The AI reports worth reading

Most reading lists on AI classify. This one chooses. Below are seven kinds of reader and one named report for each, with the reasoning, the obvious alternative it was picked over, and what it does not cover. Where a document sits in this estate's graded evidence base or classified canon, the entry links straight through to it.

Making a choice is the whole point, and the part everyone avoids. This research already publishes 145 graded studies and 84 classified works. Both sort. Neither answers the question people actually ask, which is what to read given who they are and how little time they have.

Read this before you cite any of it

The policy layer moves faster than the reading lists that describe it, and a great deal of what is currently in circulation is out of date in ways that are easy to miss. Four examples found while assembling this page, all verified against the issuing body on 28 August 2026.

• The UK's AI Safety Institute has been the AI Security Institute since 14 February 2025. Anything using the old name after that date has not been checked.
• Executive Order 14110 was revoked and is still cited as live US policy in reading lists published this year.
• The Department for Education renamed its supplier guidance from "product safety expectations" to product safety standards, and the old URL now redirects.
• DSIT is being dissolved, and GOV.UK now flags the AI Opportunities Action Plan as the work of a previous administration. UK citations need a date and an attribution.

Every entry below carries the date it was last checked. Treat anything on this page older than a quarter as needing a recheck, including this page.

If you are a policymaker or a regulator

The pick, United States: Winning the Race: America's AI Action Plan, The White House, released 23 July 2025. Roughly ninety actions across three pillars: accelerating innovation, building infrastructure, and international diplomacy and security.

Read it with Executive Order 14409 of 2 June 2026, which is the newest instrument and the one most often missing from lists. Its section 3(c) explicitly rules out mandatory licensing or preclearance for AI models, which is the single most consequential sentence for anyone modelling where US regulation is heading.

The pick, United Kingdom: AI Opportunities Action Plan: One Year On, DSIT, 29 January 2026. The original plan of January 2025 is the document everyone cites; this is the one that says what happened. Thirty-eight of fifty actions met, five AI Growth Zones, compute capacity from 2 to 21 ExaFLOPs against a 420 ExaFLOP target for 2030.

Alongside it, the AISI Frontier AI Trends Report, 18 December 2025, the AI Security Institute's first public assessment drawing on two years of testing across thirty frontier models.

What these do not cover. None of them measures what AI does to the people using it. They are strategy and capability documents, and the capability question they answer is the machine's rather than the workforce's.

If you sit on a board or run a risk function

The pick: the NIST AI Risk Management Framework 1.0, January 2023, with its separate Generative AI Profile, NIST AI 600-1, of July 2024. Govern, Map, Measure, Manage is a structure a board recognises, and the common vocabulary most other frameworks are written against.

Why this rather than ISO/IEC 42001. ISO 42001 is certifiable, which makes it attractive to procurement, and certification demonstrates that a management system exists rather than that anything is being managed well. Neither is equivalent to complying with the EU AI Act, and the two are routinely conflated in tenders. The distinction is set out at how to audit an AI-assisted decision.

Read alongside, and almost nobody does: the Information Commissioner's Office's Recruitment rewired, 31 March 2026. After engaging more than thirty employers, the ICO's own conclusion is that many are likely relying on solely automated decisions without meaningful human involvement, which engages Article 22 of the UK GDPR. If you use AI in hiring, this is the most directly consequential document published this year.

Caution. NIST's own page states the framework is being revised under the AI Action Plan. There is no version 2.0 yet, and anyone promising one is guessing.

If you lead HR, learning or workforce strategy

The pick: OECD, Skills in the AI Age. It treats skills as a system question rather than a training-budget question, which is the distinction most workforce strategies fail on.

Why this over the WEF Future of Jobs Report. Future of Jobs is the most quoted document in this field and it is an employer expectations survey, so it measures what executives believe will happen. That is useful and it is not the same as evidence. Its serial forecasts are examined at the most-quoted AI statistics, checked. Read it second, and read it as sentiment.

Also worth your time: the BCG Henderson Institute on companies losing critical thinking when everyone uses AI, which is the closest thing consulting has produced to this estate's own argument, and the PwC Global AI Jobs Barometer for scale, with the caveat that it reads job advertisements rather than outcomes and PwC sells into the market it measures.

What none of them do. Not one measures whether people can still do the work unaided. That gap is the reason capability debt is a framework here rather than a finding.

If you run a school, a college or a university

The pick: Tom Chatfield, AI and the Future of Pedagogy, Sage, 3 November 2025.

It argues from cognitive science and instructional research rather than from institutional anxiety, and it makes the case that AI should be a context for deeper engagement rather than a shortcut. Its sharpest move is to reject defensive, surveillance-based responses in favour of transparent, mastery-based assessment. Five recommendations, of which the most useful is to integrate AI only where it serves a stated pedagogical objective.

This research recommends it partly because it disagrees with the prevailing institutional instinct, and partly because its central claim, that AI must not erode critical thinking, discernment and domain expertise, is the closest external statement of the argument made across this estate.

For practice rather than philosophy: the Department for Education's Generative artificial intelligence in education, updated 12 August 2025 to include Ofsted's approach at inspection, and its support materials, four free staff training modules updated for the 2026-27 academic year. Both apply to England.

For a competency structure: UNESCO's AI competency framework for teachers.

What is missing. The assessment problem is not solved by any of these. What to do when the artefact can be generated is treated separately at assessing students when AI can do the assignment, and the detection answer is worse than most institutions assume at does AI detection work.

