Anyone can produce the work

In March 2023 a New York lawyer filed a brief in a routine personal-injury case against an airline. It cited more than half a dozen court decisions, with quotations and docket numbers, and none of them existed. When the airline’s lawyers said they could not find the cases, the court asked for copies, and the lawyer who had written the brief did the most natural thing in the world. He asked ChatGPT whether the cases were real. It assured him that they were, and that they could be found in reputable legal databases. He filed that assurance with the court. In June the judge fined the lawyers five thousand dollars and ordered them to write to every judge whose name had been attached to an opinion that was never written [1].

The fine was small and the lesson was not. The lawyer’s error was not that he used a machine to draft. It was that he asked the machine to verify its own work, and the machine said yes. A generator will always say yes. It has no way to know.

Two years later the same mistake reached a different scale. In July 2025 Deloitte delivered a report to the Australian government, a A$440,000 review of a welfare compliance system, and in August a law academic in Sydney noticed that it cited papers that did not exist and a court judgment with an invented quotation. The firm re-issued the report with a note admitting that a large language model had been used in its preparation, and agreed to repay the final instalment of the fee [2]. That same spring a Chicago newspaper printed a summer reading list in which ten of the fifteen recommended books had never been written; the authors were real, the titles were not, and the list ran in print before anyone checked [3].

A sole practitioner, a Big Four consultancy, a metropolitan daily. Three years, three professions, one failure. The work was produced faster than anyone could check it, and the people who should have checked it had assumed that something which reads as finished is finished.

What the labs conceded

In September 2026 the people building these systems said so themselves. Dario Amodei wrote that the industry “must slow the pace at which we improve the capabilities of AI models,” and committed Anthropic to giving independent evaluators permanent, employee-level access to its systems. Within hours Sam Altman agreed that “we need to pace the frontier” and made the same commitment for OpenAI; Elon Musk posted that “Dario is right” [4]. It was a striking thing for them to say. The makers of the machines were conceding that generation had outrun verification, and that the remedy was human judgment.

Three days later Scientific American asked who those humans would be. One Cambridge academic said he had audited around thirty AI companies himself. Auditors, the piece noted, earn far less than the engineers they would be checking; embedded auditors risk being absorbed by the culture they are meant to judge; what they are allowed to publish will be “a political and commercial question”; and the models behave differently depending on who is testing them, a problem the field “does not really know how to solve.” The verdict from the man who had done the most of it: the kind of auditing Amodei described is “years into the future” [5].

Notice, too, what the proposal covers. Evaluators inside the laboratories, checking the laboratories. Nothing about the essays, briefs, reports, code and applications the machines are producing in the world, and nothing about how anyone is to know which of the people judging that output deserve to be believed. The labs named the problem and left the larger half of it on the table.

The classroom

Start where the problem arrived first, because the classroom is where every artifact is also a credential. In the United Kingdom the Higher Education Policy Institute has surveyed undergraduates each year since the tools appeared. In 2026, 95 percent of students said they used AI in some form and 94 percent used it for assessed work. Most uses are benign: explaining a concept, summarising a reading. But the direction is a straight line, and every instructor knows where it leads.

The scale is visible from the other side as well. Turnitin, whose software is used to check submitted work at thousands of institutions, reported that in the first year of its AI detector it had reviewed more than 200 million papers; about 11 percent contained at least 20 percent AI writing, and about 3 percent, more than six million papers, were at least 80 percent machine-written [7]. Among American teenagers, Pew found in the autumn of 2025 that 54 percent use chatbots for schoolwork, one in ten do all or most of their schoolwork with them, and 59 percent say that cheating with AI has become a regular feature of student life [8].

The students are not confused about what is happening. In May 2025 New York Magazine spent weeks with students and faculty for a piece whose title did not hedge: “Everyone Is Cheating Their Way Through College.” Its central figure was a Columbia undergraduate who estimated that AI had written about 80 percent of his essays, treated assignments as problems to be hacked, and went on to build a tool that fed answers to candidates during coding interviews, demonstrating it on an Amazon interview before the university suspended him [9]. He was not hiding. He had concluded, correctly, that the artifact was no longer evidence of anything, and he was simply acting on it before his institution did.

