AI-DHD

Attention Directing Hyperactive Enablers.

The discourse around tech is dominated by security and coding at the moment. Those are the industries currently most disrupted. But while I’m a tech enthusiast, I’m hardly a tech professional. Below I talk about a different potential for this technology. This could be the crutch that people are looking for.

The Queue

NHS England estimates that around 2.5 million people in England have ADHD, diagnosed or otherwise. In much of the country the average wait for an assessment is now eight years; in some areas it has reached ten to fifteen. The number of children and young people waiting rose from about 21,000 in 2019 to about 270,000 by the end of 2025. Last year more than half of England’s 42 Integrated Care Boards capped how many assessments they would fund, in many cases without telling GPs or patients.

Behind the queue sits the treatment gap. The NHS’s own taskforce reports that 70-90% of people with ADHD benefit from medication, yet only 25% of children and 15% of adults receive it. It puts the avoidable cost of unsupported ADHD at £17 billion a year: educational failure, unemployment, crime, addiction, illness. And a Lancet Regional Health study published last month found diagnosis rates still running well below true prevalence, so the queue is the visible part of a larger problem.

The Impairment

ADHD is poorly named. It’s not about energy per se. The core problem is executive function. The clinical literature, from NICE’s guideline to Russell Barkley’s work, describes the same patterns: starting tasks, holding things in working memory, sensing time passing, ranking priorities, regulating emotion. A person can know exactly what needs doing, want to do it, and still be unable to begin. Very frustrating.

There is infinite guidance and information to which sufferers are ‘signposted’, but it’s of limited use. A deficit in execution is fixed by an external scaffold: something external that starts the task, keeps you on track, and watches the clock. AI can do all those things with minimal technical skill required.

The Enabler

Scaffolding is what language models are good at. Give one a task - the tax return, the insurance claim, the scary email - and it will break it into steps small enough to start, in seconds, without judgement. That last part deserves emphasis. The judgement-free quality and patience of a digital assistant removes the shame that comes with asking for the fourth reminder.

The same pattern repeats across the executive functions:

• Working memory: a model that has read your notes can answer “what was I in the middle of?”
• Time blindness: an assistant that plans backwards from a deadline supplies the time-sense the brain doesn’t.
• Task initiation: researchers are now studying AI “body doubling” - a companion presence that makes starting easier - formally, including a 2025 study testing AI against human body doubles on a simulated task in VR. Early and small, but the direction is right.

The clearest examples are domestic. Goblin Tools, a free suite built by the software engineer Bram de Buyser for his neurodivergent friends, has a Magic ToDo that turns “clean the kitchen” into a list of steps, and any step that still looms breaks down again. People use general chatbots the same way: one entrepreneur with ADHD told the Associated Press she has hers plan easy recipes with a matching shopping list, because the organising defeated her rather than the cooking.

The research is starting to catch up. A 2025 diary study followed fifteen neurodivergent adults using an off-the-shelf chatbot across their daily routines, and a 2026 analysis of 147 posts from ADHD forums found task initiation was the use people returned to most. Even the medication illustrates the stakes: keeping to a dose schedule needs exactly the routine the condition disrupts, and the clinical guidance flags adherence as its own problem in adult ADHD.

ADHD is the case I think about most, but the pattern is general. Be My Eyes wired GPT-4’s vision into a phone app so blind users get scene descriptions on demand. Live captioning is now built into every major meeting tool. Voice control operates a whole computer for people who can’t use their hands. In each case the model substitutes for a function the person is missing, at marginal cost, without an eight-year queue.

I’d be cautious, however, of rushing to build a crutch for ADHD into phones. They’re the most distracting thing humanity has ever invented. A dedicated device or launcher may be a better option. Helpfully, AI can probably help make that, too.

(This kind of support means telling an AI about your health and finances. I’d rather keep that off someone else’s servers, and it is why I run models at home behind The Privacy Gateway. A capable local model on reasonable hardware can do most of the scaffolding above in complete privacy.)

The Paperwork

There is a second, less discussed burden here: bureaucracy. The academics Pamela Herd and Donald Moynihan call it ‘administrative burden’ - the learning, compliance and psychological costs of dealing with the state. Annie Lowrey called it the time tax. Whatever the name, it lands hardest on people whose executive function is already a challenge.

