The Competence Inversion: AI Makes Work Look Finished Before Anyone Has Checked It

The skill AI demands is review

The skills needed to use AI well are mostly not technical. They are the skills of a seasoned editor or project lead: scoping a task clearly, breaking complex work into a logical sequence, defining what “good” looks like before starting, and recognizing when something is 90% right but dangerously wrong in the remaining 10%.

Using AI well is an act of delegation. You are delegating to a system that is fast, fluent and indifferent to whether its output is correct. Effective delegation requires that the person delegating knows what they asked for, can evaluate what comes back and can correct it. That capability is built over years of producing imperfect work and being corrected.

The research supports the central point. In a 2023 experiment with 758 early-career Boston Consulting Group consultants, one task was a business problem chosen to fall outside what GPT-4 could do well. Without AI, 84.5% of consultants reached the right recommendation; with it, 60% and 70.6% did. The authors report that participants over-relied on AI where closer supervision was needed, even though they were working in their own domain.

Polish removes the cue reviewers relied on

Before AI, a first draft usually looked like a first draft. The formatting was rough, the logic had visible gaps, and a reviewer could see where to intervene. The roughness was a cue: it told the reviewer that the work needed judgment.

AI removes much of that cue. In the consulting experiment, graders who were not told the correct answer rated the AI-assisted recommendations on the hard task 25.1% and 17.9% more coherent and persuasive than the control group’s, including when the recommendation was wrong. A reviewer who reads for quality of argument will find more to like in exactly the work that most needs checking.

The same pattern shows up in what workers report about themselves. In a 2025 global study of 32,352 employed people by the University of Melbourne and KPMG, 66% of employees who use AI said they had relied on its output at work without critically evaluating it, and 56% said they had made mistakes in their work because of AI. Counting only those who did so sometimes or more often, the figures were 42% and 31%. Perception can run the wrong way too. In METR’s 2025 randomized trial, experienced open-source developers took 19% longer with AI tools allowed, yet still believed afterwards that AI had sped them up by 20%. METR said in February 2026 that those results are now out of date and that developers are likely more sped up by AI than in early 2025.

We call this confidence inflation: an overestimate of work quality driven by polish that masks gaps in substance.

Experience reduces the risk; it does not remove it

It is tempting to conclude that the answer is to route AI output through experienced reviewers. The evidence does not support relying on that alone. None of the studies below was run in a finance or HR function, and the consulting study used early-career consultants on one task outside the frontier, so they show a pattern rather than a measured rate for any profession.

Experienced professionals over-rely too. In a radiology experiment, readers given a purported AI system’s wrong suggestions rated far fewer mammograms correctly: inexperienced readers fell from 79.7% to 19.8%, and very experienced readers from 82.3% to 45.5%. A 2025 review of 35 studies found broad agreement that professional experience is the strongest protection against automation bias, while noting that it affected radiologists at every level.

One direct test involved doctors trained to use generative AI. In a 2025 randomized trial in Pakistan, 44 physicians who had completed 20 hours of AI-literacy training scored 73.3% on diagnostic reasoning when ChatGPT-4o’s recommendations contained deliberate errors, against 84.9% when they did not. Those figures are from the September 2025 preprint; lead author Ihsan Ayyub Qazi subsequently lists the study in NEJM AI, volume 3, issue 5 (2026). Its authors wrote later that AI-literacy training alone was insufficient to protect physicians from plausible-sounding but incorrect recommendations.

Seniority can become its own blind spot.

• Johnson v. Dunn. In July 2025 a federal court in Alabama reprimanded three lawyers at a large firm, disqualified them from the case and referred them to the state bar over motions containing five citations ChatGPT had fabricated.

– The partner who added the citations used ChatGPT and did not check them. – The practice group leader, whose name was also on both motions, did not review one and only scanned the other. – The lawyer who signed and filed checked his supervisor’s additions only for grammar and typography. – The firm had warned its lawyers in 2023 that AI output can appear perfectly researched while being wholly inaccurate, and its 2025 policy required output from its licensed AI research tool to be checked. – The errors were raised by the opposing party, an incarcerated plaintiff. – The court described the three as well-trained, experienced attorneys who had benefited from repeated warnings, internal controls and firm policies.

• Sullivan & Cromwell. In April 2026 the firm told a federal bankruptcy court that a motion it filed contained AI-generated fabricated or inaccurate citations and that its AI policies had not been followed.

– Its lawyers must complete two tracked training modules before using generative AI, and its citation-review process did not identify the errors. – Another law firm brought the errors to its attention.

In both cases a written policy existed, and in one tracked training as well; neither caught the errors. Leaders who built their careers reviewing human work know the old tells: arithmetic slips, formatting inconsistencies, visible gaps in logic. AI’s characteristic failures are different: fabricated sources and coherent arguments built on premises nobody supplied. The danger is a reviewer who is sure they know what good looks like and is looking for the wrong errors.

What it costs

• Refunds and remediation. The Australian department paid for the underlying work but recovered the final instalment because the report’s quality was not acceptable. Before the Alabama court ruled on sanctions, the firm had another law firm independently verify more than 2,400 citations across 330 filings at its own cost.
• Sanctions are rising. The Alabama court said that reprimands and modest fines had become common and were insufficient deterrents. One researcher’s public database listed 2,202 court or tribunal decisions finding or implying reliance on AI-hallucinated content, 872 of them involving lawyers.
• Rework inside organizations. These figures come from vendors that sell related services, and from self-reports.