The institutional reflex was to detect the machine, and detection has failed. OpenAI released a classifier for AI-written text in January 2023 and withdrew it that July for “low accuracy”; it caught about a quarter of machine text and wrongly flagged human writing 9 percent of the time [10]. A Stanford study found that seven popular detectors flagged essays by non-native English speakers as machine-written 61 percent of the time on average, and one of them flagged nearly all of them [11]. The tool built to protect the honest student was most likely to accuse the one writing in a second language.

So a university faces a choice it has not wanted to make. It can keep grading the artifact and accept that the grade proves less each year. Or it can admit that the artifact was never the point.

The workplace

The office caught up fast. Microsoft’s 2024 survey of 31,000 workers in 31 countries found three in four knowledge workers already using generative AI, most of them bringing their own tools without waiting for permission [12]. Gallup found that the share of American employees using AI at work a few times a year or more nearly doubled between 2023 and 2025, from 21 percent to 40 percent [13]. And the adoption is deepest where the output is easiest to generate. By the spring of 2025 Microsoft’s chief executive said that 20 to 30 percent of the code in the company’s repositories was being written by AI [14].

What the people doing the work say about it is the interesting part. Stack Overflow’s 2025 survey of 49,000 developers found 84 percent using or planning to use AI tools. In the same survey, trust in the accuracy of what the tools produce fell to 29 percent, more developers distrusted it than trusted it, and two thirds named “almost right, but not quite” as their most common frustration [15]. Almost right is the expensive kind of wrong. A randomised trial by METR gave experienced open-source developers real tasks in codebases they knew well, with and without AI assistance. With the tools they took 19 percent longer. They had expected to be faster, and afterwards still believed they had been [16]. The time went into reading, checking and correcting output that looked finished and was not.

Outside engineering the phenomenon now has a name. Researchers at Stanford and BetterUp Labs surveyed 1,150 American desk workers and found that 40 percent had received what they call “workslop” in the previous month: AI-generated work that “masquerades as good work, but lacks the substance to meaningfully advance a given task.” Each instance cost its recipient close to two hours to untangle, and 42 percent of recipients came to see the sender as less trustworthy [17]. That last figure is the one to sit with. The cost of generated work is not only the time it wastes. It is the trust it burns between colleagues who can no longer tell, from the output alone, who understood the task.

Managers feel this most. When every report, analysis and proposal arrives fluent and complete, the signals that used to separate the capable from the merely busy have gone quiet. The leader who once judged people by what they produced is now looking at a stack of work that all reads the same, asking who to trust with the next decision.

The market for knowledge

Wherever the cost of producing something fell toward zero, the volume rose faster than anyone’s ability to check it, and the institutions on the receiving end responded the only way they could: by rationing.

In February 2023 Clarkesworld, a science-fiction magazine that had accepted open submissions for seventeen years, closed them. In the first twenty days of the month it had received 700 legitimate stories and 500 machine-written ones, and the second number was growing faster than the first [18]. That September Amazon limited self-publishers to three new books a day, after a flood of generated titles, some of them under the names of living authors [19]. In 2023 the publisher Hindawi, owned by Wiley, retracted more than 8,000 papers, more than the entire scholarly record had retracted in any previous year, after discovering that organised “paper mills” had been selling authorship on manufactured research; Wiley closed journals and put the cost at $35 to $40 million [20]. In February 2024 a peer-reviewed journal published a paper whose figures had been drawn by an image generator, including an anatomically impossible rat labelled with invented words; it was retracted three days later, after the scientists of the internet had done the review the journal had not [21].

Even the chatbot a company puts on its own website now produces claims nobody has checked. In 2022 Air Canada’s chatbot told a grieving passenger that he could apply for a bereavement fare after flying; the airline’s policy said the opposite. When he sued, the airline argued that the chatbot was “a separate legal entity that is responsible for its own actions.” The tribunal disagreed, found the airline had “not taken reasonable care to ensure its chatbot was accurate,” and ordered it to pay [22]. The amount was C$812. The principle was that generated text is a claim, and somebody is responsible for whether it is true.

The hiring funnel

Hiring was built entirely on artifacts: the résumé, the cover letter, the take-home exercise, the interview answer. Every one of them can now be produced. Gartner’s 2025 survey of 3,290 job candidates found that 39 percent had used AI in the application process, mostly to write the résumé, the cover letter and the answers to assessment questions; 6 percent admitted to interview fraud, having someone else pose as them or posing as someone else. The same firm predicts that by 2028 one in four candidate profiles worldwide will be fake [23]. The student who built the interview tool was not an outlier. He was early.

