Every research page on this site, grouped. Generated from the estate rather than maintained by hand, so nothing can be published and then be impossible to find.
Start here
- Human capability in the age of AIThe category defined: what it covers, how it differs from AI adoption, literacy, governance and training, the six dimensions it can be measured on, and five stages of organisational practice.
- The SuperSkills thesis: why capability compounds
- What we actually know about AI and human capability
- What I argue, and what the evidence showsTen interpretive positions, each separated from the evidence behind it and each stating what it does not claim. Arguments, not findings.
- 811 Questions About Humans and AI
- The best writing on AI, and what changed
- How this research worksThe method, stated so it can be challenged. How sources are selected and graded, how contradictory evidence is handled allowed to decide anything, how terms are attributed, and what commercial interests exist.
- How to cite this researchHow the stable identifiers work, what the five tier labels mean, how to cite a page, a study or the evidence base as a dataset, and the licence the data carries.
The seven SuperSkills
- What is a SuperSkill? Definition and the four testsWhat a SuperSkill is, the four tests a capability has to pass to be one, and the seven that pass. Durable across technological cycles, transferable, governing how you work with intelligent systems, and compounding.
- Curiosity
- Change Readiness
- Big Picture Thinking
- Empathy
- Global Adaptability
- Principled Innovation
- The Augmented Mindset
Frameworks for using AI well
- AI frameworks comparedSix teaching frameworks set against Bloom, SAMR, CLEAR, TCREI, SIFT, the CRAAP test and the levels of automation literature, with the peer-reviewed critique of SAMR applied to all of them including these. Offered as unique rather than best, and each page names the outside framework that does the job better.
- Think, AI, thinkWrite your own position first, use the model, then decide what to keep. Skipping the last step produces a bad answer somebody eventually catches; skipping the first produces a good one nobody can check. Nearest published relatives are Dell'Acqua's centaurs and cyborgs and Mollick's seven approaches.
- The four levels of AI useExtract, Explore, Examine, Extend: a ladder about the request rather than the tool. Bloom's taxonomy describes the learner and stops discriminating once a model can perform all six categories; SAMR describes the technology and carries a peer-reviewed critique that applies here too.
- Goal, Context, Friction, StandardFour parts to a prompt, and Friction is the reason it exists. CLEAR, TCREI and CO-STAR all ask what the model should be told, and none has a slot for what the person should be made to keep doing, which is the difference between the two arms of the Bastani trial.
- Keep, share, hand overThree columns for AI delegation, defaulting to Keep for anything you cannot classify. Sixty years of levels-of-automation research is better at everything except being usable in four seconds, and Swiss federal guideline four says the same thing in policy: responsibility must not be capable of being delegated to machines.
- The five rungs of AI useChat, Project, Skill, Automation, Agent. The rungs are not evenly spaced: one to three are conveniences, and four removes the person from the moment of execution, which is where oversight quietly stops working. Anthropic draws the same line between workflows and agents.
- The source ruleSearch, Open, Understand, Record, Cite. Written for the failure a model creates, a reference that looks perfect and does not exist. Caulfield's SIFT is the stronger instrument for a web page and this page says so; the amendment is that existence has to be checked before quality.
For students
- How to use AI at university94 per cent of wholly AI-written exam answers went undetected at Reading, and they outscored real students, so "will I get caught" is the weakest argument available. The decision that replaces it, three columns, taken before the work rather than at midnight.
- How to be honest about using AIThe question is not whether you are allowed but whether you could tell your tutor exactly what you did without leaving bits out. At Reading 94 per cent of wholly AI-written submissions went undetected; detectors flag 61.22 per cent of non-native English essays as AI. The HEPI 94 per cent and the 12 per cent answer different questions.
- Using AI when your department does not want you toHistory, English, Law and Philosophy are strict because in those subjects the writing IS the assessment. If it touches the words that get marked, do not; if it touches your understanding, do. Six things that are almost always fine, six where people get caught, and three of those six are done by people who think they are safe.
- How to handle forty readingsThe bottleneck is not reading speed, it is that you read forty things and kept none of them findable. Five free steps, fifteen minutes in week one. Step five, being tested on your own readings, is the best-evidenced study technique in psychology: testing beat restudying 56 to 42 per cent at one week.
