The best writing on AI, and what changed

On 26 August 2026, Bill Gates published a 5,700-word essay arguing that the transition to the AI era will be one of the most turbulent periods in human history, and proposing that some jobs be designated "Human Reserved", protected by agreement rather than by economics. Buried in it was a sentence that has almost nothing to do with economics: he doubted he would have put in the same work as a young man if he had had an AI companion available.

Three and a half years earlier, in March 2023, the most-quoted sentence in the field was that around 80 per cent of US workers could have at least 10 per cent of their tasks affected by large language models. Same subject. Completely different question.

This is a chronology rather than a reading list, because the order is the argument. What follows is the writing that actually moved the conversation between those two points, what each piece got right, and the thing the arc reveals that no individual piece does.

2023 · The numbers that framed everything#

The story does not begin in 2024. It begins within months of ChatGPT, and almost everything since has been an argument with four documents published in a single spring.

Eloundou, Manning, Mishkin and Rock, "GPTs are GPTs" (March 2023). The paper that supplied the field its founding statistic: around 80 per cent of US workers could have at least 10 per cent of their tasks affected, and about 19 per cent could see at least half affected. It travelled further than almost any economics paper of the decade, usually stripped of the word could and of the fact that it measures exposure, not displacement. If you only read one thing from 2023, read this one and notice how carefully it hedges.

Goldman Sachs on 300 million jobs (March 2023). The single most repeated early figure, and still cited in Goldman's own 2026 labour analysis. Its durability is instructive: a large round number attached to a reputable institution outlives every caveat attached to it.

McKinsey on 2.6 to 4.4 trillion dollars of annual value (June 2023). The headline upside number, and the mirror image of the Goldman figure. Between them these two established the shape of the entire subsequent debate: an enormous quantity of jobs on one side, an enormous quantity of value on the other, and remarkably little about what happens to the people in between.

Why this matters here: all three are exposure and value estimates. None of them is a measurement of what happened. Three years later, the estimates are still quoted more often than the measurements that now exist.

2024 · Abundance against bubble#

With the numbers established, 2024 became an argument about whether they would ever arrive.

Leopold Aschenbrenner, "Situational Awareness" (June 2024). An acceleration thesis that became far more than an essay: it seeded an investment worldview and, subsequently, a very large fund. Read it as a document of how a technical argument became a financial position.

Dario Amodei, "Machines of Loving Grace" (October 2024) and Sam Altman, "The Intelligence Age" (September 2024). The two most articulate frontier-builder cases for radical upside. Both are worth reading precisely because they are written by people with the strongest possible interest in being right, which does not make them wrong but does make them evidence of a position rather than of a fact.

David Cahn, Sequoia, "AI's $600B Question" (June 2024). The counterweight from inside venture capital. Cahn worked from Nvidia's run-rate revenue, doubled it for total data-centre cost of ownership, doubled again for end-user gross margin, and asked where the revenue to justify it was going to come from. It travelled through technical and investor audiences harder than almost anything else that year.

Goldman Sachs, "Gen AI: Too Much Spend, Too Little Benefit?" (June 2024). Jim Covello's argument that the technology is exceptionally expensive and is not designed to solve the problems that would justify the cost. This mattered because serious finance had started to question the assumption that capability automatically becomes economic value.

Why this matters here: 2024 is the year the argument was conducted almost entirely in capital expenditure and revenue. Human capability appears in none of these documents.

2025 · Measurement arrives. It is awkward#

The most important year, and the least discussed, because measurement is less shareable than prediction.

"AI 2027" (2025). The year measurement arrived is also the year the most vivid acceleration scenario yet was published, and the scenario travelled considerably further. A month-by-month fictional timeline, specific enough to feel like forecasting and unfalsifiable enough to survive contact with events. Its influence is real and its epistemic status should be clear: a scenario, and scenarios persuade through vividness rather than through evidence. Holding it beside the METR trial below is the fastest way to see the whole problem with this field.

METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity" (July 2025). A randomised trial: 16 experienced developers, 246 real tasks, randomly assigned to permit or prohibit AI tools. Developers forecast AI would make them 24 per cent faster. They were measured as 19 per cent slower. Afterwards, having experienced the slowdown, they still estimated it had sped them up by about 20 per cent.

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.

The sample is small and the population specific, and it should not be generalised to all software work. But the perception gap is the finding, and it should worry any organisation measuring AI benefit by asking people whether it helped. Most are.

