- When an AI agent proposes changes, how should a team decide which to keep?
- Why do developers use AI all day but refuse to trust it for important decisions?
- If AI diagnoses better than a doctor, what is the doctor still for?
- Does experience or expertise protect against deferring to AI?
- If AI reviews are better one by one, why would science get more homogenous?
- Who decides what counts as good science if the reviewer is a model from a lab?
- Which parts of research are scientists handing to AI, and which are they keeping?
- If a model's remaining weaknesses are judgement, taste and self-assessment, which human skills become more valuable?
- How much time do finance teams spend checking AI?
- What happens to quality when AI produces work faster than people can review it?
- Is engineering becoming more about judgement than code?
- Why does AI speed up tasks but not whole organisations?
- Who verifies machine-generated scientific claims, and can people keep up with the volume?
- Why do the same models score above 99 per cent on old benchmarks and 26 per cent on real user requests, and what should an employer test before buying?
- If coding agents meet most requirements but fully finish one task in four, who is responsible for the rest and how should teams staff for it?
- What do clients pay an accountant for once AI does the return?
- If my lawyers use AI, should my legal bill be lower?
- Does using AI change what counts as reasonable billable hours?
- What is the verification-value paradox?
- Who is responsible when a junior lawyer uses an approved AI tool and it is wrong?
- What does a law firm partner do when AI takes the delegation and checking work?
- Does a benchmark showing AI beats accountants mean we should automate month-end?
- How should a firm design a pilot so it can tell where review adds value?
- Would an AI auditor stand up to a client CFO better than a human junior would?
- Does reviewing the intent of a change rather than every line still count as oversight?
- Do people absorb confident falsehoods from AI regardless of how wrong it is, and does their trust adjust?
- Should journals replace human peer reviewers with AI?
- If AI finds and patches hundreds of vulnerabilities, who reviews the patches?
- What is a checkpoint in AI-accelerated research and who sets it?
- If AI can produce research faster than people can review it, how should reviewing time be allocated?
- Can peer review keep up with AI-generated papers?
- Who reviews AI-written code, and can they keep up?
- Who is responsible for AI-assisted work when there is no firm standard for it?
- If developers spend four hours a week reviewing AI output, has review replaced writing as the job?
- In a company where people judge and agents execute, what happens to those who learned judgement by executing?
- Is AI activity being counted as productivity, and who is designing the review step?
- When most pull requests start as agent drafts, is human review the bottleneck?
- How much of our people's time now goes on checking AI output, and who is paying for it?
- How many hours a week do our people now spend reviewing AI output, and is that where we want their judgement?
- What should a senior engineer do that a junior with unlimited tokens cannot?
- Does the white-collar hiring rebound show firms want people who contribute judgement alongside AI?
- If Amazon says its team neither wrote nor reviewed the code of a tool, what does trust mean for the engineers who use it?
- Who signs off the work at an AI-first law firm?
- Should a hospital trust an AI that writes the right plan but never places the order?
- Is a 70 per cent acceptance rate for AI-suggested edits a good or a worrying number for oversight?
- Will the in-house lawyer's job become risk management of agent teams?
In the New York Observer on 17 July 2026 I argued that the next AI power class will not be the people who build the models. It will be the people whose judgement others rely on: who decide whether what the machine produced is good enough, who can stop it, and who answer for the result.
The answer, in one line
The people whose judgement others rely on as AI does more of the producing: those who decide whether output is good enough, who can stop it, and who answer for the result.
Definition#
The judgment class is the group whose value lies in deciding, verifying and answering for work as AI does more of the producing. The term was set out in The Next A.I. Power Class Won't Build the Models (Observer, 17 July 2026). The New York Observer's American spelling is kept for the term; the rest of this site writes judgement.
Why power moves away from production#
When anyone can produce competent-looking output, competent-looking output stops being evidence of anything. Drafting, summarising and first analysis get cheap. What stays scarce is the part that cannot be generated: deciding what the output is worth, noticing what it missed, and being the person accountable when it goes wrong.