- Group work when everyone has AISomebody drops three paragraphs of slop in at midnight and the whole group wears the mark. Five things to agree in week one, and the Procter and Gamble field experiment with 776 professionals showing AI dissolved the functional split that groups usually divide along.
- What actually happens if you get caught using AIExpulsion is rare and it is the wrong thing to picture. What usually happens is a zero or a failed module, settled quietly, and at the 2026/27 English fee cap of 9,790 pounds across six modules that is roughly 1,600 pounds. The cost question survives a falling detection risk where the fear question does not.
- How often does AI invent a source?Two counted floors. One paper in 277 on PubMed cited a study that does not exist in early 2026, up from 1 in 2,828 in 2023. And 2,022 court decisions worldwide record fabricated material, of which 1,163 were filed by people representing themselves rather than by lawyers.
- Running out of messages makes you worseEvery free tier has a limit and what it does to you before you reach it is the part nobody notices. You take the first answer, stop being tested, ask worse questions and hoard the good tool. Scarcity turns a capable user back into a beginner.
The named concepts
- Capability debtWhat happens to skill when the doing is automated: capability debt, the missing rungs, synthetic seniority, the missed reps, and how humans learn with AI.
- Synthetic Seniority
- The Missing Rungs Problem
- The Missed Reps
- Cognitive debt, capability debt, and the restFive competing terms for the same worry, mapped against their primary sources. Who actually claims to have coined what, what the evidence under each really is, and three sources that could not be verified and are named rather than quietly used.
- Drift versus Design
- Human at the Start
- Usage Theatre
- The Verifier's Discount
- The Unclaimed Hour
- Outsourced recognition
- The AI Readiness LieWhy most AI transformation stalls, and what works instead: the readiness chain, transformation as a quarterly practice, and the recurring challenges.
- What is the AI Memo Test?Three questions Hirji published on 11 May 2025 after four chief executives sent AI memos in a month. Seventeen months on, one firm walked its memo back, one cut about 30 per cent of its workforce, and two held position. The test scores a decision, and four near-identical decisions produced four different outcomes.
Judgement, oversight and accountability
- What board oversight of AI actually looks likeWhat the frameworks actually require and what belongs on a quarterly agenda. Article 14 treats the capability to override as the content of oversight rather than the presence of a person. NIST expects reversal thresholds before they are needed. Ayinde (2025) made the verification duty non-delegable and upward-travelling. And the capability question sits in none of the three lines of defence.
- Does AI weaken human judgement?The flagship review. 106 experiments where human and AI pairs did worse than either alone, endoscopists losing six points of unassisted detection, students scoring lower once the tool was removed, and where AI demonstrably improves decisions instead.
- Is human intuition better than AI logic?Kahneman and Klein settled when intuition can be trusted: a learnable environment, and prolonged practice in it with fast clear feedback. AI barely touches the first condition and removes the second, because the practice is the part being automated.
- Should we trust AI over human experts?Who actually gains from AI advice. Across 140 radiologists the effect ran from strongly positive to strongly negative and nothing predicted which; across 5,172 support agents the gains went to the least skilled while the most skilled gained nothing at all.
- What happens when AI and human judgement conflict?What happens when a person and a system disagree. Access to a system raised agreement from 58.4 to 80.9 per cent and cut accuracy from 74.2 to 63.9; only 5 per cent of model answers carry any marker of doubt, and the confidence signal that would justify pushing back never reaches the reader.
- Is AI dangerous?Three questions asked as one. Fabricated references now reach 1 in 277 papers and endoscopists averaging 27.6 years of experience lost six points of unassisted detection, while the extinction estimates move from 5 to 10 per cent on one changed clause.
- Should AI development be paused or slowed?
- What is an AI kill switch, and would one work?
- What do the rogue AI agent incidents mean for an organisation using AI?
- Do AI models know when they are being tested?
- What should a board do about the AI safety warnings?
- Is it ethical to let AI judge people?Four obligations that survive being correct: an explanation the person can use, a contest somebody can win, a named human who carries it, and verification on the population it is used on. A widely deployed sepsis model scored 0.63 against a developer claim of 0.76 to 0.83.
- Can AI be unbiased, or does it reproduce human bias?It reproduces human bias at measurable scale, and unbiased is not one target: several reasonable fairness criteria are provably incompatible. The first bias audit law in the world produced published audits from about 5 per cent of employers.