If you are a parent

The pick, an unusual one: Generative AI: product safety standards, Department for Education, updated 19 January 2026.

It is written for edtech suppliers rather than for parents, which is exactly why it is worth reading. The January 2026 update added standards on cognitive development, emotional and social development, mental health and manipulation. That is a government department writing down what a product must not do to a child's development. No parenting guide will give you a list that concrete.

For the numbers, read them carefully: Internet Matters, Me, myself and AI on UK children, and Common Sense Media on US teens and AI companions. Both are the best available surveys and both have internal number problems their own press releases do not mention. Those are set out at how much should teenagers use AI, which is written for the teenager rather than about them.

For principles: UNICEF's guidance on AI and children, which is normative rather than empirical and does not pretend otherwise.

If you want the objective one

The pick: the Stanford HAI AI Index, because it is descriptive rather than advocacy.

This is the pick people most want justified, so here is the reasoning. Nearly every widely circulated report in this field is published by an organisation with a commercial or institutional stake in its conclusion: consultancies sell transformation, vendors sell tools, industry bodies lobby, and governments defend a strategy. The AI Index compiles what happened across capability, investment, adoption and policy without an argument to sell. It is long and dry, and the one to quote when you need a figure that will survive scrutiny.

The honest caveat. Descriptive is not neutral. Choosing what to count is an editorial act, and the Index counts what is countable, which under-represents everything happening to people rather than to models.

If you are a sceptic, or a journalist checking a claim

The pick: the METR developer productivity study, July 2025. Sixteen experienced open-source developers, 246 real tasks, randomised. They were measured 19 per cent slower with AI tools permitted, and afterwards still estimated it had made them about 20 per cent faster.

Updated 28 August 2026. METR withdrew this as a signal of the current effect on 24 February 2026. Their second study now estimates a speed-up of 18 per cent for returning developers, confidence interval -38 to +9, and they believe developers are likely faster with AI in 2026 than in 2025. They also say their own data is weak evidence, because 30 to 50 per cent of developers declined to submit tasks they did not want to do without AI. The 19 per cent belongs to early 2025 and is quoted here as a historical measurement. What survives untouched is the perception gap: the same participants estimated a 20 per cent speed-up while being measured slower.

It earns the slot because it is the cleanest available demonstration that self-reported productivity gain can point in the opposite direction to measured productivity gain. Any claim resting on how much time people say AI saves them has to get past this study first.

Then read the pair that shows how far estimates can diverge: Frey and Osborne's 47 per cent against Arntz, Gregory and Zierahn's 9 per cent. Same question, occupations against tasks, a fivefold difference, and neither has been scored against what actually happened.

And for the current employment signal: Canaries in the Coal Mine, which finds no economy-wide displacement alongside a 19 per cent relative decline for 22 to 25 year olds in the most exposed occupations. Both halves are the finding, and most coverage quotes one.

What is not here

Two different kinds of absence, and they should not be confused. First, the deliberate exclusions.

• Vendor research about the vendor's own product. Not because it is worthless, but because it cannot be assessed to the standard applied everywhere else on this site.
• Anything predicting a date. The forecasting record in this field is examined at the predictions record, including the misses.
• Reports whose headline number could not be traced to a stated method. Several well-known ones were considered and dropped for this reason.
• The Gulf, and most of the world. The UAE, Qatar and Saudi Arabia have substantial national AI strategies and do not yet have the measured labour-market studies that Germany, Japan and Denmark do. The wider national picture is at AI and work, country by country.

Then the ordinary kind, which is work not yet done. There is no pick for clinicians, none for model risk management in financial services, and UNESCO's student competency framework and its guidance on generative AI in education have not been assessed here. The EU AI Act runs through this estate as law and has no entry as a document to read. Each will appear once it has been read properly rather than listed, which is the difference between a curated set and a bibliography.

Key research and primary sources

• Chatfield, T. (2025). AI and the Future of Pedagogy. Sage, 3 November 2025.
• The White House (2025). Winning the Race: America's AI Action Plan, and Executive Order 14409, 2 June 2026.
• DSIT (2026). AI Opportunities Action Plan: One Year On, 29 January 2026.
• AI Security Institute (2025). AISI Frontier AI Trends Report, 18 December 2025.
• Department for Education. Generative AI: product safety standards, updated 19 January 2026, and Generative artificial intelligence in education, updated 12 August 2025.
• Information Commissioner's Office (2026). Recruitment rewired, 31 March 2026.
• Graded in this estate: NIST AI RMF, ISO/IEC 42001, METR, Canaries in the Coal Mine, UNICEF, Internet Matters, Common Sense Media.

Related SuperSkills research

The other reference layers, each doing a different job: the graded evidence base ranks, the essential works classifies, the best writing on AI sequences, and the reading strategy tells you how to read them. On checking a number before you quote it, the most-quoted AI statistics, checked. On the national picture, AI and work, country by country. On the method, how this research works.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. Every government and publisher document named here was fetched from the issuing body's own page on 28 August 2026 and its title, date and current status confirmed there rather than from secondary coverage. Documents that could not be confirmed were left out, including one US framework listed by title on a federal site whose release page could not be retrieved. The picks are judgements and are argued for rather than asserted; where a pick is contestable the alternative is named. Reviewed quarterly, and the policy entries more often than that.