Why keep a ledger

This publication keeps it.

The fear has figures, the frontier has none

Say this for the fear: it is precise. We know what people dread, in what order, and how fast it moves. Around the world, the most common fear is job loss, and it is a fear for others: 64 percent of Americans expect AI to reduce employment over the next twenty years, and 79 percent over the next ten in a 2026 Gallup survey, yet only 16 to 35 percent fear for their own position, depending on the poll.² Since 2025 the questionnaires have started measuring a newer worry, aimed not at work but at ourselves: 53 percent think AI will erode people’s ability to think creatively, 56 percent fear becoming cognitively dependent on it, and in France 44 percent cite the dulling of our faculties, level with the fear of losing jobs.³

The people who build these systems read the same numbers differently. Where 64 percent of the public expects fewer jobs, 39 percent of AI researchers do.⁴ And almost nobody trusts the builders to arbitrate: 15 percent of Americans trust AI companies to decide how the technology is developed and used.⁵

None of this says the fear is misplaced. It says something stranger: the fear is exact, and its object is vague. Every one of these worries points at a frontier, the line between what machines now do and what still asks for a person, and that line is nowhere written down. It moves, someone announces the movement with triumph or dread, the announcement is argued about for a week, and the next movement arrives before anyone has checked the last one. A public that cannot see the line has no way to size its own fear. Everything feels as if it might be taken, because nobody can say what has been.

The alarms and the reassurances share a defect

We have been announcing this frontier for sixty years, in both directions, and the announcements have a record.

Herbert Simon, 1965: “machines will be capable, within twenty years, of doing any work a man can do.”⁶ In 2013, an Oxford study put 47 percent of American employment at risk of automation within a decade or two. When the OECD redid the count task by task rather than job by job, the figure fell to 9 percent, and by 2019 employment in the supposedly threatened occupations had grown 6 percent.⁷ The World Economic Forum asked employers and announced a net loss of 14 million jobs in 2023, then asked again and announced a net gain of 78 million in 2025.⁸ The 300 million jobs in the Goldman Sachs headline, like the 40 percent in the IMF’s, were measures of exposure, of what a model could touch, not of anyone losing anything.⁹ And in 2016 Geoffrey Hinton told a room that it was “just completely obvious” that deep learning would outperform radiologists within five years, and that training them should stop. Ten years later, radiology has thousands of unfilled positions, record salaries, and more images to read than ever.¹⁰ The same pattern repeats at the scale of a single firm: Klarna announced in February 2024 that an assistant was doing the work of 700 people, and by May 2025 its chief executive was conceding “lower quality” and hiring humans back.¹¹

These were not fools. Some of them are among the most capable people of their generation. The defect is structural: a forecast has no maintenance contract. Once made, it is defended, not revised, and the language of defense is always the same: the timing was off, the direction was right, the conditions changed. Philip Tetlock spent twenty years scoring 27,450 expert predictions and found that they barely beat chance, lost to crude extrapolation, and that the experts most often wrong were the least likely to say so.¹² His remedy was not better experts. It was keeping score in public: a question with a date, a number, and a reckoning. Judgment improves only where somebody counts.

Counting also finds shapes that argument misses. After three years of claims and counterclaims about AI and employment, the clearest signal in the data is not mass unemployment, which has not appeared, but a closed door: in the most exposed occupations, employment of 22-to-25-year-olds now runs 19 percent below the trend of their less exposed peers, and the gap comes from hiring that never happens rather than from layoffs.¹³ No editorial predicted that particular shape. Payroll records found it.

What a ledger is

A ledger is the opposite instrument. It is a list of human capacities. Each line states a status, the date the status was reached, a short note on what exactly changed, and the sources that establish it. Three statuses are used. Held: no machine performs the act, because the act belongs to a person by its nature. Contested: machines match or approach the best humans in a narrow or disputed setting, and the question is still argued in good faith. Ceded: on the task as defined, the best machine result exceeds the best human result. A status can be wrong, and that is its virtue: someone can point at the line, produce a source, and the line moves, with the old status kept in the entry’s history.