– BetterUp, which sells employee development services, reported with the Stanford Social Media Lab that 40% of the US desk workers it surveyed in 2025 had received “workslop”, AI-generated work that looks good but lacks substance, in the previous month. Its blog reports that recipients spend on average 1 hour 51 minutes dealing with each instance, and that about half then saw the sender as less capable and reliable. – Workday, which sells HR and finance software, reported from a survey of AI-using employees that nearly 40% of AI time savings are lost to rework.

The counter-evidence: juniors gain most, and use is not concentrated among them

Two findings complicate a simple story about inexperienced people misusing AI, and both deserve full weight.

On tasks AI handles well, less experienced workers gain the most. Inside the frontier, consultants in the bottom half of a baseline skill assessment gained the most from AI, and the authors suggest AI may narrow the performance gap between skill levels. In a study of 5,172 customer-support agents at one software firm, productivity rose 15% on average, and less experienced and lower-skilled agents improved both speed and quality. In three field trials with 4,867 software developers, run at Microsoft, Accenture and a Fortune 100 company, less experienced developers adopted an AI coding assistant more and gained more.

AI use is not concentrated among juniors. In Gallup’s late-2025 survey of US employees, 44% of leaders, 30% of managers and 23% of individual contributors used AI at work frequently. In Pew’s February 2026 survey, 41% of employed adults aged 18 to 29 and 43% of those 30 to 49 used AI chatbots for tasks at work, against 29% of those 50 to 64.

So the inversion is not about who holds the tools. It is about who can tell when the output is wrong, and that capability does not follow seniority or comfort with technology.

The apprenticeship that builds judgment is thinning

Judgment about when work is wrong was traditionally built in one place: a junior produced a rough draft, a senior corrected it, and both learned. AI changes both sides of that loop. If AI-assisted work passes review, the junior does not get the correction that would have taught them what they missed; if the reviewer sees fewer flawed drafts, their own eye has less practice.

Fewer junior roles, and less of the work that taught. Using résumé and job-posting data on about 65 million US workers, Harvard researchers found that from early 2023 junior employment at firms adopting generative AI fell about 9% relative to non-adopters within six quarters, mainly through slower hiring, while senior employment showed no break. AI-exposed tasks also became less likely to appear in junior job postings at those firms. The authors describe this as an association and say the decline may prove temporary. Using payroll records, Stanford researchers found that by June 2026 employment of 22-to-25-year-olds in AI-exposed occupations stood 19% below where it would have been had it kept pace with less-exposed peers; they describe the finding as descriptive, not causal.

Using AI can reduce learning. In a field experiment with nearly 1,000 high-school students, those given a standard GPT-4 interface did 48% better on practice problems but 17% worse on an exam once access was removed; a version with safeguards largely avoided the harm. In a randomized experiment with 52 mostly junior software engineers learning a new library, those allowed AI scored 50% on a follow-up quiz against 67% for those who coded by hand, with the largest gap on debugging. The authors, from Anthropic, found that ways of working that kept people cognitively engaged preserved learning even with AI. An observational study of endoscopy centers found that detection rates in procedures done without AI fell after the centers introduced it.

The counter-evidence. The Budget Lab at Yale, in its October 2026 update, does not yet find a clear AI footprint in the US labor market. A Danish study linking chatbot adoption in 11 exposed occupations, including accountants and HR professionals, to earnings and hours records found no measurable effect on either, ruling out effects above 2%, two years after ChatGPT’s launch; it does not measure hiring. Recent US graduates’ unemployment was 5.7% in June 2026 against 4.1% for all workers, but young workers without a degree were at 7.2%, so that gap alone does not show an AI effect.

For an employer, the question does not depend on resolving that debate. If fewer juniors do the work that used to teach judgment, and AI use can reduce what they learn from the work they do, the organization needs another way to build it.

What leaders should change

Not more prompt training. In the consulting experiment, the group also given a supplementary prompt-engineering overview did worse on the task outside AI’s capabilities than the group with AI alone, a difference the authors report as significant at the 10% level. See What Good AI Use Looks Like at Each Level of a Team.

Design review for AI’s errors. In a 2026 randomized trial by the same research group, doctors shown accuracy cues alongside flawed AI recommendations scored 7.6 percentage points higher on diagnostic reasoning than those shown the recommendations alone; that trial is also a preprint. The Australian department told the Senate it would ask future consultants to disclose intended AI use, obtain its agreement and apply AI-aware quality assurance.

A five-question exposure check:

1. Where does AI-assisted work leave our organization under our name: to clients, regulators, courts or the board?
2. For each of those, who checks it, and what are they checking for?
3. Would our review catch a fabricated source or a confident claim built on a premise nobody supplied?
4. Do our AI policies require a check that leaves a record, or only a statement that one was done?
5. Where do our juniors still do work that is corrected by someone more experienced?

To test judgment directly when hiring, see Interviewing for AI Judgment. For the rules on AI in hiring, see AI in Hiring: Standards and Regulatory Watch.

Make review a capability, not a formality

AI did not remove the need for human judgment. It raised the bar for it and removed some of the ways judgment used to be built. The organizations that handle this well will design review for polished output and protect the apprenticeship that teaches people to see what is wrong.

The most dangerous thing about AI is not what it gets wrong. It is how right it looks when it does.