- Human and AI decision making
- Decision Quality in the AI Era
- AI agents and human judgementWhen AI can plan and act, the human job becomes deciding what to delegate and owning the result. Agents, human at the start, and who is on the hook.
- Should I let an AI agent act on my behalf?Not a trust question. A system that answers gives you something to reject; a system that acts has already done it. Every oversight model assumes a pause that agents remove.
- Human in the loop is not a safeguard
- What is meaningful human oversight?Article 14 of the EU AI Act came into force on 2 August 2026 and names automation bias in the legislation. The five things an overseer must be enabled to do, why most arrangements fail the test, and why it is a capability problem wearing a governance costume.
- Who owns verification when AI does the work?In most organisations, nobody. It is not in a job description, a budget line or an org chart. The capability test that settles whether verification is real, and four ownership models with their actual costs.
- Who supervises work they cannot do themselves?A person three years into a career reviews AI-generated work of a kind they have never produced, and signs it off. The organisation records a control as satisfied. Nothing has been checked.
- When should I override AI?
- How do I know when AI is wrong?
- How do you audit an AI-assisted decision?
- What should a board ask about AI?
- The Delegation Boundary Map
- Who can override an AI system?Oversight without authority is ceremony. What Article 14 actually requires, the one paragraph that names competence, training and authority together, and why overriding costs something while deferring is free.
- Does explaining an AI's reasoning help?NIST separates explanation accuracy from decision accuracy: a true account of how a system reached a wrong answer is an ordinary outcome. Why a reason can substitute for a check, and the three things an explanation is genuinely good for.
- How should AI decision rights be allocated?Doing the work of deciding and holding the right to decide were always different. What the regulation already allocates between provider and deployer, and why an allocation arrived at by drift cannot be reviewed.
- Which decisions should become slower because of AI?Cognitive forcing cut overreliance in a 199-participant experiment, and the designs that worked best were the ones people rated worst. Article 14(5) already mandates a two-person check for one category. Four conditions that justify putting the friction back.
- When agents become part of the workforce, who manages them?Article 26 already names the person: competence, training and authority, plus the power to suspend. The gap sits underneath it. The FAccT visibility paper states we lack methods for determining when an agent has created a sub-agent, which breaks every span-of-control assumption in management.
- How do you design a stop button people will actually use?Article 14(4)(e) requires the button and decides nothing about whether it gets pressed. One ICU study annotated 12,671 arrhythmia alarms and found 88.8 per cent false. A regulator has on record that stop-work authority went unused for fear of reprisal. Five properties that make a stop control real.
- How should an AI agent communicate uncertainty to a human?Two problems, usually confused. GPT-3's verbalised confidence carries an expected calibration error of 0.52, and readers put "very likely" at 62 per cent where the guidelines mean above 90. Giving readers a translation table made them worse. First-person hedging raised accuracy from 63.9 to 72.8 per cent and lowered intention to use.
- Can a human approve an AI decision at machine speed?Only when the machine proposes no faster than a person can check, and most deployments have measured neither. A congressional hearing, a Spanish regulator's first agent-executed breach and a US-China proposal reached the same point in one week; thirty years of automation studies say what happens to a checker who cannot keep up.
Thinking, learning and capability
- Does using AI stop you learning?No, and delegating the work to it does. Four designs that withdraw the assistance before measuring split on the same line: 52 developers at 50 per cent against 67, 1,222 participants losing persistence inside ten minutes, a school arm 17 per cent below students who never had access, and 26,811 students whose entrance-exam penalty arrives in year two.
- AI and critical thinking
- How humans learn with AI
- Using AI without dependency
- Am I becoming dependent on AI?
- Does AI make everyone think alike?
- How do I get AI to challenge me rather than agree with me?
- How do I keep my own voice when using AI?
- Is attention a trainable skill?
- Assessing students when AI can do the assignment
- Does AI detection work?Not well enough to accuse anyone. Detectors flagged more than half of essays by non-native English speakers as AI-generated while classifying US eighth-grade essays almost perfectly. Plus the base-rate arithmetic nobody runs.
- Should children use AI?
- What stays human
- Human skills in the age of AI
- Can you regain a skill you have lost?
- How do you assess capability rather than output?