The discipline is old, and it was invented every time by people who had something at stake. Lloyd’s Register began in 1764 because underwriters needed to know the condition of hulls before insuring them: paid surveyors, a public scale, a new column each year for revisions.¹⁴ The recommended values of the physical constants are revised on a fixed cycle, every four years since 1998. Even a constant, in this discipline, is a state of knowledge with a closing date.¹⁵ The closest model is the IUCN Red List, which has tracked the status of species since 1964: statuses move in both directions, the hooded seal up to Endangered in the same 2025 update that moved the green turtle down to Least Concern, and the method distinguishes a change in the world from a change in our knowledge of it, correcting the past without erasing it.¹⁶ The ledger borrows that rule exactly.

Half of this work already exists for machines, and it is done well. The Stanford AI Index reports each year on scores and their decay, and its recent editions read like a warning against photography: evaluations built to last years now saturate in months, some widely used tests turn out to contain up to 42 percent invalid questions, and the same systems that win gold at the olympiad still fail, in the Index’s words, to reliably tell the time.¹⁷ METR converts machine capability into units of human work, the length of task a system completes half the time, and finds it doubling roughly every seven months since 2019, faster recently, publishing its confidence intervals and its limits with each revision.¹⁸ What none of these instruments records is the other side of the line: not the scores but the positions, who still frames, who still consents, who still answers. And even good trackers die with their hosts, as Papers with Code did in 2025 when its owner lost interest.¹⁹ A registry is also a promise of persistence.

One rule matters more than the rest: the task as defined. Every honest claim of machine parity was born narrow. The 2015 ImageNet result beat a specific human error rate, 5.1 percent, on a specific test set. The 2016 speech milestone was parity with professional transcribers on one telephone corpus, and the 2018 translation claim held for one language pair on one year’s news test.²⁰ The trouble begins when a narrow result is stretched over a whole profession. As Arvind Narayanan and Sayash Kapoor put it, “the easier a task is to measure via benchmarks, the less likely it is to represent the kind of complex, contextual work that defines professional practice.”²¹ So every ceded line in this ledger names its task exactly, and the note says what is not ceded.

What the ledger already shows

Read down the ceded column and a shape appears. Almost every line is a task someone else posed: a game with rules, an exam with answer keys, a benchmark with a scoring function agreed in advance. Chess in 1997, Go in 2016, Protein folding in 2020, IMO mathematics in 2025. Even the Turing test, when it finally fell in 2025, fell as a defined task: five-minute conversations, naive judges, GPT-4.5 taken for the human 73 percent of the time. Change the protocol and the result no longer applies. The definition makes the finding.²²

The contested column is where the century is being decided, and its lines share the opposite shape: the human baseline is fuzzy and the stakes are personal. In a randomized study of clinical reasoning, the model alone scored 92 percent, physicians 74, and physicians given the model 76: the machine did not fail, the coupling did (Medical diagnosis).²³ In structured debates, GPT-4 with a few demographic facts about its opponent out-persuaded humans 64 percent of the time (Persuasion).²⁴ Ordinary readers could not tell machine poems from those of celebrated poets, and rated the machine’s higher (Poetry and prose).²⁵ Each result is real. Each is narrow. Each is argued about by serious people, which is exactly what the status is for.