That is the verifier's role, and organisations tend to underpay it, because verification looks like checking rather than like producing. The argument is developed on staying valuable in the age of AI.
The trap at the start of a career#
Membership is earned through practice. AI lets someone produce work that looks like it carries fifteen years of judgement when it carries eighteen months, and the people who skip the hard early work never build what the judgment class is paid for. This research calls that synthetic seniority, and the disappearing junior work behind it the missing rungs.
What it asks of an organisation#
Two things. First, decide which decisions stay with people and who those people are, which is judgement allocation, written down as custody. Second, keep building the judgement those people need, on purpose, because the routine work that used to build it is the work being automated first.
2 to 8 October 2026: the checking turns up as a cost#
Two reports in one week put figures on the part of the work this page says stays scarce. Scarlett Evans at CIO Dive reported on 2 October, from Bain’s 2026 Global Technology Report, that AI coding tools are helping developers complete “roughly 21% more tasks, but time spent reviewing outputs rose 91%”. A Bain partner, Purna Doddapaneni, told her: “The bottleneck has moved from writing code to trusting it”, and “the system settles at the speed of its slowest human checkpoint.” Alexei Alexis at CFO Dive reported on 6 October a Datarails survey of 270 chief financial officers and finance leaders at large US organisations. They put the share of their teams’ work time spent “verifying or correcting AI-generated finance outputs” at an average of 26 per cent. One figure is from a consultancy’s report and the other from a company’s own survey, and the second is the leaders’ estimate, not a timesheet. They point the same way.
Two more describe what is then sold. Legal Cheek reported on 7 October that Falcon, a firm co-founded by the former Allen & Overy senior partner Wim Dejonghe, will review up to six commercial contracts a month for €1,000, returned the same day, and that “Every contract review is signed off by a qualified lawyer”, because clients still want what Dejonghe called “the judgment of a lawyer”. The Thomson Reuters Institute, writing on 2 October about the tax, audit and accounting firms its parent company serves, said hourly billing is still “used by 78% of firms” and forecast that AI will force them “to stop billing by the hour and start pricing for their judgment instead”. Christian Catalini’s essay in Harvard Business Review the same day gives the pattern a name. In the words of HBR’s summary, firms turn into “verification factories: institutions capable of properly steering AI-generated output and standing behind the results.” Only the opening of that essay could be read without a subscription.
The same summary carries the warning in the second half of this page. Verification factories, it says, rely on “unique ground truth and talent”, and “many firms are adopting AI in ways that erode both of these critical assets.” A trial published the same week shows where the erosion would start. It is set out on whether AI makes junior staff better or only makes their work look better. One caution about who is in the class today. Gallup reported on 6 October, from a survey of US employees fielded from January to March, that 37 per cent of managers use AI at least weekly against 25 per cent of individual contributors. Among employees who use it, 63 per cent say it makes them faster, 46 per cent say it raises the quality of their work and 43 per cent say it does not. Speed is widely reported. Whether the work is better is contested, and deciding that is the job this page describes.
9 to 15 October 2026: the reviewer is the limit, and accountability is what stays#
Three accounts from the people building the tools put the reviewer where this page puts the judgement. George Weale, Bo Yang and Andrew Biernat at Google Cloud reported on 13 October that an agent ran unattended for eleven hours on a 205,000-line codebase and proposed six optimisations; the maintainers shipped two and rejected four in review. Their summary: “the agent explores, the maintainer subtracts.” Anthropic’s source control team wrote the same day that the company averaged around 2,500 code changes merged a day in May 2026 and over 30,000 a day by September, and that for the system they built to carry it, “Our team members still review every change request for the system, but they now examine the intent of the change as opposed to reading every line of code.” Kelly Hong and Greg Burnham at Epoch AI, testing whether models could do their own organisation’s work, concluded on 8 October that the best models are reliable on well-defined tasks but “consistently fall short on the judgment that drives real work”.
