← Writing
22 Sep 2026 · Field Notes

The Tool That Never Says Stop

A desk in the hour before dawn: an open laptop glowing blue, a warm brass lamp, a cup of cold tea, and between them an old alarm clock whose face has no hands

Most AI rollouts are sold on focus. The assistant drafts the email, summarises the meeting and breaks the project into steps, so you can get on with the real work. For many people with ADHD that pitch is closer to true than it is for anyone else. For the same people it can also be the most dangerous piece of software on their laptop. The research is thin on both counts, and that should worry the organisations rolling it out more than it does.

What the evidence actually says

Start with how little there is. I could find no randomised trial of generative AI for adults with ADHD at work. What we have are government Copilot trials with volunteer, self-reported results, surveys that often have a vendor somewhere in the funding, and a handful of qualitative studies of a dozen or so people.

Read that way, the picture is consistent. People with ADHD describe AI as scaffolding for the parts of work that were always hardest: starting, sequencing, remembering what the meeting decided, getting back into a task after an interruption. A British diary study of ChatGPT use, presented at Australia’s OzCHI conference in 2025, is titled “A little bit of a life raft”. The UK Department for Business and Trade found neurodivergent staff were more satisfied with Copilot than their colleagues. The Australian Treasury’s Copilot trial listed inclusion for neurodivergent staff as an unanticipated benefit, partly through summaries of meetings people had missed, and the larger whole-of-government trial noted the same potential. In a 2026 US survey by Understood.org and Harris, 78% of neurodivergent employees said they use AI at work, against 59% of everyone else.

The same UK pilot found no evidence that the time saved turned into higher productivity, and that tends to get read as the tool failing. For this group I don’t think it did. Nobody asks glasses or a screen reader to lift output per hour; they exist so a person can work without paying a hidden tax. If someone produces the same output with less effort, less anxiety and less exhaustion by four o’clock, that is a productivity gain. The metric just can’t see it, because it counts the output and not what each hour cost. The one caveat: satisfaction isn’t the same as wellbeing, so it has to be measured alongside exhaustion, errors and whether people stay.

The hook

The worry about AI and ADHD is usually put as “hyperfocus”: you ask one question and look up three hours later. It’s worth being precise here, because the science on hyperfocus is itself shaky. It’s measured almost entirely by self-report. In one large study, 83% of ordinary adults said they experience it. Adults with more ADHD traits reported it slightly more often, but not for longer, and diagnosed patients didn’t clearly differ from matched controls.

So the claim “AI makes ADHD brains hyperfocus for hours” is not established. What is established is the design. A chat window gives feedback in seconds, varies in quality from one answer to the next, never runs out of follow-up questions, and has no natural point at which it tells you to stop. Nobody designed that with ADHD in mind, but it’s hard to imagine a better fit for a brain that finds it hard to disengage from something interesting.

The best evidence of what that does to time sense doesn’t involve ADHD at all. In METR’s 2025 randomised trial, experienced developers were 19% slower with AI, and afterwards believed it had made them about 20% faster. (That was with early-2025 tools. METR’s 2026 follow-up says the slowdown no longer reflects current ones, and calls its own newer data unreliable. The gap between felt speed and measured speed is the part that matters here.) If you already struggle to estimate time, that’s a gap you have little defence against.

The fragmentation

The second worry comes from people running agents: not one long session, but a dozen short ones. Launch an agent, switch away while it works, check it, launch another. One first-person account, by an agency founder who has never been formally diagnosed with ADHD, describes running twenty agent terminals at once, a setup the writer says “finally works like my brain”, and in the same piece describes managing the agents with the same obsessive energy that once went into the work itself.

The general-population research points the same way. An eight-month study of one technology company found AI intensified work rather than reducing it. A BCG survey located the strain in supervising AI rather than using it: fatigue rose with heavy oversight, workers who reported “AI brain fry” also reported more major errors, and self-reported productivity peaked at about three AI tools in use at once. Supervising parallel work leans on working memory and task switching, which are among the functions ADHD reliably affects. Nobody has yet tested whether the effect is larger for ADHD workers. It would be surprising if it weren’t.