Then there is the held column, and it is the reason this site exists. It contains no talents. Nothing in it says humans write better, judge better, feel better. It contains positions: Responsibility, Choosing ends, Promising, forgiving, consent, witness, office. Joseph Weizenbaum drew this line in 1976, in the middle of the first AI wave, and it has not moved since: “there is all the difference between deciding and choosing.” Deciding is computation, choosing engages values, and “since we do not now have any ways of making computers wise, we ought not now to give computers tasks that demand wisdom.”²⁶ Hannah Arendt had named the deeper ground twenty years earlier: promising and forgiving, the two remedies for an unpredictable and irreversible world, “depend upon plurality, on the presence and acting of others.” One promises to someone, before others, and can be held to it.²⁷

What makes these lines strong is that they hold by decision, not by lag, and the decisions have dates. The European Parliament floated “electronic persons” for sophisticated robots in 2017, was rebuked by 150 experts, and buried the idea by 2020. The 2024 product liability directive settles the point by routing responsibility for software and AI squarely to persons, natural and legal.²⁸ A federal judge fined the lawyers who filed ChatGPT’s invented cases, writing that existing rules impose “a gatekeeping role” on attorneys, in June 2023.²⁹ A tribunal ordered Air Canada to honour its chatbot’s invented refund policy in February 2024, calling the airline’s suggestion that the bot was “a separate legal entity that is responsible for its own actions” a remarkable submission.³⁰ Asked, each time, whether the machine could be the one who answers, the institutions said no, and wrote it down. A ledger can date the refusals as precisely as the records.

One caution belongs in the same column. A position without control is not a position. It is a fuse. Madeleine Clare Elish gave the failure mode a name, the moral crumple zone: when a system fails, blame collapses onto the nearest human, however little authority they had over it.³¹ Held, in this ledger, means held with the means to answer. The entry for a capacity must say whether the person kept the controls or just the liability.

What delegation reveals

There is a second reason to keep the record, and it is the one the fear points at without naming. Every delegation creates work on the near side of the line. Lisanne Bainbridge wrote it down in 1983, studying industrial automation: “the more advanced a control system is, so the more crucial may be the contribution of the human operator,” who is left with exactly the tasks the designer could not automate, while the skills for them decay unused.³² Forty years later the pattern repeats on schedule. Gastroenterologists who spent months with an AI assistant detected fewer adenomas once it was switched off, 22.4 percent against 28.4 before.³³ The skill did not vanish because the machine was bad. It faded because it was good.

Watch what is appearing instead, because these are the newest lines in the ledger. Verification is becoming a job in itself: in one survey, 41 percent of American employees had received AI-generated work that looked finished and was not, at a cost of nearly two hours of sorting each time.³⁴ Supervision is becoming a job: entry-level postings in exposed occupations are now seven times more likely to demand the judgment once asked of seniors.³⁵ And signature is becoming a job, the oldest one, sharpened: more than 1,600 court decisions have now dealt with invented citations, and European law requires that high-risk systems be overseen by natural persons who understand their limits and can stop them.³⁶ Framing the question, checking what returns, putting a name under the result: none of this was a listed occupation ten years ago, because there was no one to delegate to.

Read this way, the ledger is not an inventory of losses. It is a map with traffic in both directions: capacities crossing to the machines with a date, and positions surfacing on the human side, some ancient and newly visible, some genuinely new. This is what it means to call the site a compass of who we are as a species at this moment: it does not say where to go. It says where we stand, and it shows the last surveyed position each time we ask.

The rules of the house

Every date and figure above carries its source, and that is the standard for every line the ledger will ever hold. A claim that cannot be checked is marked contested, and the note says why. A status, once changed, keeps its past in the open. The tone stays flat on purpose: the frontier moving is neither a catastrophe nor a liberation, it is a fact with a date, and the point of this publication is to keep the dates straight while everyone argues about the feeling. The ledger is identical in English and French. When an essay exists in only one language, the site says so.

How to use it

Start with the held column. It is short, and it is not made of the things people usually list when asked what makes us human. Then read the contested column, where the next decade will be decided line by line. Then, if a line looks wrong, write: the address is at the bottom of every page, and a correction with a source is the most useful thing a reader can send. The record improves the way it was built, one dated entry at a time.