The people doing the reviewing say what it costs them to trust. Stack Overflow’s 2026 developer survey, from over 30,000 respondents in 169 countries, found that 48 per cent trust AI when they can easily validate the answers and 6.6 per cent would trust it for important decisions; 78.7 per cent say source attribution matters. The survey’s own reading: “human-in-the-loop is the dominant mode of working with these tools.” Rabihah Butler at the Thomson Reuters Institute reported on 12 October, from her company’s 2026 Future of Professionals Report, that 47 per cent of professionals say they hold responsibility for errors in AI-assisted work, 48 per cent fear a negative effect on the development of their independent judgement, and legal professionals “expect the timeline to trusted judgment to extend by nearly two years under current conditions”. Her diagnosis: “The disease is the absence of clear professional standards for what AI-assisted work should look like in practice.”
What stays with the person was named most plainly in medicine. Andrew Parsons of the University of Virginia and Adam Rodman of Harvard, in a commentary in the Annals of Internal Medicine published on 12 October, accept that AI can match or exceed doctors at some diagnostic tasks and contend that what stays with the clinician is accountability, “owning the risk of acting under uncertainty and answering for the decision that follows”. They add: “A clinician who knows they will answer for the result weighs options differently than a system with nothing to lose.” And: “Deviating from the technically correct answer is often good medicine, and someone must answer for that deviation.” That is the class this page describes, in a profession’s own words. The week’s clearest case of the reviewer being removed rather than named is on whether a decision can rest on an AI summary.
16 to 22 October 2026: the tier is set by the change, and the review is the job#
The team that has moved furthest from reading code to judging it described what is left. Addy Osmani, Morgan Westlee Lunt and Qing Wang of Anthropic’s Claude Code team wrote on 15 October that Claude writes about 90 per cent of the product’s code, that review runs through seven layers, most of them machines, and that approvals are tiered by risk: none for low-risk changes, one person for medium, two for high. “The change sets the tier, not the author.” “An agent’s approval never counts as a person’s, either.” What people keep is deciding whether to build, risky code, architecture others will copy, reading the evidence and owning the outcome. The post reports merged changes more than doubling between June and September while the share reverted fell by over 40 per cent; it is the company’s own account, unverified. The full reading, and what the week’s other sources add, is on whether AI can check AI’s work.
Three surveys put numbers on the review that remains. Atlassian’s State of Developer Experience 2026, from more than 1,500 practitioners, found 87 per cent say AI helps them ship faster, 49 per cent spend four or more hours a week reviewing AI output and correcting prompts, 75 per cent say they now carry more responsibility for validating AI output, and 41 per cent say they use AI more than necessary because they feel expected to. The report’s own sentence: “Engineering is becoming more about judgment than code.” Richard Speed at The Register reported on 21 October a Linux Foundation survey in which maintainers estimated that about a quarter of incoming pull requests are low-quality AI work, 31 per cent at large projects, and only 35 per cent of projects have a formal policy permitting generative AI; Daniel Stenberg, who maintains curl, said all its key maintainers “feel strained, overloaded and brought to the brink of what we can muster”. Pratik Wadher, chief technology officer of Kaseya, told Paul Kunert at The Register the same day that after rewriting a storage layer with agents in six weeks, “the new code is the spec”, and that “the best engineers are still critical” for what they know rather than what they type.
Medicine supplied the week’s cleanest statement of why the class is needed. Nigam Shah and Jonathan Chen at Stanford HAI, summarising benchmarks their group has built from real clinical and administrative work, reported on 19 October that the best model completed 46 per cent of long-horizon physician tasks at one attempt and 28 per cent across three, and that on administrative tasks one agent passed 82.8 per cent of subtasks and 26.7 per cent of whole tasks. One agent “wrote the correct action in its note and never placed the order in the EHR”. Their line: “Competence on the parts does not add up to reliability in a whole workflow.” Somebody has to hold the whole, and answer for it. In law, the chief executive of a legal AI company wrote in Fortune on 22 October that his agents take a whole task from start to finish and “Lawyers step in where judgment is needed.” The capability debt page records what the Law Society said the same week about where those lawyers will have learned it.
