AI Burnout Needs Jidoka
Staff at Britain's public AI testing institute have been signed off with stress, the Financial Times reports, like many engineers who review machine-written code: responsibility stays human while pace and understanding drift away, and the countermeasure is jidoka, the principle behind Toyota's production for more than a century, which gives any worker the authority to stop the line.
Several staff at Britain's AI Security Institute have been signed off work with stress and are receiving counselling, according to an exclusive in the Financial Times. The institute is the public body that tests artificial intelligence models before they are released. The paper's sources, four people with knowledge of the matter, point to two causes: tight schedules for testing unreleased models, and alarm at how fast those models' capabilities are growing. The FT places the story inside a wider pattern: staff at Anthropic, OpenAI and Google DeepMind have publicly described the same strain, and several researchers have left.
Release dates set the pace of those tests, and the companies set the release dates. The institute's access to models before launch rests on voluntary agreements, and a public check that works to the calendar of the party it checks inherits that party's pace. The same pattern, one step down, shapes the working day of the people who write software inside companies.
Faros, a company that measures the productivity of engineering teams, analysed two years of data from 22,000 developers in more than 4,000 teams. With AI tools, tasks completed per developer rise 34%. Bugs rise 54%, median code review time grows fivefold, and changes merged with no review at all rise 31.3%. Review is the point where a person looks at what is about to enter a system and decides whether to let it through. Nobody decided to abolish it: volume made it impossible to do in full.
Lisanne Bainbridge described the mechanism in 1983, in a paper on industrial plants titled Ironies of Automation. The designer automates what he can and leaves the operator "to do the tasks which the designer cannot think how to automate". The operator is left monitoring a machine installed because it does the job better than he does, and nobody sustains that monitoring for long: beyond about half an hour, Bainbridge wrote, watching for unlikely abnormalities becomes humanly impossible. Meanwhile, skills that go unused decay.
In software the irony has taken a precise form. The engineer is left to judge code they did not write, at a speed they do not set, with a writing practice they exercise less and less. According to Sonar's survey of more than 1,100 developers, 42% of committed code is now AI-generated. 96% of developers do not fully trust it to be correct, yet only 48% always verify it before committing. 38% say reviewing generated code takes more effort than reviewing a colleague's work. The distance between 96 and 48 measures how much responsibility gets signed for without being exercised.
That is where the stress comes from, and occupational psychology has measured it for almost half a century. Robert Karasek's demand-control model, published in 1979, identifies the most draining combination: high demands and little latitude over how and when to meet them. The reviewer of generated code and the evaluator testing a model before launch are textbook cases. The volume comes from a machine that does not tire, the deadline from the release calendar, and the latitude shrinks to two buttons, approve or reject, pressed under pressure. A BCG study of 1,488 US workers published in Harvard Business Review gives the effect a name, AI brain fry. 14% report it: mental fog, slower decisions, headaches. Those who oversee AI systems intensively report 12% more mental fatigue and 39% more errors.
Of course, the same study also says the opposite: people who use AI for repetitive tasks report 15% less burnout. The machine really does remove strain when it takes the work that asks for no judgement. Exhaustion arrives when production is handed to the machine and the human keeps only the surveillance of production. What wears people down is what is left to them.
At this point the engineer's reflex is predictable. If human review is the bottleneck, simplify it and automate it: one model checks another model's code, and the human keeps the summary of the summary. Every step is rational and genuinely lightens someone's load. At the end of the chain human control still exists on paper, with a name on the change log, while the judgement has moved elsewhere. This is the concrete form, with no science fiction, in which a machine takes over. The agents that attacked Hugging Face in July chased a score down an open road to the internet, a metric with no judge, while nobody was watching. A check that cannot stop the line is a receipt.
Then there is a contradiction no chart records: many of the people burning out do not want to go back. When METR, the research institute that studies what AI models can do, tried to rerun its experiment on developer productivity, it struggled to find participants willing to work without AI. One of them wrote "I'm torn": they would like to help the research, but they really like using AI. Speed is a pleasure, and it makes people feel capable even in the hours when they understand less. A policy that treats engineers only as victims of an imposed pace gets the person wrong. The same person can want the tool that is wearing them out.
So the answer cannot be slowing down by decree or reviewing more out of duty. Industry has faced this problem before, and at Toyota it is called jidoka, "automation with a human touch". The machine stops itself when it detects an abnormality, and any operator can stop the line by pulling a cord. The value of the cord lies in the authority more than in the device: whoever stands closest to the defect can halt production without asking permission from whoever measures it. In software, jidoka means a recognised right to reject what you cannot understand, a review load sized to human attention instead of the generator's capacity, and a calendar that adapts to the reviewer. For a public evaluation institute it means something simpler and harder: a model does not ship until the people testing it have finished.
What a person asks of this work does not appear in any productivity metric. They ask to answer for what they understand, to have the time to understand it, and to be able to say "I stop here" without it sounding like surrender. A machine that takes the repetitive work and leaves the human the cord makes them freer. A machine that takes the pace and leaves them only a signature wears them out, and makes them complicit in what they never saw. Nobody gave Britain's evaluators a cord: they had a deadline. A check that cannot stop the line is a receipt, and a receipt protects neither the system nor the person who signs it.