Notes

1. Pew Research Center, “How Americans View AI and Its Impact on People and Society,” September 17, 2025; fieldwork June 9 to 15, 2025, 5,023 adults. 50 percent more concerned than excited, against 37 percent in 2021. ↩

2. Twenty-year expectation: Pew Research Center, “How the US Public and AI Experts View Artificial Intelligence,” April 3, 2025. Ten-year figure: Bentley University and Gallup, July 28, 2026, 3,270 adults. Own-position fears: Ipsos AI Monitor 2025 (16 percent expect their own job to get worse, 30 countries); YouGov, June 2025 (35 percent of US workers). ↩

3. Creativity: Pew, September 2025 (53 percent). Cognitive dependency: Anthropic Public Record with YouGov, 51,993 US respondents, November to December 2025 (56 percent; the sponsor is an AI laboratory). France: Ipsos for CESI, January 2025 (44 percent). ↩

4. Pew Research Center, April 3, 2025: 64 percent of the public versus 39 percent of surveyed AI experts expect fewer jobs over twenty years. ↩

5. Anthropic Public Record, 2025: 15 percent of Americans trust AI companies to make decisions about how AI is developed and used. ↩

6. Herbert A. Simon, The Shape of Automation for Men and Management, 1965. ↩

7. Carl Benedikt Frey and Michael A. Osborne, “The Future of Employment,” Oxford Martin School, 2013 (47 percent). Melanie Arntz, Terry Gregory and Ulrich Zierahn, OECD Working Paper 189, 2016 (9 percent, task-based). Alexandre Georgieff and Anna Milanez, OECD Working Paper 255, 2021: employment in high-risk occupations grew 6 percent from 2012 to 2019, against 18 percent elsewhere. ↩

8. World Economic Forum, Future of Jobs Report, April 2023 (net minus 14 million by 2027) and January 2025 (net plus 78 million by 2030). Both are employer surveys. ↩

9. Goldman Sachs, Briggs and Kodnani, March 26, 2023: some 300 million full-time jobs exposed. IMF staff analysis presented by Kristalina Georgieva, January 14, 2024: 40 percent of jobs worldwide exposed, 60 percent in advanced economies. ↩

10. Hinton in 2016: “It’s just completely obvious that within five years deep learning is going to do better than radiologists” (quoted in The New Republic, October 25, 2024). In 2026: 4,333 active radiologist postings in March, about 130 days to fill a position, average salary 571,000 dollars, volumes up 25 percent since 2018 (Fortune, May 4, 2026). Hinton has acknowledged speaking “too broadly.” ↩

11. Klarna, February 2024 (assistant “doing the work of 700 agents”); its chief executive to Bloomberg, May 8, 2025: “what you end up having is lower quality,” with human agents hired back. ↩

12. Philip E. Tetlock, Expert Political Judgment (Princeton, 2005): 284 experts, roughly 27,450 probability judgments, 1988 to 2003. The follow-up tournaments are in Tetlock and Gardner, Superforecasting (2015). ↩

13. Erik Brynjolfsson, Bharat Chandar and Ruyu Chen, “Canaries in the Coal Mine,” Stanford Digital Economy Lab, revision of August 12, 2026, on ADP payroll data: no economy-wide displacement, but employment of 22-to-25-year-olds in the most exposed occupations 19 percent below the trend of less exposed peers, driven by reduced hiring. ↩

14. Lloyd’s Register, first Register of Ships, 1764: hull condition rated by vowels, masts and rigging by letters, revised annually by paid surveyors. ↩

15. CODATA Task Group on Fundamental Constants: adjustments in 1973 and 1986, then every four years since 1998; the 2022 adjustment closes on data available through December 31, 2022. ↩

16. IUCN Red List, October 10, 2025 update: 172,620 species assessed, hooded seal moved from Vulnerable to Endangered, green turtle from Endangered to Least Concern. On separating real change from knowledge change: Butchart et al., “Improvements to the Red List Index,” PLoS ONE, 2007. ↩

