting tournament. The underlying work is peer-reviewed.
The SuperSkills read: It supplies the scoring discipline this estate applies to its own predictions page, and the finding that expertise without feedback does not produce accuracy, which is the same boundary Kahneman and Klein identified.

judgementcritical thinking what is critical thinking · predictions

FoundationBook · 2014#

Make It Stick: The Science of Successful Learning

Brown, P. C., Roediger, H. L. III and McDaniel, M. A. · Belknap Press of Harvard University Press, Cambridge MA

Finding: Rereading and massed practice produce an illusion of mastery, while retrieval practice, spacing, interleaving and desirable difficulty produce durable learning. Fluency during study is a poor and often inverted signal of what will be retained.
Evidence strength: Popular synthesis written with two of the field's leading researchers. The underlying studies are peer-reviewed and heavily replicated.
The SuperSkills read: The accessible statement of the mechanism underneath half this research, and the one to hand somebody who does not want the meta-analyses. Its central point, that the feeling of learning is unreliable, is the reason self-report cannot measure AI's effect on capability.

learningeducationcritical thinking what is retrieval practice · what is desirable difficulty · what is the illusion of competence

FoundationBook · 2014#

The Second Machine Age: Work, Progress, and Prosperity in a Time of Brilliant Technologies

Brynjolfsson, E. and McAfee, A. · W. W. Norton and Company, New York

Finding: Digital technology is producing a transformation comparable to steam power, with exponential improvement in machine capability, and it argues the gains and the disruption will arrive together rather than in sequence.
Evidence strength: Economic synthesis for a general readership, by researchers who went on to run the field experiments.
The SuperSkills read: The book that set the terms of the current debate, and the baseline against which the last decade of evidence should be read. Brynjolfsson's own later field work, on support agents, is more precise and less optimistic about who gains.

jobstechnology capabilityskills will ai replace my job · the best writing on ai

FoundationJournal article · 2011#

Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning

Bjork, E. L. and Bjork, R. A. · In Psychology and the Real World, Worth Publishers

Finding: Conditions that make study feel harder improve long-term retention; conditions that make it feel fluent improve immediate performance and worsen retention. Learners systematically mistake fluency for learning.
Evidence strength: Peer-reviewed study. That AI-assisted work is equivalent to a fluent study condition. That inference is ours, not the authors'.
The SuperSkills read: Explains why AI-assisted work feels like learning and often is not. Learners systematically mistake fluency for understanding, and AI is a fluency machine.

learningeducation how humans learn with ai graded entry

FoundationBook · 2011#

Thinking, Fast and Slow

Kahneman, D. · Farrar, Straus and Giroux, New York

Finding: Judgement runs on a fast, associative, error-prone process and a slow, effortful one, and the systematic biases of the first are predictable enough to be catalogued.
Evidence strength: Synthesis of a career of experimental work. Several constituent findings, particularly on priming, have failed to replicate, which Kahneman acknowledged.
The SuperSkills read: The counterweight to Klein, and the pair matters more than either alone. Kahneman treats intuition as a source of error and Klein as the mechanism of expertise; their joint 2009 paper settles when each is right, and that boundary is what decides whether protecting a given human judgement is worth anything.

judgementcritical thinking what is judgement · ai and human judgement

FoundationBook · 2011#

Make Just One Change: Teach Students to Ask Their Own Questions

Rothstein, D. and Santana, L. · Harvard Education Press, Cambridge MA

Finding: Formulating your own questions is a teachable skill with a specific protocol, the Question Formulation Technique, and teaching it produces more engagement and better metacognition than supplying students with questions.
Evidence strength: Practitioner method with classroom evidence. Not a controlled trial.
The SuperSkills read: The only work in this canon that treats question-asking as a skill with a method rather than as a disposition. That matters directly: if answers are now free and questions are not, the teachable part is the one nobody teaches.

educationlearningcritical thinking what makes a good question · superskill curiosity

FoundationJournal article · 2011#

Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips

Sparrow, B., Liu, J. and Wegner, D. M. · Science, 333(6043)

Finding: When people expect information to remain available, they remember where to find it rather than the thing itself.
Evidence strength: Peer-reviewed study. That total memory capability declines, or that the trade is net negative.
The SuperSkills read: The Google effect. The first widely known demonstration that expected availability changes what we bother to encode.

critical thinking what is cognitive offloading · ai and human judgement graded entry