Employers are answering the only way an artifact-based process can: more screening of artifacts. That is the detection mistake again, in a different building. The question an employer actually wants answered is not whether the candidate wrote the cover letter. It is whether the candidate would know if the work in front of them was wrong.

The wider economy

Adoption across the economy is early and uneven, which makes the pattern clearer rather than fainter. The Census Bureau’s survey of businesses put AI use at roughly 18 percent of firms in the spring of 2026, but at 37 percent among firms with 250 or more employees, and at around half in information, professional services and finance [24]. Those are the sectors whose product is knowledge work: writing, analysis, code, advice. They are adopting first because the gains are largest there, and they are where the question of who actually understands the work will bite hardest.

The economics are not complicated. When the supply of something becomes effectively unlimited, its price falls toward zero and the premium moves to whatever remains scarce. The supply of competent-looking text, code and analysis is now effectively unlimited. What remains scarce is the ability to say which of it is right, which of it is safe to act on, and which of the people who produced it would know the difference.

The question left behind

Put the scenes together and a single question falls out of all of them. If anyone can produce the work, who do we believe actually understands it? Who do we admit, hire, promote, fund, cite, or follow? Who carries judgment, now that the doing has been automated?

The old answer was the artifact. The essay stood for the student’s understanding, the brief for the lawyer’s competence, the report for the consultant’s, the paper for the researcher’s rigour, the application for the candidate’s ability. That inference was never perfect, but it was good enough to build institutions on: transcripts, credentials, peer review, performance reviews, hiring. Every one of them assumed that producing the work was evidence of knowing how. That assumption is now false, and the institutions built on it have not yet noticed that their measurements have gone quiet.

What has not been automated is judgment: the ability to look at a piece of work and say, correctly, whether it is good, what is wrong with it, and what should change. The lawyer in New York did not lack a drafting tool. He lacked the judgment to know that a drafting tool cannot be its own check, and the court supplied that judgment for him, expensively and late. Judgment leaves a record, if anyone keeps one. A person who evaluates well is right more often than chance, consistently, in ways that can be checked against what the work turned out to be worth. It can be measured. It simply has not been, because until recently we did not need to.

Verification is not detection

It is worth separating two words that the last three years have run together. Detection asks whether a machine made this. Verification asks whether this is right, and who would know. The first question is unanswerable in principle, as the classifiers proved, and beside the point in practice: the Deloitte report was not wrong because a model helped write it, it was wrong because nobody with the knowledge to catch the invented citations read it before it shipped. The second question is answerable, and it was answerable before the machines arrived. It has only one requirement: that the check come from outside the generator, from people whose judgment has a track record.

That is what the labs were reaching for when they asked for independent evaluators with a record and real access. They were right about the shape of the answer. They were describing it for one industry, for a handful of laboratories, and for a date years away. The rest of the world has the same problem today, at every desk where a document arrives finished and nobody can say whether to believe it.

The tools for generating have been built. The tools for knowing whom to believe have not. Whoever builds them, and proves them where the answer can be checked, will decide how much of what the machines produce can be trusted at all.