- What professions can learn from aviation
- Does GPS damage your brain?The London taxi studies measured acquisition: four years of the Knowledge, structural change only in the trainees who qualified. The removal side is one behavioural study with 13 people at follow-up and no scanner. Vivienne Ming's own dated words are milder than the version in circulation.
- What is the human signal?The trace of a mind inside a piece of work, named on 30 November 2025, with three tests a reader applies without being asked: was a real decision made, was something difficult carried with care, is there the risk of a personal point of view. What has been measured is convergence, and how badly people detect the thing they are sure they can feel.
- Is it still my idea if AI helped me write it?The credit question is the easy half. Four randomised experiments show writing assistance moving the writer's own views: 1,506 participants whose opinions followed their tool, and 80 per cent who said the advice had not moved them when it had. Ownership tracks how much of the shape you supplied.
- How good are you at using AI?Five yes-or-no questions and a banded reading at 0-1, 2-3 and 4-5, put to live audiences before it was written down. Question three is the only one you cannot acquire in an afternoon. Unvalidated, and the page says so in a section rather than a footnote.
Work, careers and the labour market
- Who AI leaves behindAI's gains land on the least skilled, in customer support and in taxi driving. The detectors built to police it misclassify non-native English writers at 61.22 per cent. The same people the tool helps are the ones the checking penalises.
- Proving you did the workDetection misclassifies over half of non-native English essays, a 61.22 per cent false positive rate, while labelling a reply as AI removes its advantage even where it beat humans. Why both detection and blanket disclosure fail, and what proof of process looks like instead.
- The mid-career squeezeThe measured displacement is at 22 to 25, through reduced hiring. Mid-career exposure is different: AI's gains land on the least experienced, compressing the gap a mid-career salary pays for. Plus why a randomised trial found experienced developers 19 per cent slower and unable to detect it.
- Will AI replace my job?
- Will AI replace entry-level jobs?
- Which jobs are safest from AI?
- What is the AI employment gap?Employment of 22 to 25 year olds in AI-exposed occupations sits about 19 per cent below where it would have been had it tracked their less-exposed peers. The 19 is the distance between a fall of 11 and a growth of 10, not a fall of 19, and it runs through hiring rather than redundancy.
- Does AI actually make people more productive?Large measured gains on narrow tasks: writing 40 per cent faster, a standardised coding task 55.8 per cent faster, support resolutions up 15 per cent an hour. Scattered or negative in real work. Nothing yet in the national statistics, and the perception gap runs in both directions.
- Will AI replace programmers?No study shows replacement. The two most cited coding experiments sit seventy-five points apart because one built something new and the other changed a system somebody already knew. The measured risk is to how developers are made.
- Should juniors use AI at all?Yes, and the evidence says which version of use is safe. The three experiments that removed the tool afterwards, and the four rules that follow. Plus four for whoever manages them, because the burden is in the wrong place.
- Do apprenticeships still work?Germany has record unplaced applicants and 54,400 empty places in the same year. England shortened the apprenticeship and its flagship replacement drew 74 starts against a target of 1,000. Four systems read at source, and three claims the evidence does not support.
- Should I still learn to code?
- Staying valuable in the age of AIWhere human value moves as AI spreads: staying valuable, the entry-level question, and the human skills that matter most.
- What should I tell my children to study?
- Why "learn to prompt" is weak career advice
- Four Generations of Disruption
- Who captures the productivity gains from AI?
- Which tasks do workers not want automated?Asked task by task about their own occupations, 1,500 US workers were positive about automating 46.1 per cent of them. Where they refused, distrust of accuracy outranked fear of replacement by two to one, and workers wanted more human involvement than experts thought necessary on 47.5 per cent of tasks.
- Is deskilling real, or a rescaling of what counts as skill?Deskilling and rescaling have both been measured, in different people. Nineteen Polish endoscopists lost six percentage points of unassisted detection; a Japanese taxi fleet's skill gap closed by 14 per cent. The strong claim that there is no deskilling fails against a measurement, and the rescaling claim survives only as a question about which tasks were automated.
- How do juniors become senior if AI does the junior work?Junior work was a by-product of senior workload, not a training scheme, so AI removes the reason it existed. Access is not the harm: in randomised trials with 1,222 people the withdrawal effect appeared after ten minutes, and across 26,811 Chinese students the exam losses fell almost entirely on those whose homework time collapsed.