Anthropic Economic Index (2025 onwards). Actual usage data rather than forecasts, which makes it a different category of document from everything in 2023. Read alongside the exposure estimates it is a useful corrective: what people do with these systems is narrower and stranger than what they could do.

PwC Global AI Jobs Barometer (2025 and 2026). Close to a billion job adverts across six continents, and pointing the opposite way to the doom: a 56 per cent wage premium for AI skills, more than double the previous year; jobs still growing in the most exposed occupations; and skills requirements changing 66 per cent faster in the most exposed jobs. It belongs beside the pessimistic evidence rather than instead of it. The 2026 edition takes this further, describing a labour market splitting into two paths and rewarding human skills.

World Economic Forum, Future of Jobs Report (January 2025). A survey of 1,000 companies across 22 industries and 55 economies: 170 million roles created by 2030, 92 million displaced, a net gain of 78 million, and nearly two-fifths of current skills obsolete within five years. It is the most-cited institutional jobs number in circulation. It is also employer expectation rather than measurement, which is a different kind of evidence from the payroll data below and is almost never described as such.

Ethan Mollick on the jagged frontier. A concept, not a paper, and the reason it belongs here is that it escaped academia. Executives now use the phrase in meetings. That is a rarer achievement than a good study.

Harvard Business Review, "How People Are Really Using Gen AI". Included for what it reveals about the discourse rather than its rigour: an infographic of self-reported use cases that circulated further than most peer-reviewed work published that year.

2026 · It gets visceral#

The year the argument stopped being about capital expenditure and started being about people, and the year the gap between the evidence and the noise became impossible to ignore.

Matt Shumer, "Something Big Is Happening" (February 2026). Seen more than 80 million times on X. It argued that AI coding and agent tools would displace lawyers and wealth managers, and that everyone should practise using AI for an hour a day. It provoked a Cato Institute rebuttal and a Forbes piece calling it a manifesto. Shumer subsequently told CNBC it was not meant to scare people and that he would have rewritten parts had he known how far it would travel. That coda is the most interesting thing about it.

Citrini Research, "The 2028 Global Intelligence Crisis" (22 February 2026). A 7,175-word scenario projecting 10.2 per cent unemployment and a 38 per cent S&P 500 drawdown by June 2028, through a self-reinforcing displacement spiral. Roughly 16 million views, amplified by Michael Burry, and the major indices opened sharply lower. A thought experiment that moved markets.

And then the correction. Citadel Securities pointed to Indeed data showing demand for software engineers up 11 per cent year on year in early 2026, and Austan Goolsbee of the Chicago Fed said plainly that it is simply too early for the data to show AI eating into jobs. The viral piece was directionally opposed to the contemporaneous hiring data, and it still moved prices. That is the clearest single illustration of the first finding on this page.

Stanford Digital Economy Lab, "Canaries in the Coal Mine?" (updated August 2026). The most important labour-market document currently available, and it says two things at once. There is no widespread economy-wide displacement. And employment among 22 to 25 year olds in highly AI-exposed occupations now sits about 19 per cent below where it would be had it tracked similarly aged workers in less-exposed occupations.

Two further details are routinely dropped when it is quoted, and both matter enormously. The divergence runs through reduced hiring rather than increased separations, which is a slower and quieter mechanism than firing. And the declines concentrate in occupations where AI substitutes for human tasks; where it complements, employment is flat or rising, especially for experienced workers.

Dario Amodei on the entry-level "bloodbath". The figure of up to 50 per cent of entry-level white-collar jobs became the story, largely detached from the interview it came from. Worth reading against the Stanford data, which is measuring the same population and finding something real but considerably narrower.

Leo XIV, Magnifica Humanitas (15 May 2026). An encyclical on safeguarding the human person in the time of artificial intelligence: 245 paragraphs, five chapters, signed on the 135th anniversary of Rerum Novarum. It is the only document in this chronology addressed to a global rather than an Anglophone audience, and it reaches the deskilling argument by a route that touches none of the evidence above. Paragraph 100 holds that the ease of obtaining ready-made answers can weaken personal creativity and judgment. Paragraph 150, quoting the Vatican's 2025 note Antiqua et Nova, states that current approaches to technology can paradoxically de-skill workers, subject them to automated surveillance and relegate them to rigid and repetitive tasks. Paragraph 156 says it is not enough to react only when jobs disappear.

The passage worth reading twice is 140, on what the speed of an answer does to the appetite for a question:

This is a fundamental issue because every technology shapes those who use it. Educating people about the use of AI, then, involves teaching them to decide when and for what purpose it ought not to be used. The speed and ease with which answers or summaries can be obtained risk extinguishing the desire to ask questions, which is a process that bears fruit only over time.

