Evidence and the wider picture: questions about AI

AnsweredWhat do we actually know about AI and human capability?#

Nineteen claims in three bands: strong evidence, emerging evidence, and what remains unknown. Reviewed quarterly.

AnsweredWhere is the evidence on AI and human capability?#

Fifty-six graded studies, each with its method, its finding, and what it does not support. Every entry separately citable.

AnsweredWhich books and papers on AI and human capability actually matter?#

Eighty-four works, classified by role in the field rather than by how much they agree with anything here.

AnsweredWere past predictions about AI and work correct?#

Dated positions, kept whether or not they held up. Removing the misses would defeat the purpose.

AnsweredWhen did each idea about AI and human capability first appear?#

Weekly since January 2017, five years and ten months before ChatGPT. Curation first, thesis later, described plainly.

AnsweredWho is researching AI and human capability?#

The people whose work this research draws on, and the reading that goes with them.

AnsweredWhat do the key terms about AI and human capability mean?#

The terms used here, including which are established concepts and which are coinages from this work.

AnsweredWhat is the SuperSkills argument about AI and human capability?#

The whole case in one place, with the parts that are contested marked as contested.

AnsweredHow strong is the evidence that AI weakens human capability?#

Weaker than most commentary implies in some places and stronger in others. The bands are there so the difference is visible.

AnsweredWhat does the evidence on AI and human capability not show?#

Every entry carries a what-it-does-not-support field, which is the reason the base exists.

AnsweredWhat is happening outside the US and UK?#

The first country page is Japan: the strongest economic case for adoption anywhere, and adoption running at roughly 18 per cent.

Better answered elsewhere Global coverage, including the low and middle income countries that almost every exposure study omits. · ILO work on generative AI and jobs

AnsweredCan capability loss from AI actually be measured?#

Three times now, and only by removing the tool: six percentage points of endoscopist detection, 77 per cent against 39 on a no-AI task, 17 per cent below a control once access was withdrawn. No validated organisational index exists.

AnsweredWhat should I read about AI and human capability?#

A four-quadrant shortlist for a leadership team, chosen to close gaps rather than to cover the field. Reading is only useful here if it produces shared language.

AnsweredWhich AI reports are worth reading?#

One named report for each kind of reader, with the reasoning and the limits. Choosing is the value; listing is the easy half.

AnsweredAre the most quoted AI statistics true?#

Ten numbers repeated constantly, each traced to what its source actually says. Four are misquoted, two cannot be traced at all, and three trackers give ChatGPT market shares from under half to over three quarters for the same spring.

AnsweredDoes automation always cause deskilling?#

No. Robots in Japanese nursing homes raised employment, improved retention and cut physical restraint and pressure ulcers. The condition was an acute labour shortage, so the machine filled vacancies rather than replacing people.

AnsweredDoes the EU AI Act cover AI in schools?#

Yes. Annex III paragraph 3 makes AI that decides admission, evaluates learning outcomes or monitors test behaviour high-risk, with the full obligations applying from 2 August 2026.

AnsweredIs official guidance on AI in education based on evidence?#

Mostly not. Of thirteen documents read at source, three present any original data. The UK Department for Education says on the face of its own policy that it has limited evidence.

Better answered elsewhere A research programme with direct access to teaching populations, which this research reviews from the outside. · MIT AI and Education work

AnsweredDo US states regulate AI in schools?#

Three published guidance and only one has statutory force. California says compliance is not mandatory; Ohio requires every district to adopt a policy by 1 July 2026 under Revised Code 3301.24.

AnsweredWhat is the lag between losing practice and losing performance?#

Substantial degradation within the first year on the CPR evidence, six percentage points in months for endoscopists using AI, and d = -1.4 beyond 365 days of disuse. Nobody has measured it for professional judgement.

AnsweredCan capability be measured without testing people unaided?#

Not directly. Confidence fails because the reporting faculty is the impaired one; output fails because assisted output is what stays high.

AnsweredDo most people expect AI to cost jobs?#

In 34 of the 37 countries Pew surveyed in 2026, majorities expect AI to mean fewer jobs, and the expectation is strongest in the richest countries. It is expectation, not outcome: the national evidence on the same page mostly finds small effects concentrated on the young and the educated.