Organisations and leadership
- The shape of the organisation after AIAcemoglu puts ten-year productivity gains under 0.66 per cent, Danish records find precise nulls on pay two years in, and US payroll data rules out widespread displacement. What the evidence supports about redesign and fragility, and why middle management is the layer nobody has counted.
- What AI does to a teamEveryone improves and the room converges. Doshi and Hauser on individual creativity rising while collective diversity falls, Dell'Acqua on AI flattening the difference between an engineer and a marketer, and the sycophancy evidence on what happens when challenge moves from colleagues to a model that affirms half again as often.
- Who should own AI strategy in an organisation?Ownership is decided by inheritance rather than argument, and it silently answers a bigger question: augmentation or replacement. Autor and Thompson on why which tasks you automate matters more than how many, what each placement is blind to, and the role missing from most implementation teams.
- AI workforce strategy
- The CHRO guide to AI
- How leaders should respond to AI
- What does AI literacy mean for leaders?A legal obligation since February 2025, enforced since this month, at every risk tier. What Article 4 requires, the clause almost every programme misses, and why a completed training module is not evidence of capability.
- How do you write an AI use policy that works?Most are unenforceable and everyone involved knows it. Six elements that survive every model release, what EU law now makes auditable, and the one-afternoon test that beats legal review.
- How do you measure AI adoption properly?
- Why reskilling programmes mostly failA position page against the near-universal institutional answer to AI. Five structural reasons, the evidence stated honestly including where it is indirect, and the published data that would change the position.
- AI Transformation Is Not a Change-Management Problem
- Common AI Transformation Challenges
- Should AI attend my meetings?
- The Third Way
- How do you keep expertise in an organisation?Documentation preserves what experts can say, and most of what they know is not that. Polanyi on the tacit part, Lave and Wenger on how it transfers, and why the work juniors learned from is the work most easily automated.
- What happens to institutional memory?Retrieval is not retention. An organisation can get better at finding things while getting worse at knowing them, because the reasons, the rejected options and the trust calibration were never in the archive.
- What happens when AI removes the visible work?Bainbridge, 1983: taking away the easy parts of a task can make the difficult parts harder. Why time saved prices only what was removed, why task productivity and workload are different variables, and what responsible de-automation requires.
- What should we tell employees about AI and headcount?Which decisions a machine may make in the organisation's name, and what will and will not be done with the time it saves, declared by domain and dated. Deloitte's 25,000-worker UK survey found 31 per cent concealing use and 64 per cent of weekly users afraid their manager will decide AI can do their job; the PNAS experiments show the penalty they fear is real. Concealed use is delegation nobody has decided, checked, or could reconstruct.
- What is a Shared Prompt Review?Four things on the table once a week: the prompt as typed, the raw output, what a person cut and why, and where it might be wrong. A field experiment with 776 professionals found AI erasing the difference between what a technical and a commercial specialist proposed. The problem is measured; the remedy is not.
- What is automating versus informating?Zuboff, 1988. Automating replaces human judgement with a machine; informating makes the work more visible to the person doing it. The same model configured two ways, so output can improve while capability erodes with nothing on any dashboard showing the second half.
- What is a capability audit?A proposed method, labelled as one. What people can still do unaided, where expertise actually sits, what has no redundancy and what would be hard to rebuild. Why confidence and output are both broken instruments.
- When should an organisation reverse an AI deployment?NIST's five conditions for deactivating a system, Perrow on why adding a safeguard can reduce safety, and why a fallback that exists only as a document is not a fallback.
- How do you preserve capability across vendors?Prahalad and Hamel's 1990 warning, applied. Keep enough to specify the work, judge it, handle exceptions and replace the supplier. Three things that are genuinely new, and one that is not new at all.
- If every competitor has the same AI, where does the advantage come from?A licence every rival can buy at list price passes one of Barney's four tests. US Census data puts firm use at 18 per cent with 57 per cent of adopters in three or fewer functions, so parity has not arrived, and the assets that pass the test are the ones the invoice does not include.
- How long should we give an AI investment before deciding whether it worked?The J-curve says an early read understates and a late one overstates, for the same reason. Danish administrative records found precise nulls on pay two years in, next to heavy task reorganisation. Three questions, three clocks, and most business cases wind only one.