17. Stanford HAI, AI Index 2025 and 2026: SWE-bench from 4.4 to 71.7 percent in one year; new evaluations saturating within months; error rates up to 42 percent of invalid questions on widely used tests; “jagged intelligence.” ↩

18. METR, “Measuring AI Ability to Complete Long Software Tasks,” March 2025, and Time Horizon 1.1, January 29, 2026: 50 percent time horizon doubling roughly every 7 months since 2019, about every 4 months since 2024, with published confidence intervals and stated limits (software-heavy task set). ↩

19. Papers with Code was shut down and redirected to Hugging Face in July 2025. ↩

20. He, Zhang, Ren and Sun, February 2015 (4.94 percent top-5 error against a measured human 5.1 percent on ImageNet); Xiong et al., “Achieving Human Parity in Conversational Speech Recognition,” 2016 (Switchboard corpus); Hassan et al., “Achieving Human Parity on Automatic Chinese to English News Translation,” 2018 (WMT 2017 news test). ↩

21. Arvind Narayanan and Sayash Kapoor, “AI as Normal Technology,” Knight First Amendment Institute, April 15, 2025. See also their AI Snake Oil (Princeton, 2024). ↩

22. Cameron R. Jones and Benjamin K. Bergen, “Large Language Models Pass the Turing Test,” March 31, 2025: three-party, five-minute conversations; GPT-4.5 judged human 73 percent of the time. ↩

23. Goh et al., JAMA Network Open, October 28, 2024: 50 physicians, written cases; model alone 92 percent on a diagnostic reasoning rubric, physicians with conventional resources 74, physicians with the model 76. ↩

24. Salvi et al., Nature Human Behaviour, 2025: preregistered study, 900 US participants; GPT-4 with minimal demographic information more persuasive than the human debater in 64 percent of comparable pairings. ↩

25. Porter and Machery, Scientific Reports, 2024: non-expert readers distinguished AI poems from those of ten well-known poets at 46.6 percent accuracy, below chance, and rated the AI poems higher on several qualities. ↩

26. Joseph Weizenbaum, Computer Power and Human Reason (1976), pp. 227 and 259. ↩

27. Hannah Arendt, The Human Condition (1958), pp. 236 to 247. ↩

28. European Parliament resolution of February 16, 2017 (electronic personhood explored); open letter of some 150 experts, April 2018; resolution of October 20, 2020 (liability routed to persons); Directive (EU) 2024/2853 on liability for defective products, which treats software, including AI systems, as products. ↩

29. Mata v. Avianca, Inc., S.D.N.Y., opinion of Judge P. Kevin Castel, June 22, 2023: 5,000 dollar sanction; “existing rules impose a gatekeeping role on attorneys to ensure the accuracy of their filings.” ↩

30. Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, 2024 BCCRT 149, February 14, 2024. ↩

31. Madeleine Clare Elish, “Moral Crumple Zones: Cautionary Tales in Human-Robot Interaction,” Engaging Science, Technology, and Society, 2019. ↩

32. Lisanne Bainbridge, “Ironies of Automation,” Automatica 19(6), 1983. ↩

33. Budzyń et al., The Lancet Gastroenterology and Hepatology, August 2025: adenoma detection without assistance fell from 28.4 to 22.4 percent in four Polish centres after months of AI-assisted practice. ↩

34. BetterUp Labs and Stanford Social Media Lab, Harvard Business Review, September 2025: 41 percent of 1,150 surveyed employees had received such work in the past month, at an average 1 hour 56 minutes of handling time each. ↩

35. PwC Global AI Jobs Barometer 2026, on 2.4 million US postings: exposed entry-level roles are seven times more likely to require senior-level judgment and oversight skills than in 2019. ↩

36. Tally of court decisions involving fabricated citations: database kept by Damien Charlotin (HEC Paris), over 1,600 by mid-2026. Human oversight: Regulation (EU) 2024/1689, Article 14. ↩