AnsweredWhat do world leaders say about AI?#

Almost all of them say AI must stay under human control and serve people, and almost none says which decisions or who can stop it. The register holds 125 verified quotations from 121 people outside the AI industry, sliced by kind of leader and by region; they divide on jobs, creativity and risk.

AnsweredWhat do presidents and prime ministers say about AI?#

The countries that build the models talk about winning, the countries that buy them about being left out or dominated, and London, Rome, Washington, New Delhi and the UN all ask for human control. Thirteen heads of government, read at the transcript.

AnsweredWhat has King Charles said about AI?#

At Dumfries House on 17 September 2026 he asked 'Surely, we need sufficient means of control before it is all too late?' and told the developers to keep the technology in the service of humanity, community and the natural world. Seven royals in all, with the Dutch and Jordanian monarchs and the rulers of Dubai.

AnsweredWhat has the Pope said about AI?#

Francis told the G7 that human dignity depends on a space for proper human control over the choices of AI; Leo XIV wrote that technology is never neutral because it takes on the characteristics of those who devise, finance, regulate and use it. Ten religious leaders and bodies, read at the Holy See, Hansard and the outlet.

AnsweredWhat do central bankers say about AI and jobs?#

They split: Georgieva's tsunami and Macklem's net loss against Bailey's warning about oversimplified conclusions and Brynjolfsson's tasks-not-jobs. All say skills decide the outcome. Eighteen voices, every figure the speaker's own.

AnsweredWhat do commentators and analysts say about AI?#

The sharpest disagreement in the register: Galloway and Evans expect jobs to change and return, Harari calls it alien intelligence, Webb describes what you give up when you hand decisions to an agent, Haidt and Chiang describe students handing over the thinking. Seventeen voices.

AnsweredWhat do artists, actors and musicians say about AI?#

Three complaints, one question: voices cloned without consent, work taken or priced away, and a machine that has been nowhere and endured nothing. All three reduce to who decided. Sixteen voices, with Ishiguro and Grimes as the dissenters.

AnsweredWhat do other chief executives say about AI and jobs?#

They forecast other people's jobs, from Farley's half of white-collar workers to Dimon's deployment outrunning adaptation, and say nothing about which decisions their own machines now make. Seventeen chief executives, chairs and investors outside the technology industry, read at the letter or the report.

AnsweredWhat do trade unions say about AI?#

A seat and a share: workers must be in the room when tasks go to the machine, and the gains negotiated rather than taken as layoffs. Eight voices from the unions, the ILO and the Kenyan data workers behind the machine.

AnsweredWhat is the best writing on AI?#

A chronology from March 2023 to August 2026, because the order is the argument.

AnsweredWhat is the method behind the SuperSkills research?#

Source selection, evidence tiers, where AI is used and where it is not allowed to decide, attribution rules, and what commercial interests exist.

AnsweredWhat has the SuperSkills research got wrong?#

A public log with the original claim, what was found, and what changed. Nothing is removed.

AnsweredWhat is happening with AI and work in Asia?#

Singapore names deskilling in national guidance. Hong Kong has no AI strategy and no AI statistic. Japan and Korea show adoption far below the discourse.

AnsweredDoes any government address AI deskilling in policy?#

Singapore does. IMDA's agentic framework names the loss of entry-level training grounds and requires training so people retain foundational skills.

Better answered elsewhere Section 2.4.3 is the actual text, short enough to read in full rather than take on this research's summary. · Model AI Governance Framework for Agentic AI

AnsweredWhat is happening with AI and work in the Gulf?#

Of the Gulf instruments reviewed, Saudi Arabia's is the one that names over-reliance on AI. None of the three requires that a human overseer be able to do the work.

AnsweredDo Gulf AI frameworks address deskilling?#

No. The word appears in no verified document from Saudi Arabia, the UAE or Qatar.

AnsweredWhat is happening with AI and work in Japan?#

The strongest economic case for adoption anywhere, and adoption at roughly 18 per cent. The control condition this debate never had.

AnsweredWhat is time horizon (METR)?#

The 50%-task-completion time horizon is the length of task, measured by how long a human expert takes to do it, that an AI model completes with about 50 per cent success.

AnsweredWhat is the capability-reliability gap?#

The capability-reliability gap is the distance between what a model can do once and what it can do dependably. It is the main barrier to agents automating real work.

AnsweredWhat is the jagged technological frontier?#

The irregular boundary between tasks an AI system performs well and tasks it performs badly, where the two can be almost indistinguishable in apparent difficulty and the system gives no signal of having crossed from one to the other.

