The verification bottleneck

The verification bottleneck is the point at which checking work, rather than producing it, becomes the limit on what an organisation can do with AI. Generation scales with compute. Verification scales with skilled human attention, which does not scale. The idea has a harder edge than "remember to check the output": a small 2026 pilot found people leaned on AI most for the hardest problems, which were also the problems where their answers were least accurate. If that pattern holds, checking is weakest where it matters most. The phrase appears in a January 2026 research paper and in software engineering writing from the same month, and the idea was circulating through 2025. Nobody can fairly claim to have coined it.

Where the idea comes from#

Three lines of work meet in it. The first is AI safety research on scalable oversight, which Samuel Bowman and colleagues at Anthropic described in 2022 as "the problem of supervising systems that potentially outperform us on most skills relevant to the task at hand" (arXiv 2211.03540). The second is software engineering, where AI coding tools made writing code fast and left review as the slow step. Andrej Karpathy told a Y Combinator audience in June 2025 that he was "still the bottleneck", because he had to make sure the tool was not introducing bugs. In January 2026 a Sonar-sponsored article in The New Stack was titled "The AI verification bottleneck", and by June a TypeDB engineering blog was stating it flatly: "code verification, not generation, is now the bottleneck".

The third is economics. Christian Catalini, Xiang Hui and Jane Wu argued in Some Simple Economics of AGI (February 2026) that "the binding constraint on growth is no longer intelligence but human verification bandwidth". Their framing sets a cost to automate that keeps falling against a cost to verify that is held up by human biology.

The phrase also appears in the title of a January 2026 paper by Matthias Huemmer and colleagues, discussed below. Neither they nor the engineering writers coined the idea, and this page credits no one with it.

The evidence that checking is the constraint#

The Huemmer pilot. Huemmer, Durner, Shyiramunda and Cummings-Koether followed an academic group through three waves over six months, giving participants maths and analytical problems (arXiv 2601.17055). Daily AI use rose from 52.4 to 95.7 per cent. In the final wave 73.9 per cent relied on AI most for difficult tasks. Participants' accuracy, working with AI, fell with difficulty from 95.2 per cent on the easiest problems to 47.8 per cent on complex ones, and the gap between how well they thought they had done and how well they had done widened to 34.6 points. The authors conclude that the bottleneck had moved "from solution generation to solution verification". The samples were 21 to 36 people per wave, there was no control group, and the authors say this precludes causal inference. It describes a pattern; it does not measure an effect.

The METR trial. In a randomised trial published in July 2025, METR found that 16 experienced open-source developers took 19 per cent longer to complete real issues with AI tools, while believing afterwards that the tools had sped them up by 20 per cent. A larger follow-up published in February 2026 pointed towards speed-ups, but METR called the result an unreliable signal and is redesigning the study. The lasting finding is the gap between perceived and measured speed: people misjudged their own work with the tool. See what is the METR study.

Review time in software teams. Faros AI's telemetry across more than 10,000 developers, published in July 2025, found that teams with high AI adoption merged 98 per cent more pull requests while time spent reviewing them rose 91 per cent. The company called human approval the "critical bottleneck". This is vendor data and correlational.

Experts misled by wrong suggestions. When radiologists read mammograms with deliberately wrong AI suggestions, in a study published in Radiology in 2023, the accuracy of the most experienced readers fell from 82 per cent to 45.5 per cent. Expertise did not protect them from being misled.

Courts. Damien Charlotin's database of decisions noting hallucinated material in court filings passed 2,200 cases in October 2026. Each one marks a check somebody skipped.

Why checking gets harder when it matters most#

Verification is not equally hard everywhere. The AI researcher Jason Wei set out the other side in July 2025 in an essay on the asymmetry of verification: "some tasks are much easier to verify than to solve". A sudoku, a passing test suite, a reconciled ledger. Where an output can be checked cheaply and mechanically, the bottleneck barely exists.