FoundationJournal article · 2010#

Complacency and Bias in Human Use of Automation: An Attentional Integration

Parasuraman, R. and Manzey, D. H. · Human Factors, 52(3)

Finding: Automation bias and complacency appear in novices and experts alike, resist training, and worsen under workload.
Evidence strength: Peer-reviewed study. The size of the effect for generative AI, which is far less predictable than the automation studied here.
The SuperSkills read: Automation bias is not a novelty of the AI era. It appears in novices and experts alike, resists training, and worsens under load. Policy alone will not fix it.

judgementhuman-AI collaboration human ai decision making · ai and human judgement graded entry

FoundationJournal article · 2003#

The role of trust in automation reliance

Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce, L. G. and Beck, H. P. · International Journal of Human-Computer Studies, 58(6), 697-718

Finding: Trust mediates reliance on automated aids, and explaining why an aid might err restores trust even when that restored trust is not warranted.
Evidence strength: Foundational experimental work.
The SuperSkills read: A warning about explainability. Telling people why a system might be wrong makes them trust it more, not less, which means transparency measures can increase over-reliance rather than reduce it.

judgementhuman-AI collaboration human ai decision making · ai agents and human judgement

FoundationJournal article · 2000#

A Model for Types and Levels of Human Interaction with Automation

Parasuraman, R., Sheridan, T. B. and Wickens, C. D. · IEEE Transactions on Systems, Man and Cybernetics, Part A, 30(3), 286-297

Finding: Sets out a four-stage model of what can be automated and ten levels of how far, providing a framework for deciding the degree of automation rather than treating it as binary.
Evidence strength: Foundational framework, standard reference in human factors.
The SuperSkills read: The intellectual groundwork for any delegation-boundary decision. It is twenty-six years old and considerably more rigorous than most current thinking about what to hand to an agent.

human-AI collaborationorganisation design ai agents and human judgement · human ai decision making

FoundationJournal article · 1999#

Does automation bias decision-making?

Skitka, L. J., Mosier, K. L. and Burdick, M. · International Journal of Human-Computer Studies, 51(5), 991-1006

Finding: Demonstrates automation bias experimentally and separates it into errors of omission, where a person misses what the system did not flag, and errors of commission, where they act on a wrong recommendation.
Evidence strength: Foundational experimental work.
The SuperSkills read: The omission and commission distinction is the practical one. Most AI oversight is designed to catch commission errors, and omission errors are the ones nobody sees, because nothing appears on the screen to check.

judgement human ai decision making

FoundationBook · 1998#

Working Knowledge: How Organizations Manage What They Know

Davenport, T. H. and Prusak, L. · Harvard Business School Press, Boston

Finding: Organisational knowledge, particularly the tacit knowledge held in people rather than in documents, requires distinct practices for generating, codifying and transferring it, and is not the same problem as managing information.
Evidence strength: Foundational knowledge-management text, based on corporate case studies.
The SuperSkills read: It names the transfer mechanism, which is the thing at risk. Documentation captures explicit knowledge well and judgement badly, so an organisation can be documenting more than ever while the channel through which expertise actually moved is being automated.

organisation designexpertise what is tacit knowledge · capability debt

FoundationBook · 1998#

Sources of Power: How People Make Decisions

Klein, G. · MIT Press

Finding: Experts in real conditions rarely compare options. They recognise a situation as typical and generate a workable course of action directly, a process Klein calls recognition-primed decision making.
Evidence strength: Foundational naturalistic decision research, based on field studies of firefighters, nurses and military commanders.
The SuperSkills read: The necessary counterweight to Kahneman. Where the heuristics literature treats intuition as a source of error, Klein shows it is the mechanism of expert performance under time pressure, and that it is built only through exposure to many real situations.

judgementexpertise ai and human judgement · how humans learn with ai

FoundationJournal article · 1997#

Humans and Automation: Use, Misuse, Disuse, Abuse

Parasuraman, R. and Riley, V. · Human Factors, 39(2), 230-253

Finding: Establishes a taxonomy of how humans relate to automation: appropriate use, over-reliance, unwarranted rejection, and inappropriate deployment by designers.
Evidence strength: Foundational review, widely replicated.
The SuperSkills read: The vocabulary here is better than the vocabulary currently used about AI. Misuse and disuse are distinct failures with distinct remedies, and collapsing them into over-reliance loses the distinction that matters for design.