References

1. Mata v. Avianca, Inc., S.D.N.Y., June 2023; summary and sanctions order via Wikipedia and LawNext. https://en.wikipedia.org/wiki/Mata_v._Avianca,_Inc. and https://www.lawnext.com/2023/06/court-imposes-sanctions-on-lawyers-who-filed-bogus-cases-after-relying-on-chatgpt-for-legal-research.html
2. City A.M., “Deloitte refunds Australian government after AI ‘made up citations’ in report,” October 2025. https://www.cityam.com/?p=2367690
3. NPR, “How an AI-generated summer reading list got published in major newspapers,” 20 May 2025. https://www.npr.org/2025/05/20/nx-s1-5405022/fake-book-list-ai-chicago-sun-times
4. MediaNama, “Anthropic wants independent evaluators inside frontier AI labs; OpenAI agrees,” September 2026. https://www.medianama.com/2026/09/223-anthropic-evaluators-frontier-ai-labs/
5. Scientific American, “Can independent AI auditors keep up with frontier AI development?,” 17 September 2026. https://www.scientificamerican.com/article/can-independent-ai-auditors-keep-up-with-frontier-ai-development/
6. Higher Education Policy Institute, “Student Generative AI Survey 2026,” Report 199, March 2026. https://www.hepi.ac.uk/reports/student-generative-ai-survey-2026/
7. Turnitin, “Anniversary and number of papers reviewed sheds new light on data around AI writing use in education,” 9 April 2024. https://www.aap.com.au/aapreleases/cision20240409ae83139/
8. Pew Research Center, “How Teens Use and View AI,” 24 February 2026, fieldwork September to October 2025. https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/
9. Walsh, J., “Everyone Is Cheating Their Way Through College,” New York Magazine, 7 May 2025; discussed on CBC Front Burner. https://www.cbc.ca/radio/frontburner/ai-cheating-runs-wild-on-campus-transcript-1.7548309
10. Decrypt, “OpenAI quietly shutters its AI text classifier due to low accuracy,” July 2023. https://decrypt.co/149826/openai-quietly-shutters-its-ai-text-classifier-due-to-low-accuracy
11. Liang, W., et al., “GPT detectors are biased against non-native English writers,” Patterns, 2023; reported in The Markup, August 2023. https://themarkup.org/machine-learning/2023/08/14/ai-detection-tools-falsely-accuse-international-students-of-cheating
12. Microsoft and LinkedIn, “2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part,” 8 May 2024. https://blogs.microsoft.com/blog/2024/05/08/microsoft-and-linkedin-release-the-2024-work-trend-index-on-the-state-of-ai-at-work/
13. Gallup, “AI Use at Work Has Nearly Doubled in Two Years,” June 2025; reported by HR Brew. https://www.hr-brew.com/stories/2025/06/20/new-data-reveals-ai-use-has-doubled-in-the-last-two-years-gallup
14. SiliconAngle, “Satya Nadella says AI now writing 30% of Microsoft’s code,” April 2025. https://siliconangle.com/2025/04/29/satya-nadella-says-ai-now-writing-30-microsofts-code-warns-real-change-still-many-years-away/
15. Stack Overflow, 2025 Developer Survey, as reported by InfoWorld, July 2025. https://infoworld.com/article/4031673/ai-use-among-software-developers-grows-but-trust-remains-an-issue-stack-overflow-survey.html
16. METR, “Measuring the impact of early-2025 AI on experienced open-source developer productivity,” July 2025. https://getdx.com/blog/metr-study-on-how-ai-affects-developer-productivity/
17. Niederhoffer, K., Kellerman, G. R., et al., “AI-Generated ‘Workslop’ Is Destroying Productivity,” Harvard Business Review, September 2025. https://hbr.org/2025/09/ai-generated-workslop-is-destroying-productivity
18. NPR, “A sci-fi magazine has cut off submissions after a flood of AI-generated stories,” 24 February 2023. https://www.npr.org/2023/02/24/1159286436/ai-chatbot-chatgpt-magazine-clarkesworld-artificial-intelligence
19. The Authors Guild, “Amazon Adds to KDP Generative AI Policy, Caps Daily Self-Publishing Uploads,” September 2023. https://authorsguild.org/news/amazon-adds-to-kdp-generative-ai-policy-caps-daily-self-publishing-uploads/
20. Retraction Watch, “Hindawi reveals process for retracting more than 8,000 paper mill articles,” December 2023, and “Wiley to stop using Hindawi name amid $18 million revenue decline.” https://retractionwatch.com/2023/12/06/wiley-to-stop-using-hindawi-name-amid-18-million-revenue-decline
21. Popular Science, “A ridiculous AI-generated rat penis made it into a peer-reviewed journal,” February 2024. https://popsci.com/technology/ai-rat-journal
22. Moffatt v. Air Canada, British Columbia Civil Resolution Tribunal, 14 February 2024; reported by Daily Hive. https://dailyhive.com/canada/air-canada-chatbot-decision
23. Gartner, candidate survey and prediction, July 2025; reported by HR Dive, “Fake job candidates are on the rise.” https://www.hrdive.com/news/fake-job-candidates-ai/757126/
24. United States Census Bureau, Business Trends and Outlook Survey, “Large Firms With at Least 20 Employees Biggest AI Users,” May 2026. https://census.gov/library/stories/2026/05/ai-use-businesses.html