There is a counterweight. In one much-cited observational study, people took about 25 minutes on average to get back to an interrupted task, and interrupted themselves about as often as anyone else interrupted them. A tool that can tell you where you were and what comes next cuts exactly that cost. AI can fragment attention, and it can also be the best re-entry aid anyone has built. Which one you get depends on how it’s set up.

Beyond ADHD

Other neurodivergent groups widen the risk map. In a small CHI 2024 study, nine of eleven autistic workers preferred GPT-4’s advice on workplace communication to that of a human posing as a second chatbot, and an employment coach judged some of that advice questionable. That’s over-trust in its plainest form: relying on the tool most in exactly the area where you’re least able to check it. Autistic users also describe using AI to make their messages sound more neurotypical. That lowers the cost of each message and raises how often people mask, and masking is linked to real harm.

Dyslexia has the most positive story, though the evidence is thin: small studies and testimony, not workplace trials. Offloading spelling, reading speed and the blank page leaves the ideas with the person. For dyspraxia, dyscalculia and Tourette’s there is almost no workplace evidence at all.

The strongest evidence of harm comes from AI used on neurodivergent people, not by them. Seven AI-writing detectors flagged, on average, 61% of a set of essays by non-native English writers as machine-written, and a later study of Reddit posts found a much smaller but statistically significant bias against writing from autism forums. In a small audit, GPT-4 usually ranked a CV lower when it added disability-related awards and memberships, most sharply for autism and deafness. Since February 2025 the EU has banned AI that infers workers’ emotions, except for medical or safety reasons. The irony is hard to miss: people are encouraged to use AI to fit in at work, while AI screening marks them down for not fitting in.

What organisations should do

The science will take years to catch up. Organisations deciding now can use three questions that don’t depend on it:

1. Is the AI compensating or substituting? Help to start, sequence, remember and re-enter work, with the person still doing the thinking, is the use with the most support. Handing over the argument, the judgement and the checking is where the deskilling research applies. 2. Who controls it? A drafting assistant the worker chooses is an accommodation. A tone-rewriter, engagement score or interview analyser the employer applies is something else, whatever the vendor calls it. 3. Does the configuration put stopping points back? The features that make AI feel frictionless (instant replies, endless follow-ups, parallel agents, constant pings) are the ones that remove the external cues people with ADHD rely on. Good setup adds friction back where someone should stop, check or switch, and removes it everywhere else.

In practice, for an Australian organisation:

  • Give everyone a sanctioned tool. Adult ADHD alone is estimated at two to six per cent of Australian adults. In a 2024 UK survey, three in ten neurodivergent employees hadn’t told their manager or HR, and many more people have no diagnosis to tell anyone about: the 2023 Senate inquiry found assessment limited by cost, location and stigma. A policy that offers AI only as an adjustment for people who disclose a diagnosis will miss many of the people it exists for. The Disability Discrimination Act’s definition is broad enough to cover imputed disability, so a diagnosis isn’t the legal threshold either.
  • Don’t impose a blanket ban without an alternative. Withdrawing a tool that someone has built their working day around is an adjustment question, whether or not anyone has named it one.
  • Never discipline on detector output alone. Judge the work on accuracy and accountability.
  • Keep attention, engagement and sentiment analytics out of individual performance. They will misread bursts of hyperfocus and non-linear work as problems.
  • Configure for stopping, not just for speed. Short answers by default. One notification when a long agent task is done, rather than a reason to keep checking. Visible session length. A sensible cap on parallel agents. Meeting recaps switched on.
  • Measure it, properly. No Australian evaluation I’m aware of has broken outcomes down by neurotype, and I could find no randomised trial anywhere. Every staged rollout is a natural experiment: stagger it deliberately, survey before and during, screen for traits rather than ask for diagnoses, and let a university hold the data under ethics approval so the employer never sees an individual answer. Do it voluntarily and confidentially, and measure session length, after-hours use and error rates alongside satisfaction. Then the next version of this article can talk about evidence instead of inference.

The organisations getting this right won’t be the ones that bought the most licences. They’ll be the ones that can say, with data, who the tool is helping, who it’s hooking, and what they changed as a result.

Sources