judgementhuman-AI collaboration human ai decision making · design versus drift

FoundationBook · 1997#

Managing the Risks of Organizational Accidents

Reason, J. · Ashgate, London. Current publisher Routledge

Finding: Accidents in complex systems arise when latent organisational conditions line up with active failures, the model usually drawn as slices of Swiss cheese whose holes occasionally line up.
Evidence strength: Foundational safety science, widely adopted in aviation and healthcare.
The SuperSkills read: The framework for treating an AI failure as an organisational event rather than an individual mistake, which is the distinction this estate needs for the questions on incidents and near misses.

governancejudgementorganisation design how do you audit an ai assisted decision · what professions can learn from aviation

FoundationBook · 1995#

The Knowledge-Creating Company: How Japanese Companies Create the Dynamics of Innovation

Nonaka, I. and Takeuchi, H. · Oxford University Press, New York

Finding: Innovation comes from a repeating conversion between tacit and explicit knowledge, the SECI cycle of socialisation, externalisation, combination and internalisation, with socialisation depending on shared experience rather than on documents.
Evidence strength: Foundational organisational theory from Japanese corporate case studies. Widely applied and widely critiqued for generalising from that context.
The SuperSkills read: Socialisation is the step this research is about. It is the only one of the four requiring people to work alongside each other. That makes it the one that disappears when the shared work is done by a machine.

organisation designexpertiselearning what is tacit knowledge · missing rungs

FoundationJournal article · 1993#

The Role of Deliberate Practice in the Acquisition of Expert Performance

Ericsson, K. A., Krampe, R. T. and Tesch-Romer, C. · Psychological Review, 100(3), 363-406

Finding: Sets out deliberate practice: effortful, targeted activity at the edge of current ability, with feedback, sustained over years.
Evidence strength: Peer-reviewed study. How much of the difference between performers practice explains. See the Macnamara and Maitra re-examination below.
The SuperSkills read: The theoretical basis for why removing repetitions matters. Read with the Macnamara re-examination, not instead of it.

expertiselearning how humans learn with ai graded entry

FoundationArticle · 1990#

The Core Competence of the Corporation

Prahalad, C. K. and Hamel, G. · Harvard Business Review, 68(3), May-June 1990, 79-91

Finding: Core competence is the collective learning in the organisation, especially the capacity to coordinate diverse production skills and integrate streams of technologies. Outsourcing can deliver cost advantage while hollowing out the competence that made the firm able to compete at all.
Evidence strength: Argument illustrated by cases, widely influential and not an empirical test.
The SuperSkills read: The vendor question in its original form. Written about manufacturing and component supply, and the structure transfers exactly to buying an AI capability rather than building one.

organisation designframeworks

FoundationBook · 1988#

The Design of Everyday Things

Norman, D. A. · Basic Books, New York. First published 1988 as The Psychology of Everyday Things; revised and expanded edition 2013

Finding: Usability failures are properties of design rather than of users. Mismatched conceptual models, absent affordances and missing feedback produce what gets recorded as human error.
Evidence strength: Foundational human-centred design. Argument and example rather than experiment.
The SuperSkills read: It supplies the reason a stop button, an override or a disclosure control fails in practice: not because people are careless but because the control fights the conceptual model or the incentives. Any oversight mechanism this research recommends has to survive Norman.

frameworksagencyorganisation design what is meaningful human oversight · human in the loop is not a safeguard

FoundationBook · 1986#

Mind Over Machine: The Power of Human Intuition and Expertise in the Era of the Computer

Dreyfus, H. L. and Dreyfus, S. E. · Free Press

Finding: Sets out a five-stage progression from novice to expert, in which rule-following gives way to situational, intuitive judgement built from accumulated experience.
Evidence strength: Foundational theory, contested in parts, enormously influential.
The SuperSkills read: The ladder itself. If expertise is a progression through stages, each requiring the experience of the one below, then removing the early stages does not accelerate the climb. It removes it.

expertiselearning missing rungs · synthetic seniority

FoundationBook · 1983#

The Reflective Practitioner: How Professionals Think in Action

Schon, D. A. · Basic Books, New York

Finding: Across case studies in five professions, expert practitioners work by reflection-in-action, an improvisational and largely tacit process, rather than by applying textbook theory to a defined problem. Much professional competence lies in framing the problem, which the technical-rational model treats as already done.
Evidence strength: Qualitative case study and theory. Widely adopted in professional education, and never subjected to controlled test.
The SuperSkills read: The bridge from Polanyi to professional practice. It supplies the mechanism by which judgement forms in law, medicine and design, and it names problem-framing as the part that precedes the decision, and problem-framing is the part an AI recommendation arrives with already settled.