AnsweredWhat did the METR study actually find?#

Sixteen developers were 19 per cent slower with AI while forecasting a 24 per cent speed-up, and METR marked the result out of date on 24 February 2026. The perception gap survives the withdrawal; the number does not.

AnsweredHow long a task can AI complete on its own?#

The measured answer carries a success rate with it and is usually quoted without one. METR put frontier models near fifty minutes at 50 per cent success, doubling every 207 days, and the same authors put the 80 per cent horizon four to six times shorter on a suite they scored at 3.2 out of 16 for messiness.

AnsweredIs the 19 per cent AI slowdown figure still valid?#

No, and METR say so on their own page. It measures early-2025 tooling in one setting. Quote the 40 percentage point gap between what developers felt and what the clock showed instead.

AnsweredCan AI stand in for real people in a survey?#

Not yet. Pew Research Center reported on 30 September 2026 that AI respondents modelled on its own panel members differed from the real answers by an average of 12 percentage points across nearly 300 questions, and by more than 15 points on about 28 per cent of them. A simulated respondent is a model's guess about people who were not asked.

AnsweredHow accurate are synthetic survey respondents?#

In Pew's test of September 2026, off by 12 points on average, with larger errors for some groups: 16.1 points for Republicans and 15.1 for Black adults. On nearly half the questions at least one answer was chosen by no AI respondent, and two models erred in opposite directions, one describing a more extreme public and one a more moderate one. One organisation's test on US political questions.

AnsweredShould a company use AI personas instead of asking customers?#

Not as the finding. Pew's errors of September 2026 fall where research earns its cost: rare views disappear, people appear better informed than they are, and some groups are described worst. A simulated respondent can help draft and test a questionnaire; a figure about what customers or employees think should say whether people were asked.

AnsweredDoes Pew Research Center use AI in its polls?#

Not to generate opinion. Its statement of 28 September 2026 says humans decide what it studies and write and review its reports, and that it does not use AI to create or model synthetic public opinion; it uses AI for code, for sorting open-ended answers and in the first stage of copy editing. The page treats the statement as a model of an AI policy written by task.

AnsweredWho verifies machine-generated scientific claims, and can people keep up with the volume?#

Nobody and no, on Niam Yaraghi's account at Brookings of 29 October 2026: after OpenAI released 722 machine-generated manuscripts, three withdrawn after a sign error, he wrote that absorption does not scale and that there is no institution for verifying machine-generated scientific claims outside pure mathematics. The page holds the checking as the job; his proposals, funded verification teams, are proposals.

PartialAre third-party AI safety evaluations trustworthy just because evaluators got access?#

No, on Elham Tabassi's argument at Brookings on 24 September 2026: access is not evidence, re-evaluations of the same benchmarks produce different scores and rankings, and the fix is minimum reporting standards, interlaboratory comparison and shared infrastructure. The evaluator page holds the independence question; the measurement question is open.

PartialHow does this research compare with the WEF Future of Jobs report?#

Partly. The chronology sets the WEF numbers in context as employer expectation rather than measurement, which is almost never said.

Better answered elsewhere The employer-expectation series itself, which is broader across countries and sectors than anything here. · The Future of Jobs Report

PartialWhat evidence would change your mind about AI and capability?#

Stated per claim on the state of the evidence, and maintained as a standing register rather than offered when challenged.

PartialCan an organisation improve AI productivity while losing the capability it needs to survive without AI?#

The central tension of this research, and yes: measured in endoscopists whose assisted performance rose while their unassisted detection fell six percentage points.

PartialHow would you know whether AI caused the capability loss?#

Only by removing the tool. Every measurement here that found a loss found it that way, and no observational design has separated it from ordinary decline.

PartialWhat would falsify the capability debt argument?#

Unassisted performance holding steady under sustained AI use, or recovery on removal that is fast and complete. Stated per claim rather than offered when challenged.

PartialHow much of the AI and work evidence comes from high-income English-speaking countries?#

Most of it. The research now carries verified national data from Japan, Korea, Singapore, France, Germany and the Gulf partly to correct for that.

PartialWhich AI studies measure output but not learning?#

Most of them. That single gap explains more contradictory headlines than anything else on this map.

PartialWhy is the AI debate so much louder than the evidence?#

A viral scenario moved markets in February 2026 while hiring data pointed the other way. Alarm is tracking narrative quality, not evidence.