23 to 29 October 2026: the reviewer who defers is not reviewing#
This page argues that the judgment class is whoever decides whether to accept. The week produced the first controlled measurement of how often the person in that seat decides anything. In Anthropic’s randomised trial of 27 October, 50 causal-inference professionals, most with graduate degrees or six years’ experience, matched a coding agent’s randomly assigned lean 86 per cent of the time across four methodological calls, and 54 per cent recommended the final course of action when the agent leaned that way against 36 per cent when it leaned the other. Experience, expertise, education and stated trust in AI did not predict deference; only age did, and the effect disappeared on adjustment. Seventy-seven per cent repeated the agent’s arguments and 15 per cent credited it. The authors’ sentence for the class is “Institutions often treat an expert’s sign-off as evidence of independent judgment.” The eleven who declined the agent made serious errors more often (55 per cent against 6 per cent) and took longer. The class is the person whose view existed before the tool spoke, which is not the same as the person who worked without one.
A preprint posted on 22 October measured the same deference from the other side. Thirty-two participants learned from assistants whose briefings had between none and all of their facts replaced with confident falsehoods; participants chose the planted falsehood on 60 per cent of falsified items, their confidence stopped predicting their correctness, their stated trust barely moved, and about 30 per cent of turns pasted the quiz question back into the source being checked. Checking the tool against the tool felt like verification. It is a small study with one model and a learning task. It measures what this page has called the trap at the start of a career.
The volume the class must handle kept rising. Ana Maria Constantin at The Next Web reported GitHub’s Octoverse figures of 28 October: more than 40 per cent of pull requests began with an agent draft in September, up from 17 per cent in March, 110 million agent-started pull requests in the month, and a company raising $143 million to check AI-written code. Steef-Jan Wiggers at InfoQ added on 27 October that 65 per cent of developers in Atlassian’s survey spend time correcting AI output, and quoted a reader’s question: “Is anyone designing the review step yet, or are teams just absorbing it?” Niam Yaraghi at Brookings made the same point about science on 29 October, after OpenAI’s release of 722 machine-generated manuscripts: “Absorption does not scale,” and “We have no institution for verifying machine-generated scientific claims outside of pure mathematics.” Tim Fernholz at TechCrunch reported MIT’s Ryan Williams on a proof two labs claimed in the same week: parts routine, and “two or three alien-looking parts”. The checking is the job, and the week showed the job outgrowing the people, and the people adopting the answer they were meant to check.
30 October to 5 November 2026: the reviewer is outscored, and the bill stays the same#
The seat this page describes was contested in two professions this week, and in both the answer was the same: the judgment class is whoever decides what the review is for. Gary King and Steven Worthington of Harvard’s Institute for Quantitative Social Science posted a paper on 28 October, reported by Frances Jones at Times Higher Education on 4 November, comparing 134 paid human reviews of 124 papers submitted to their rapidPeer service between 2019 and 2026 with reviews of the same papers by Anthropic’s Claude Opus. Judged by OpenAI’s GPT-5.5, the AI review won in all 132 decided pairs; judged by three former journal editors on ten random pairs, it won 87 per cent, with a higher mean rating in all seven areas scored. The paper is not peer reviewed, uses one model and is judged mainly by another company’s model. King said they “thought [that] we would find good justification for human reviewing”, that this was “not what we found”, and asked: “should OpenAI and Anthropic be deciding what’s good science?” The paper’s own warning is that “AI benefits us individually but risks making the collective efforts of science more homogenous and ordinary.” Its most realistic proposal has an editor judge a dossier of the paper, an AI review from the journal’s official prompt and the author’s response, adding a human review “only if deemed necessary”. The reviewer leaves the class. The editor who writes the prompt and decides when a person is needed joins it.