Two notes on how to read it. It is a moral and doctrinal argument and it measures nothing, so it cannot be cited as evidence that AI de-skills anyone; the deskilling sentence is itself a quotation from an earlier Vatican note rather than from a study. And paragraph 106, holding that a slower pace in adopting AI does not mean opposing progress, is the same argument as which decisions should become slower, arrived at from a completely different tradition.

Bill Gates, "The turbulent AI era is here. The choices we make now are critical" (26 August 2026). Three risks: permanent job losses, empowered bad actors, and damage to children's development and human relationships. The proposal of Human Reserved occupations, childcare and jury service among them, plus taxes on AI tokens and robots. The remark about the AI companion he is glad he never had is the thesis of this entire page in one sentence from someone with no reason to make it.

The counterweight, running throughout#

Arvind Narayanan and Sayash Kapoor, "AI as Normal Technology". The most serious intellectual objection to the acceleration frame, arguing that diffusion is slow, institutions are the bottleneck, and the appropriate historical comparison is electricity rather than a new species. Its authors have called it the most influential thing they have written, and it drew direct engagement from the New York Times, the Economist and the New Yorker.

Oliver Burkeman belongs in this list for a reason unlike any other entry. He is not forecasting unemployment. He is arguing for deliberate non-use, and for protecting what is distinctively human because it is worth protecting rather than because it is economically defensible. In 2023 that would have looked like a marginal position. In 2026 it looks like the front of the argument.

Set the encyclical beside him. Burkeman argues for deliberate non-use on secular grounds, from attention and finitude. Leo XIV argues for it from Catholic social teaching, and lands on the same instruction: learn to decide when the tool ought not to be used. Two traditions with nothing methodological in common, converging on restraint, is a stronger signal than either reached alone.

What the arc shows#

Read in order, three things become visible that no single piece contains.

Estimates have outlived measurements. The 2023 exposure figures are still quoted more often than the 2025 and 2026 measurements that now exist, including measurements that complicate them. A field that keeps citing its founding forecasts over its subsequent evidence is not learning at the speed it thinks it is.

The noise and the signal have decoupled. A scenario piece moved markets in February 2026 while contemporaneous hiring data pointed the other way. Meanwhile the genuinely alarming finding, a 19 per cent employment gap for young workers, arrived by payroll data and generated a fraction of the attention. Alarm tracks narrative quality rather than evidence.

The question changed underneath everyone. In 2023 it was about GDP, tasks and headcount. By 2026, the most-read pieces are about children's development, human relationships, what we should agree to keep human, and whether people can still do things unaided. Gates, Burkeman, Mollick and the deskilling literature are now closer to each other than any of them are to the 2023 forecasts.

What this chronology is missing, and why that matters#

Read the entries again and notice who is speaking. With one exception, every piece here is American or written for an Anglophone audience. The exception is the encyclical, which is addressed to the whole world and was published in several languages at once, and it took a papal document to break the pattern. There is still nothing from Japan, where the demographic position makes the labour question structurally different. Nothing from India, where the services-export exposure is the whole argument. Nothing from Germany, where works councils have been negotiating this in enterprise agreements while the Anglosphere wrote essays about it. Nothing from China, Brazil, the Gulf or anywhere in Africa.

That is an accurate description of what travels in English rather than an oversight in the selection, which is a different thing from what is being thought. A chronology of the AI discourse assembled this way is a chronology of the American AI discourse, and the fact that it reads as the global one is itself worth noticing.

The same is true of who gets amplified. The most-shared pieces here are overwhelmingly by men, while the most rigorous entry on the list, the exposure study that started everything, has a woman as first author and is usually cited without any author named at all. Reach and rigour are selecting different people.

Why the migration matters#

The migration in that third point is the whole reason this research exists. The economic question, how many jobs, was always going to be answered slowly and ambiguously, and after three and a half years it still has not been answered. The capability question, what does living with these systems do to what people can do, was barely asked until recently and is now arriving from every direction at once.

The Stanford nuance is where the two meet. It is the most under-quoted finding in the field: the damage concentrates where AI substitutes and not where it complements, and it runs through hiring rather than firing. Which makes it a story about organisations choosing substitution over complementarity, one requisition at a time, without anyone announcing a decision, rather than a story about a technology destroying jobs. Which is drift rather than design, and the reason the missing entry-level roles show up as missing rungs long before they show up as unemployment.