The bottleneck bites where outputs cannot be checked that way: a strategy recommendation, a legal argument, a diagnosis, a hiring judgement, a policy summary. Checking these needs the same expertise that producing them needed. So the work most often handed to AI because it is hard is also the work that is hardest to verify, and verifying it requires capability the person may no longer be building. A Microsoft Research survey of 319 knowledge workers (CHI 2025) adds a warning: in that survey, people who trusted AI more reported thinking less critically.

So "a human will review it" is weak as a control. The reviewer is fastest and least careful on fluent, confident output, which is what these systems produce. See human in the loop is not a safeguard and automation bias.

Where this sits in my own argument#

The core claim of SuperSkills is that human judgement, not production, is the bottleneck in the AI era. The verification bottleneck is that claim at the level of a single task: production has become cheap, and the scarce thing is a person able to say whether the product is right and willing to put their name to it.

I have been caught by it myself. This month a charity asked for my view on its AI strategy. The document was good, so I asked, half-joking, whether the new colleague called Gemini had drafted it. It had. I told them they were heading in the right direction, which was true, and missed the question they needed answering, which was what they were trying to change and who would have to change it. As I wrote in Pre-answered: "I read their answer before I had one of my own." The verification bottleneck is not only about time. A confident, well-organised answer changes what the checker looks for.

It also explains why I worry most about the missing rungs. If juniors stop doing the work by hand, the organisation stops producing people who can check it, and a verification bottleneck today becomes a verification collapse a few years later. AI slop is what an organisation gets in the meantime, when it scales generation and never pays for the checking.

What an organisation can do#

Sort work by how checkable it is. Where outputs can be verified mechanically, automate hard. Where they cannot, treat AI output as a draft owned by a named person. The use-or-keep test starts with this question because it helps separate the safe uses from the dangerous ones.

Budget the checking. If a team's output doubles and its review time does not, either quality has fallen or someone is working unpaid hours. Plan review capacity as you would plan production capacity.

Measure verification as well as volume. Track review time, rework and how often reviewers reject work. If that last number is zero, find out why.

Keep the people who can check. Protect the repetitions that build expertise, and keep the experienced reviewers whose judgement is now the scarce input. See who owns verification when AI does the work.

Underneath all four sits the same position as the rest of this research. The risk that can be measured is in handing decisions to machines and what happens to human judgement afterwards. What an organisation controls is which decisions a machine may make, who can stop each one, what people must remain able to do, and how anyone would know if it went wrong. Verification is how anyone would know.

September and October 2026: confidence rising as controls fall#

Three surveys published in the autumn suggest that organisations' belief in their own checking is running ahead of the checking itself. KPMG's Q3 AI Pulse (24 September), from 314 US leaders at companies with revenue of $1 billion or more, found confidence in governance rising from 57 to 73 per cent in a quarter, while the share building controls into their agents fell from 43 per cent two quarters earlier to 30 per cent. Only 49 per cent had defined the high-risk uses where agents may not decide alone.

Accenture's Future of Work report, drawing on its July 2026 survey of senior executives, found only 45 per cent of organisations with monitoring capable of questioning AI outputs, and only 39 per cent of leaders saying human accountability is clearly established as agents take on more work. PwC's Hopes and Fears survey (29 September) of 49,364 workers in 48 countries found that 31 per cent of daily AI users name concerns about accuracy or quality as their top barrier, and only 34 per cent say AI helps them with the most complex and difficult parts of their job.

Read together: the people using the tools worry most about quality, the people running the organisations feel more confident than they did, and the controls that would justify the confidence are thinning. That is the pattern a verification bottleneck predicts, seen from the boardroom side.

What this does not show#

There is no controlled study that measures the verification bottleneck as such. The Huemmer pilot is small, has no control group and studied maths and analytical problems in one academic setting. METR's own follow-up weakened its headline result. The Faros figures come from a company that sells engineering analytics. The radiology study used deliberately wrong suggestions, which tells you about the worst case rather than the usual one. The evidence agrees on direction: producing is getting cheaper faster than checking. How much that costs, and in which kinds of work, has not been measured.