PartialWhat can AI still not do reliably?#

The boundary is jagged rather than smooth, and invisible from the output.

PartialHow much of the workforce is actually exposed to AI?#

Partly answered, and the honest summary is that the estimates are wide and the definitions differ. The UK government's January 2026 assessment splits the workforce into 35 per cent high exposure with high complementarity, 32 per cent high exposure with low complementarity and 33 per cent low exposure, and says plainly that ex ante measures overstate the adoption actually observed.

PartialWho else is mapping this territory, and what do they cover?#

Partly answered. The neighbouring work divides into research projects answering a defined question, institutional thought leadership, short board question sets and government evidence assessments. Several are better than this research on their own ground, and where they are, the map now links out to them rather than paraphrasing.

PartialShould a human verify data an AI has extracted?#

One unreviewed benchmark of 16 models on synthetic web pages (Earn an Honest Dollar, 27 September 2026) found that the instruction 'do not guess' cut invented fields from 70.7 per cent to 20.2 per cent, and that a cheap second-model check caught 77.6 per cent of the remainder. A single run by the tool's own maker; the page holds the general answer, which is that the output cannot tell you.

PartialIs AI chatbot use bad for teenagers' mental health?#

A review in Current Pediatrics Reports summarised by RAND on 28 September 2026 found associations between generative AI use and depression, anxiety and other difficulties in adolescents, and says the evidence 'remains predominantly cross-sectional, limiting causal inference': distress may lead to use as much as the reverse. The teenagers page holds the practical answer while the longitudinal studies are missing.

PartialHow reliable are studies that say AI can do a share of all jobs?#

Less than their headlines. Lyu and Thompson reported on 1 October 2026 that 33 configurations of AI judge, set against 45,796 worker ratings, put between 3.0 and 97.9 per cent of responses as acceptable where workers in the occupation said 61.1 per cent. Agreement on ranking did not carry over to the totals. A preprint, read as an abstract.

PartialHow reliable is a forecast that AI will weaken judgement in 30 per cent of organisations by 2030?#

It is Gartner's projection, reported at second hand on 12 October 2026, from a survey of 533 IT leaders and the firm's own model; no method is published and the report sells advice on the risk it forecasts. The page records it as the first dated forecast of capability debt and nothing more.

PartialWhy do AI insiders and the public disagree about how fast AI will take over work?#

Anthropic's survey of 10,980 US adults and 1,334 of its own staff, published 16 October 2026, found both groups agree on what AI can do today and differ on pace: staff expect AI to do 75 per cent of their work hours by 2030 against 28 per cent for US workers. The page holds the figures in circulation; neither group's expectation is evidence of what will happen.

PartialIf 90 per cent of working hours are not covered by any benchmark, how should an employer judge whether a model is good at the work?#

On the organisation's own work. Mercor's benchmark map, published in October 2026, finds existing benchmarks cover tasks in 208 of 923 occupations and that '90% of American working hours are spent on tasks with little or no coverage'. The page holds the test for a piece of work; Mercor sells evaluation and calls its results provisional.

OpenHow does this research compare with the OECD's work on skills?#

Open. The OECD's adult skills work measures capability as a stock at a point in time. This research asks what sustained AI use does to that stock, which is a different question and has not been reconciled with theirs.

Better answered elsewhere Cross-country skills measurement built on survey instruments this research has no equivalent of. · OECD work on skills and AI

OpenHow do you tell AI-related deskilling from ordinary organisational decline?#

Open, and unanswered anywhere. Without it, capability debt is a story that fits any struggling organisation.

OpenWhat would a credible control group for AI adoption look like?#

Open. Japan is the closest thing available, at roughly 18 per cent adoption in an economy with the strongest case for it.

OpenHow often are AI and work studies independently replicated?#

Open, and rarely. Several of the most quoted findings rest on one design that nobody has repeated.

OpenWhat evidence is missing because nobody is funding it?#

Open. Longitudinal capability measurement has no obvious funder: vendors will not pay for it and employers do not want the answer on record.

OpenHow do we regulate AI decision-making?#

Open here. The EU AI Act tiers obligations by risk and the enforcement questions are only now arriving, so for the next two years nobody can say with confidence how it will be applied in practice.