The law reached the same seat from the bill. In Flux Drive v Rong, decided on 30 October, Judge Carolyn McHugh wrote for a unanimous Tenth Circuit panel upholding a fee award of 114,663 dollars that the losing side said should be cut because the winning lawyers had used AI: “Plaintiffs have not shown that defense counsel’s use of AI rendered the hours that they actually expended unreasonable.” The firm said its lawyers had checked every piece of AI-assisted research; its partner Christopher Ottele told Above the Law that AI “is not a substitute for the judgment and analysis that attorneys still must exercise”. Michael Berman at E-Discovery LLC connected the ruling on 2 November to Joshua Yuvaraj’s verification-value paradox: “the net value of AI to legal practice is grossly overestimated, due to an underestimation of the verification cost.” The New York State Bar Association’s governing body adopted a report on 31 October, reported by Victor Li at the ABA Journal, restricting billing for time AI saved or time spent learning it, and reminding supervising lawyers that they “should not assume that a law firm-approved tool or policy removes the need to review work on each matter”. Dan Robinson at The Register reported the Productivity Institute’s finding, in a report commissioned by Bytes, that coding activity rose about 240 per cent where releases rose about 30 per cent, “due to humans still needing to check and test their work before release”; Bart van Ark: “AI’s productivity potential becomes real when faster tasks lead to better performance across the business.” The hours did not fall. They moved to the class.
Who sits in it was described from inside. Richard Tromans at Artificial Lawyer wrote on 2 November that equity partners have been “the final check on quality” and now become planners who “triage in terms of AI and non-AI”, which “makes them feel like an associate all over again”; “The buck stops with them.” Microsoft told the same site the next day that its Copilot for Legal sees “70%-plus acceptance rates already for suggested review changes”, a company figure with no account of what the lawyers who accepted had read. Yunita Anwar at Accounting Today described, on 5 November, a Mercor study in which an agent scored perfectly across 20 runs of four month-end tasks, unassisted accountants averaged about 37 per cent, and “assisted accountants were slightly less accurate and took longer than the AI agent alone”; her prescription is the class’s job description: preserve the first answer, “record each change, its reason and its supporting evidence”, agree a stopping rule with “a named person responsible for the decision”, and hold that “An attractive answer should have to survive a clear explanation.” Richard Zhe Wang’s working paper of 31 October cast model agents as audit senior associates and found they “yield to pressure from the engagement partner, from peers and from a reviewer from another firm” but largely not to the client, and “do not respond to social influence pressures as human auditors do”. And Anthropic’s account of its own users, published on 5 November from 207,049 sessions and a survey of 528 scientists, found that “Scientists primarily use Claude to analyze data and to write”, that for ideation and hypothesis generation 17 per cent report no AI use and 53 per cent only assistive use, that ideation is the task they value most, and that about 40 per cent expect AI to run full research projects by 2030. Its authors write that “deliberate checkpoints in the research process may be warranted, requiring us to accept slower progress to reduce risk”, and that AI “could take the lead in verification, refereeing, and communicating important results”. It is a company’s own data about its own product. It is also the scientists choosing, for now, to keep the question and hand over the checking.
The class after this week is the editor who writes the prompt, the partner who triages, the reviewer whose changes are recorded with reasons, and the scientist who keeps the question. Where the acceptance rate is seventy per cent and nobody writes down why, the seat is occupied and empty. What an organisation controls is whether the review is a decision with a record or a click, and the questions that settle it are at Rules Before Tools.
The limits of the argument#
It is an argument about where value moves, and no count of the class is claimed here. The evidence underneath it, on what AI does to expertise and how judgement is built, is graded on AI and human judgement. The longer treatment is in SuperSkills (Kogan Page, 2026).
Explainer · SS-2026-405 · Graded against the published rubric · How I use AI here · 3 working papers, 3 institutional surveys, 2 argued perspectives and 4 of other kinds
Hirji, R. (2026). What is the judgment class?. The SuperSkills evidence base, SS-2026-405. https://thesuperskills.com/research/what-is-the-judgment-class. Last reviewed 13 November 2026.
An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.
How citations and IDs work