- What becomes more valuable in a business as AI gets cheaper?Three randomised trials found the same compression: the tool raises the floor far more than the ceiling. So the capability losing value fastest is being unusually good at exactly the work the model does well. Four complements that appreciate, chosen by standing rather than by capability.
- What happens to work whose purpose was moving information around?Garicano's model says a layer exists because matching problems to knowledge is costly. Resume data on 3,100 firms finds hierarchies flattening after AI adoption, on tests the authors call under-powered. What the reporting job was doing that the report never showed.
- Which AI investments should we stop?Between 30 and 40 per cent of information systems projects show some degree of escalation, and the founding case study of that literature was an expert system that ran for a decade. There is no credible non-vendor base rate for AI programme failure: the 80 per cent figure traces to a press article. Three questions that do not need a counterfactual.
- How do humans and agents divide work across a process?Task-level allocation has been on the wall since 1951 and has never worked, because every assignment creates new work for the other party. The unanswered part is the boundary. Medicine calls those discontinuities gaps, the best multi-agent failure data puts 36.9 per cent of failure at the joins, and Article 14 says nothing about transitions.
Professions and sectors
- Which professions face the greatest deskilling risk?
- How will AI change medicine?
- How will AI change law?
- How will AI change consulting?
- How will AI change accounting and audit?The oversight precedent everybody cites was replaced in April 2026, and its replacement puts generative AI expressly out of scope. Meanwhile the UK audit regulator found the six largest firms had not measured what their tools do to audit quality.
- How will AI change journalism?Twenty-two broadcasters, 14 languages, 2,709 graded answers. Sourcing failed at 31 per cent against 20 per cent for accuracy, and the reputational cost arrives at the masthead that was cited rather than the assistant that cited it.
- How will AI change customer service?The best-evidenced occupation there is, and the study watched the tool break. Gains ran to novices, the best agents got slightly worse, and the learning survived an outage only for the workers who had engaged with the suggestions.
- How will AI change teaching?A school-randomised trial in England cut lesson planning time 31 per cent with no quality change a blinded panel could see. English teachers plan with it at 35 per cent and mark with it at 5 per cent, which is the profession quietly putting the tool where the risk is lowest.
- How will AI change the public sector?Both flagship UK figures, 26 minutes a day and 19, are self-reported, and one was calculated from tick-box midpoints with its top tail capped by the analysts. The sector's deepest precedent is not a productivity study, it is the presumption that the computer is right.
- How will AI change human resources?Retrieval models favoured White-associated names in 85.1 per cent of comparisons. The first algorithmic hiring audit law in the world produced published reports from 18 of 391 employers checked. And the function advising on capability is the one most exposed.
Everyday life
- Should AI remember everything about me?
- Is it safe to use AI for therapy or advice?
- Should I use AI to write personal messages?
- Should I let AI summarise everything I read?
- Do I still need to remember things?
- How much should teenagers use AI?
- How do I raise a child who thinks for themselves?The only experiment that took the AI tutor away found unrestricted users 17 per cent below students who never had it, while the guardrailed group was largely spared. Configuration decided it. A rule about screen time cannot reach the thing that matters.
- Should I let AI make personal decisions for me?Advice from a model with no settled view still moves yours, disclosure does not protect you, and 80 per cent believe they would have decided the same way unaided. Three questions that separate a delegable decision from one that stops existing if you delegate it.
- Is screen time the same argument as AI use?No. Self-reported screen time correlates with logged use at r = 0.38, and across 355,358 adolescents technology use explains at most 0.4 per cent of the variance in wellbeing. Przybylski's own group has published the warning against counting hours of AI. The variables that survive are order and substitution.
International
- AI and work, country by country
- AI and work in AsiaSingapore has written the loss of entry-level training grounds into a national AI framework. Hong Kong measures augmented reality adoption and does not ask about AI at all. Japan, Korea and the adoption gap, all read at source.
- AI and work in the GulfSaudi Arabia, the UAE and Qatar read at the issuing body's own pages. The one Gulf instrument that names over-reliance, the region's only official AI adoption statistic, and the widely quoted claims that could not be verified.
- AI and work in JapanThe country with the strongest imaginable economic case for adopting AI, an 11 million worker shortfall by 2040, and adoption running at roughly 18 per cent. The control condition this debate never had.