OpenWhat is happening with AI and work in India?#

Services-export exposure is the whole argument, and roughly five million graduates enter the labour market each year. Next in the international cluster.

OpenWhat is happening with AI and work in Germany?#

Works councils have been negotiating this in enterprise agreements while the Anglosphere wrote essays about it. Next in the international cluster.

OpenWhat obligations do companies have to preserve human capability?#

Nobody is asking this and it is the natural end point of the whole argument.

OpenShould governments care about AI-driven deskilling?#

Policy territory, largely unoccupied. Open.

OpenWhere is this research weaker than the alternatives?#

Open, and deliberately listed rather than answered in passing. The honest short version: no primary task-level data, no cross-country survey instrument, no access to a teaching population, and no legal force. What this research has instead is a maintained architecture of the questions, which none of them keeps.

OpenCan AI benchmark scores be trusted?#

Stanford HAI reported on 25 September 2026 an analysis of 56 AI benchmarks by Koyejo and Truong in which benchmarks claiming to measure the same thing often disagreed and some bias benchmarks measured reading comprehension: 'We're calibrating these instruments against each other and never against reality.' A summary of work posted for review. For a board relying on a score, validation of the benchmark against an outcome it cares about would settle it.

OpenCan AI check published research for errors?#

Schwartz, Andrews and Shapiro (NBER Working Paper 35782, September 2026) ran a language-model workflow over 4,452 replication packages from five economics journals and report that it flagged discrepancies in 3,460 articles or appendices; the abstract does not say how many matter. What would settle it is a human audit of a sample of the flags.

OpenWhy do people keep using AI tools they do not trust?#

Viana and colleagues interviewed 36 graduate students (arXiv 2609.30699, 25 September 2026) who say the tools degrade their writing, thinking and skills and keep using them; the authors describe a cycle of failure, extra checking and justification, and call continued use 'compliance sustained by invisible labor'. A qualitative preprint. A measure of the checking time against the time saved would test it.

OpenShould governments tax automation, subsidise it or stay neutral?#

The IMF's managing director wrote on 15 October 2026 that tax should be calibrated so it neither favours nor penalises automation, alongside help for displaced workers. No government has published such a calibration; the question turns on what the labour data show, which the IMF says it is too early to read.

OpenShould companies that take AI subsidies report on public wellbeing, including skill loss?#

Sandra Wachter and colleagues at the Oxford Internet Institute argue so in Royal Society Open Science, as The Register reported on 21 October 2026, naming competence and autonomy as things AI can undermine. A proposal, not a rule; a jurisdiction that adopted it would be the first test.

OpenCan a company's study of its own workforce be trusted as evidence for other employers?#

Anthropic's report of 16 October 2026 describes its own staff, graded by its own models, with adoption it says is not random, and the authors and Fortune both caution that the company may not predict other workplaces. A replication at an employer that does not build AI would settle it.

OpenHas the singularity begun, and what should an employer do about it?#

Jack Clark of Anthropic told the IMF on 17 October 2026 that he believes 'this year we have gained convincing evidence that it has now begun'; it is a lab leader's claim built on his company's benchmarks and output. An employer's answer turns on evidence of work changing, which the rest of this map tracks; a measured change in what organisations do would settle it, not a declaration.

OpenWho gets credit for an AI-generated proof, and what norms should govern AI contributions to research?#

Being argued now. TechCrunch reported on 27 October 2026 that two labs published refutations of the same hypothesis hours apart, that one lab's mathematician suggested the other had built on prior work, and that three computer scientists published an open letter on norms for AI contributions; MIT's Ryan Williams described parts of one proof as routine and two or three parts as alien-looking. No journal or society has published a rule.

OpenShould a legislature be told when the state authorises an autonomous system to act?#

No legislature does. A preprint of 23 October 2026 reports that no domestic legislature has reviewed a state's authorisation of an autonomous cyber capability before use, argues that attribution is necessary but not sufficient for democratic accountability, and proposes notification within 72 hours, noting that at most one of six oversight bodies could lawfully receive it today. It is an unreviewed proposal.

OpenWho evaluates AI for children and older adults, and is internal evaluation ever enough?#

A preprint of 26 October 2026 traces a decade of harms to testing against an implicit adult user model, argues the gap cannot be closed through internal evaluations alone, and proposes a lifespan vulnerability matrix with three tiers of external evaluation. No regulator requires lifespan-calibrated testing, and the paper is unreviewed.