The cruelty is circular. The forms that gate ADHD support - referrals, PIP applications, appeals, evidence bundles - demand sustained planning, working memory and deadline tracking. They are a test of the exact faculties the condition impairs! More than 100,000 people now claim PIP with ADHD as their main condition, up 40% in two years; every one of them navigated that paperwork or found someone to do it for them.

The state has started to admit the problem. GOV.UK Chat launched in May: an AI assistant over more than 80,000 pages of government guidance, aimed at call centres taking around 100,000 calls a day, with the government’s own research suggesting up to half of those questions could be answered by the tool.

The bureaucracy had outgrown any human’s ability to navigate it, and the fix the government chose was a language model. The same fix works from the citizen’s side: a model that drafts the appeal letter, translates the eligibility rules, or just sits with you while the form gets filled in. But rather than providing one faceless central tool to help the Whitehall bureaucrats with their working lives, personal-issue AI could help the everyman with everything.

The Floor

Geoffrey Rose observed that a large number of people at small risk may give rise to more cases than a small number at high risk, and concluded that the biggest gains come from moving the whole distribution slightly rather than the tail a lot.

England’s ADHD system (such as it is) is a pure high-risk strategy: an eight-year queue to join the treated tail, while the estimated 2.5 million sit spread across the population. Rose was writing about blood pressure, a risk carried continuously by everyone.

The ILO defines a social protection floor as four guarantees, held by all residents: essential health care, and income security in childhood, working age and old age. Help with the administrative business of living appears on no list; treating AI assistance as part of the floor would be a relatively cheap extension when costed over a lifetime.

The lesson from decades of targeted benefits is that people who qualify don’t claim. Pension Credit reaches between 55% and 71% of eligible pensioners depending on region. Broadband social tariffs, the nearest precedent to subsidised AI, reach fewer than 10% of eligible households - 5.1% of those on Universal Credit.

Access to Work, the scheme that already funds assistive technology for disabled workers, has a backlog so severe the National Audit Office says it now threatens the jobs it exists to protect. Gate AI support behind a diagnosis and it inherits all of this, with the eight-year queue as the front door.

So make it universal. When Korpi and Palme studied welfare systems in the 1990s they found a paradox: programmes aimed tightly at the poor tended to help them less than programmes open to everyone. A benefit the middle class also uses keeps its budget and its quality; a benefit for “them” gets cut. The finding is still argued over, but it suits a tool whose cost per additional user is small.

Amartya Sen built a whole framework on the observation that a resource only counts when a person can convert it into something they can actually do. In the UK 7.9 million adults lack basic digital skills, so the offer has to include the on-ramp. The first piece already exists: GOV.UK Chat is free and unconditional, and the LLMs themselves are like a universal natural-language interface (Microsoft was on to something there, they were just too early). The cost of investment sits opposite £17 billion a year cost of unsupported ADHD, and 14.3 million people in poverty.

Limits

In my experience the hardest problem for AI tooling is persistence - every system works for a fortnight, and then it goes. An assistant that prompts you autonomously may survive where one that waits to be opened will not; but that remains to be tested (blog post coming up!). Doctors studying these tools make a related point: AI is one tool in a toolbox alongside medication and skills training.

Cost is an issue, as always. The people with the most to gain are the least able to pay for subscriptions. Local models help, but a used GPU is still a lot. Hallucination also matters more here than in most consumer uses, because a confidently wrong form response can cost someone months.

And the state’s AI cuts both ways: the same technology that helps a claimant navigate the system lets the department automate its side, and critics have already noted the fear that when something important goes wrong, the only route left will be the automated one. There is also the ever-present risk of benefit fraud.

What Next

The UK’s disability employment gap is 29.7 percentage points. For autistic people the employment rate is 30.2%, against 82% for the non-disabled. Nobody publishes an ADHD-specific employment rate at all. Some large share of that gap is an execution-support gap, and executive support is now cheap, patient and available at 3am.

I want to find out how far the local version can go: a model on my own hardware acting as the executive assistant - capture, recall, planning, paperwork - for my partner, who suffers from ADHD. Watch this space!