One honest note about my own position. I have argued for years that AI's effect on human capability matters more than its effect on the economy, so a chronology showing the discourse arriving at that conclusion is a chronology that flatters me. Treat it accordingly. The dates and figures here are checkable, and the interpretation is mine.

What is missing from all of it#

Nothing in this chronology measures the thing that matters most: what happens to a person's unaided capability after sustained use, once the tool is taken away. Almost every study measures performance with AI against performance without it. Almost none measures performance after.

Corrected 2 September 2026. This page previously said the exceptions were rare enough to name, and named two, both of which still stand: the Bastani field experiment found students who had used an unrestricted interface scored 17 per cent lower once access was withdrawn, and the Budzyń study found endoscopists' unassisted adenoma detection fell after AI exposure. Those remain the two peer-reviewed anchors. What is no longer true is that they are alone. A deliberate sweep of the literature found at least eight studies measuring unaided performance after AI assistance, most of them published in 2026, and the two strongest are larger than anything this page had.

Strömberg, Lei and Wu (CEPR Discussion Paper 21577, 2 June 2026) is the one that matters most. Thirty months of panel data on 26,811 Chinese students in grades 7 to 12, with monthly closed-book exams and entrance exams as the outcome, which makes the measurement unaided by construction. AI adoption raised homework scores 18 per cent and cut completion time 30 per cent, and lowered monthly exam scores 20 per cent within six months. Entrance-exam scores fell 18 and 24 per cent, with the full penalty emerging only after about two years. The losses concentrate among the roughly 80 per cent of users whose behaviour looks like homework outsourcing; those who kept working at their usual pace were largely spared.

Liu, Christian, Dumbalska, Bakker and Dubey (arXiv, April 2026) supplies the causal version: randomised trials, N = 1,222, AI withdrawn before measurement. People performed significantly worse without it and were more likely to give up, and the authors report the effect emerging after roughly ten minutes of interaction. Their reading is that the loss is of persistence rather than of knowledge, because the tool conditions people to expect an immediate answer.

Three things follow, and the third is the one to keep. The gap is real and now replicated across students, developers, novice programmers and professionals. The damage tracks delegation rather than tool presence, in every study that looked: users who ask for explanations retain, users who ask for answers do not. And the reason the literature looked contradictory is invigilation. Studies whose post-test was unproctored find AI helps; studies whose post-test was proctored find it hurts. One 2026 analysis of 3.2 million learning interactions reports the same estimator producing a 25 per cent decline in odds of a correct answer on proctored items and a large opposite-signed increase on unproctored ones. Most of the apparent disagreement in this field is a measurement artefact.

The honest caveat on all of it: almost every study named here is a preprint or a working paper, none is peer reviewed, and the largest is not randomised. What has actually strengthened is a narrower claim than the headlines carry: assisted output and unaided capability come apart, and the gap grows with the horizon you measure over.

That is the gap. Until it is filled, everything written about AI and human capability, including everything on this site, rests on inference from adjacent evidence. So this page ends by saying what it does not know. See what we actually know about AI and human capability.

Related SuperSkills research#

The classified reading list, by role in the field, is the essential works. The graded studies are in the evidence base. On the entry-level evidence specifically, will AI replace entry-level jobs. On the noise, neither hype nor doom. On the people behind the work, AI people. See AI and work in Japan.

Key sources

• Brynjolfsson, E., Chandar, B. and Chen, R. (2026). Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab.
• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.
• Eloundou, T. et al. (2023). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702).
• PwC (2025). The Fearless Future: Global AI Jobs Barometer, and the 2026 edition.
• Cahn, D. (2024). AI's $600B Question. Sequoia Capital.
• Leo XIV (2026). Magnifica Humanitas: Encyclical Letter on Safeguarding the Human Person in the Time of Artificial Intelligence. The Holy See, 15 May 2026. Graded in the evidence base as institutional modelling, which is to say a position rather than a measurement.
• Gates, B. (2026). The turbulent AI era is here. The choices we make now are critical. Gates Notes, 26 August 2026. Reported by CNBC.
• On the Citrini scenario and the market reaction: Citadel Securities' response.
• On the Shumer post: CNBC interview and the Cato Institute rebuttal.

About this page#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. Every date, figure and view count here has been checked against a primary or reported source and linked. Where a piece is included for its reach rather than its rigour, the page says so. Inclusion is not endorsement, and several entries here are things I think are wrong. Reviewed quarterly, and this one will date faster than anything else on the site.

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