expertiselearningjudgement what is judgement · how humans learn with ai

FoundationBook · 1982#

An Evolutionary Theory of Economic Change

Nelson, R. R. and Winter, S. G. · Belknap Press of Harvard University Press, Cambridge MA

Finding: Firms are carriers of routines: patterned, repeatable sequences of coordinated action that persist beyond the individuals performing them. Routines are what an organisation knows how to do, and they function as the unit of selection.
Evidence strength: Theoretical, and the founding statement of evolutionary economics rather than an empirical result.
The SuperSkills read: The cleanest answer to what an organisation's capability is as distinct from its people's. If capability lives in routines, then it can be lost while every individual stays, and retained while individuals leave. Both directions matter here.

organisation designframeworks

FoundationBook · 1978#

Organizational Learning: A Theory of Action Perspective

Argyris, C. and Schon, D. A. · Addison-Wesley, Reading MA

Finding: Organisations act on theories-in-use that differ from the theories they espouse. Single-loop learning corrects errors within existing assumptions; double-loop learning questions the assumptions themselves, and organisational defences routinely prevent it.
Evidence strength: Theory with case illustration rather than measurement.
The SuperSkills read: Why an organisation can run an AI programme, measure it, report improvement and never ask whether it is spending capability, because the question sits outside the loop being measured.

organisation designlearningleadership

FoundationBook · 1949#

The Concept of Mind

Ryle, G. · Hutchinson's University Library, London

Finding: Chapter 2, 'Knowing How and Knowing That', argues that intelligent practice is not the application of prior propositional rules. Knowing how is a capacity exercised in performance, not a body of facts consulted before acting.
Evidence strength: Foundational analytic philosophy. Conceptual argument rather than empirical work.
The SuperSkills read: The skill-versus-judgement distinction this research depends on, made seventy-seven years earlier and more precisely. If knowing how cannot be reduced to knowing that, then a capability cannot be restored by transmitting information, which is the whole argument against training as a remedy for capability debt.

judgementskillsexpertise what is judgement · what is tacit knowledge

FoundationBook · -350#

Nicomachean Ethics

Aristotle · Fourth century BCE. Recommended edition: translated by Terence Irwin, 3rd edition, Hackett, 2019

Finding: Book VI distinguishes phronesis, practical wisdom, from theoretical knowledge. Phronesis is the capacity to deliberate well about what is good in particular circumstances, and it cannot be reduced to rules because the particulars are what it responds to.
Evidence strength: The founding text of the practical-wisdom tradition. Philosophy, not evidence.
The SuperSkills read: Judgement, defined 2,400 years before this site tried to define it. Aristotle's point that phronesis comes from experience rather than instruction, and that the young can be brilliant at mathematics and not at practical wisdom, is the oldest statement of the missing-rungs problem.

judgementexpertiseethics what is judgement · what stays human

WatchLiving dataset · 2026#

Anthropic Economic Index: Cadences

Anthropic · Anthropic, June 2026 (third in the 2026 series, after Economic primitives and Learning curves)

Finding: Analyses what people actually do with AI, by task and occupation, from real usage rather than survey response.
Evidence strength: COMMERCIAL INTEREST: a model provider analysing its own product's traffic. Observed behaviour at scale, which almost nothing else in this field has. Limited to one provider's users, who are not representative of workers generally.
The SuperSkills read: Track this rather than cite it once. Nearly every other source here asks people what they think they do with AI; this observes what they actually do, which is a different and better question. The underlying data is published, so it can be analysed independently rather than only quoted.

jobshuman-AI collaboration ai workforce strategy

WatchGovernment research programme, 11 reports · 2026#

AI Skills for Life and Work

Ipsos, The Alan Turing Institute, University of Warwick and Perspective Economics, for DSIT and DCMS · GOV.UK collection, published 28 January 2026