Definitions
- Human Reserved
- Frontier Firm
- Shallow jobs
- Moral crumple zone
- What is decision provenance?A record of what fed a decision, what was decided and what it set off downstream. Singh, Cobbe and Norval's term from IEEE Access in 2019, a smaller claim than an explanation and a checkable one, and the thing Article 14 of the EU AI Act quietly depends on.
- What is silent failure?A failure that tells nobody: the job runs, the answer reads normally, and nothing marks where it stopped being true. Huang's 2017 definition as differential observability, the 2026 loop in which models caught 12 per cent, repaired none and invented a figure in 12, and why the only detector that fired was a person.
- Oversight readiness
- Instruments that changed
- Capacitating and alienating configurations
- What is AGI?
- Is AI conscious?No system has been shown to be conscious and no agreed test exists that could show it, in machines or in people. Twenty researchers published a method that produces credences rather than verdicts, and Chalmers puts current language models under one in ten while warning against the figure. Consciousness, sentience and understanding are three claims, and only the third changes a decision at work.
- AI leadership
- Rules before tools
- Do you need to be technical to lead AI?
- Is AI-first a strategy or a slogan?
- Is it true that 95 per cent of AI pilots fail?
- AI governance versus AI leadership
- What does an AI-capable manager do differently?
- What counts as a serious AI incident?
- Should AI oversight sit with the full board or a committee?
- Should directors put board papers into AI, and rely on the summary?
- Our AI strategy was written eighteen months ago. What is now wrong with it?
- Should we appoint a director with AI expertise?
- What AI capability should we look for when acquiring a company?
- How does a board know management's claims about AI are true?
- What to do when people work around the AI policy
- What happened to the companies that cut staff for AI?
- What are the ironies of automation?Bainbridge, 1983, five pages. Automating the routine work hands the operator the exceptions and the watching, and removes the practice that built the competence for either. Quotations sourced through the reviewed retrospective that carries them, and the one place where the transfer to AI breaks, which is the place that makes things worse.
- What is the vigilance decrement?Detection of rare signals falls as a watch goes on, and that much has held for seventy-five years. The half-hour figure everybody quotes is the resolution of Mackworth's analysis blocks, restated as a human limit in 1983 and repeated with no citation at all in 2018.
- What is alarm fatigue?2,558,760 alarms in five intensive care units in one month, and 88.8 per cent of the annotated arrhythmia alarms false. Ignoring them is the rational response to that base rate, so routing flagged AI output to a reviewer is a design problem before it is an answer.
- What is algorithm appreciation?Identical advice, two labels, and people took more of it under the machine label. Seven studies, the researchers who predicted the exact opposite, and the seventy forecasting professionals who discounted everything and were less accurate for it. The mirror of algorithm aversion, and the same curve read before the error rather than after it.
- What is knowledge collapse?Twenty-five simulated people choosing between an expensive way to learn and a cheap one, over a hundred rounds. Where the figure of 2.3 times further from the truth comes from, the generational condition it depends on and nobody quotes, why it is the inverse of model collapse, and the second economic sense of the same two words.
- What are use, misuse, disuse and abuse?Four words from 1997 that still separate the operator's failures from the organisation's. Misuse is over-reliance, disuse is neglecting a capable system, and abuse is deploying one without asking what the remaining job will consist of. Only the fourth points at anybody who could have prevented it.
- What is the expertise reversal effect?
- What are core skills?
- Cognitive offloading
- Automation bias
- What is automation complacency?
- What is the out-of-the-loop performance problem?The loss of a person's ability to take over when an automated system fails, named by Endsley and Kiris in 1995. Their participants still saw the data and no longer understood what it meant, and the damage tracked the level of automation rather than their attention.
- What is algorithm aversion?
- What is human-AI collaboration?
- What is the jagged frontier?
- What is an AI hallucination?
- Why does AI sound so confident?
- What is deskilling?
- What is desirable difficulty?
- What is cognitive load?
- What is tacit knowledge?
- How fast do skills decay?From d = -0.01 immediately after training to d = -1.4 after a year of non-use, and cognitive tasks decay faster than physical ones. The meta-analysis, the cockpit study and the intervals aviation regulates to.
- What is judgement?Recognising what a situation is before any option is weighed. Klein, Dreyfus and Polanyi on where it comes from, the distinction from skill and capability, and where Kahneman says it should not be trusted.