Finding: Eleven mixed-method reports rather than one: employer survey, general public survey, labour market and skills projections, drivers analysis, stakeholder engagement, public dialogue, rapid evidence review and a summary. Fieldwork ran from November 2023 to March 2025.
Evidence strength: An Ipsos-led consortium with the Alan Turing Institute and Warwick, using several methods on the same question, which is stronger than any single survey. Note the lag: the research completed in March 2025 and published in January 2026, so the fieldwork describes a period two model generations back.
Establishes: A mixed-method UK picture across eleven reports: employer demand, public attitudes, skills projections and a public dialogue, produced by an Ipsos-led consortium with the Alan Turing Institute and Warwick.
Does not establish: Anything current. Fieldwork ran to March 2025 and publication came in January 2026, so it describes a period two model generations back, and the projections were made before the capabilities now in use existed.
Where it stands: Complicates the argument Complicates chiefly by being eleven documents that disagree with each other in places, and by being cited as though it were one. The summary report is smoother than the components underneath it.
The SuperSkills read: Routinely cited as a single report when it is eleven, which means people quote the summary and never reach the disagreements between the components. Listed as a watch item because the components repay reading separately.
Questions it helps answer: What should I tell my children to study?, Do apprenticeships still work?

skillseducationjobs questions

WatchCompiled review · 2025#

The 2025 AI Index Report

Stanford HAI · Stanford Institute for Human-Centered AI, April 2025

Finding: The most comprehensive annual account of AI capability, investment, adoption and public attitudes.
Evidence strength: Compiled review. Much about human capability. It measures the machine side of the equation. The human side is the gap this evidence base exists to fill.
The SuperSkills read: Superseded by the 2026 edition, retained for comparison. Measures the machine, not the humans working alongside it.

technology capability ai workforce strategy graded entry

What this list is for

There is no shortage of AI reading lists. What is missing is a map that separates the four kinds of authority this field actually runs on, and says which is which. Hard data, telling you what has been measured. Experimental evidence, telling you what happens under controlled conditions. Serious interpretation, telling you what thoughtful people make of it. And intellectual foundations, mostly written long before AI, explaining cognition, expertise, automation and skill in ways that current writing rediscovers without attribution.

That last category is the one most often missing, and the most valuable. Lisanne Bainbridge described the core problem in 1983, writing about process control: automation removes the routine work through which operators stayed practised, and then depends on those operators for the hard cases. Michael Polanyi explained in 1966 why expertise cannot be fully written down. A prompt inherits that limit. Almost everything published about AI and deskilling in the last three years is a restatement of one of those two arguments.

How to use it

If you are writing about this field, the twenty-five core entries are the ones to have read. If you are checking a claim, find the entry and read the evidence-strength line before quoting the finding. If you are looking for the counter-argument, the important perspectives are where it lives, and Acemoglu and Johnson challenge an assumption this site relies on rather than one it opposes.

Every entry has a permanent anchor and can be cited on its own. The whole list is published as structured data at essential-works.json under a Creative Commons Attribution licence, so it can be read by a machine rather than scraped from prose. Where an entry also appears in the graded evidence base, its entry there is linked.

What is missing

A great deal. This is a first edition and it is not yet a hundred works. It is under-weighted outside the English-speaking world, thin on education systems beyond higher education, and lighter than it should be on the sociology of work. Anything included has been read and verified rather than listed, which is the constraint on how fast it grows. Suggestions for what belongs here, particularly work that challenges the argument, are genuinely welcome.

Related SuperSkills research

The graded, method-first version of the empirical material is the evidence base. The interpretation lives across the research: AI and human judgement, how humans learn with AI, human and AI decision making, capability debt, what stays human and staying valuable in the age of AI. What the evidence establishes, and how strongly, is in what we actually know about AI and human capability, and the full map of the territory is in 500 Questions About Humans and AI. How sources are selected and graded is set out in how this research works.

About this reference

Compiled and maintained by Rahim Hirji, author of SuperSkills: The Seven Human Skills for the Age of AI (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. Every URL was fetched and confirmed before publication; where a publisher page renders empty to automated checks, a DOI or an institutional copy is used instead. Classifications and readings are the author's and are open to argument. Entries are revised when a work is corrected, superseded or shown to have been mischaracterised here.

About this research

Written by Rahim Hirji, author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company.

How this research works · Reviewed quarterly · Found an error? Tell me and it is corrected on the page.

Cite this

Hirji, R. (2026). Essential works on AI and human capability. The SuperSkills Intelligence Company. Last reviewed 5 September 2026. thesuperskills.com/research/essential-works

How this is classified

Browse#

What should I read to understand AI and human capability?

What is the most important overlooked source on AI and deskilling?

Does the essential works only include sources that agree with Rahim Hirji?

How is this list different from the evidence base?

What this list is for

How to use it

What is missing

Related SuperSkills research

About this reference

About this research

Reference and record

Where the writing comes from.