- What is critical thinking?Not scepticism but orientation: whether you are defending a position or improving your picture of what is true. Galef, Tetlock and Kahneman, and what changes when arguing either side costs nothing.
- What is metacognition?Knowing what you know, and why the signal people use is the wrong one. Roediger and Karpicke's reversal at one week, Rowland on why feedback nearly doubles the effect, and the group that did not know it was the weaker one.
- What is an organisation's capability?Capability lives in routines, not in the sum of what people can do, so it can be lost while everyone stays, and kept while people leave. Nelson and Winter, Prahalad and Hamel, Teece, and why the dashboard misses it.
- What is intellectual humility?Confidence and accuracy are separate quantities. The four habits that show up in scored forecasting, Galef on why people update when revising costs less than defending, and why changing your mind is not weak judgement.
- What is deliberate practice?Not repetition. The four conditions Ericsson specified, the re-analysis that cut the claim down, Epstein on kind and wicked environments, and the precise question AI raises about which attempts disappear.
- What makes a good question?Investigable, not self-answering, consequential. Rothstein and Santana on question-asking as a teachable method, why a question is not a prompt, and what changed when answers stopped being expensive.
- Over-reliance on AIDepending on a system beyond the point where you could catch it being wrong, which makes frequency of use the wrong diagnostic. In one pre-registered experiment, people with access to a 50 per cent accurate AI scored 63.9 per cent against 74.2 for people with no AI at all.
- CalibrationA system is calibrated when its stated confidence matches how often it is right, which is a different property from accuracy. Expected calibration error for verbalised confidence runs from 0.520 for GPT-3 to 0.180 for GPT-4, and a calibrated model still does not produce a calibrated reader.
- The substitution mythDekker and Woods' term for the assumption that automation swaps a machine for a person and leaves the rest intact. Capitalising on a strength of automation does not replace a human weakness, it creates new ones. Substitution-form business cases are therefore wrong in a known direction.
- Escalation of commitmentCommitting further resources to a failing course of action because you chose it. Staw measured it in 1976; Keil, Mann and Rai put 30 to 40 per cent of IS projects in it; and the four-phase de-escalation model says organisations reach phase one and skip phase two.
- What is the Google effect?
- What is retrieval practice?
- What is productive struggle?
- What is the illusion of competence?
- What is the METR study?Sixteen developers took 19 per cent longer with AI while forecasting a 24 per cent speed-up, and still believed afterwards that they had been faster. METR marked the result out of date on 24 February 2026 and most people quoting it have not noticed. The perception gap is the part nothing has overturned, and METR themselves put it at 40 percentage points.
- What is the judgement premium?PwC's 2026 barometer splits a billion job advertisements in two: roles where AI takes the routine work and leaves the judgement grow at twice the rate with 42 per cent faster advertised salary growth. Autor and Thompson supply the direction test. Nobody has priced judgement itself, so the premium is a direction with no magnitude.
Reference and record
- The most-quoted AI statistics, checked
- The AI reports worth reading
- The official guidance on AI in educationUNESCO, UNICEF, the UK, the EU, Australia, the US and MIT, read at source and labelled by what each document is. Mostly schools: no government here has issued guidance for universities. Five primary-education mandates checked, one of which has already been reversed.
- Does the brain mature at 25?It does not, and the number came from a 2004 magazine interview rather than a finding. Traced to source, checked against a 3,802-person 2025 study that puts the end of the adolescent epoch nearer 32, and reframed on the question that can actually be answered.
- The evidence on AI and human capabilityA curated, dated reference to the best research in the field, from the World Economic Forum, PwC, MIT, the NBER, Harvard and BCG and the key academic studies.
- Essential works on AI and human capabilityEssential works on human capability in the age of AI: hard data, experimental evidence, serious interpretation and the intellectual foundations most current writing rediscovers without attribution. Classified, and read rather than listed.
- The timelineWeekly since January 2017, five years and ten months before ChatGPT. When each idea first surfaced, what form it took, and how it developed. The early years described honestly as curation rather than thesis.
- The predictions recordThe dated provenance record since 2017, plus the reference layer: the annual predictions, the glossary, the people who shape AI, and the reading strategy.
- The SuperSkills Glossary
- AI People
- The AI Reading Strategy
- Accuracy and correctionsA public log of what this research got wrong: the original claim, what was found, what changed. Nothing is removed, including the corrections that weaken an argument made here.