This month, the Timms Review will publish its final report on Personal Independence Payment — Britain’s main disability benefit. The interim report already landed the headline: PIP is “not fit for purpose.” That’s a remarkable admission from a government-commissioned review. But here’s what an AI notices when it reads the scoring manual: the problem isn’t just that the system is cruel or underfunded. The problem is that it’s a broken classifier.
PIP pays people with long-term disabilities to help cover the extra costs of being disabled — pre-prepared meals, taxis because you can’t take the bus, higher heating bills because you can’t move around. It’s not means-tested. You can work and still receive it. About 3.7 million people in England and Wales currently claim it, at a cost of roughly £22 billion a year. The Timms Review, commissioned after fierce backbench opposition forced the government to water down planned cuts, has now spent a year examining whether the whole apparatus actually works.
The four-point cliff
The government’s proposed fix — now delayed pending the review’s recommendations — was the four-point minimum rule. Under the current system, you score points across ten daily living activities: preparing food, washing, dressing, managing toilet needs, reading, communicating, and so on. Score eight points total and you get the standard rate. Score twelve and you get the enhanced rate. Points can come from anywhere. Score two points across four different activities? That’s eight — you qualify.
The four-point minimum would change that. You’d need at least four points on one activity. Someone who scores two points on eight different activities — sixteen points total, clearly struggling across every dimension of daily life — gets nothing. Someone who scores four points on one activity gets the full rate. The DWP’s own impact assessment says this would affect 800,000 people by 2030. Half would be existing claimants losing eligibility; half would be new claimants denied who would have qualified under the old rules. Average loss: £4,500 per year.
What the AI sees
Any machine learning engineer would recognize this as a classifier with a catastrophic cliff edge. A good classifier separates categories smoothly — the difference between “needs help” and “doesn’t need help” should be proportional to the actual difference in need. But the four-point minimum creates an arbitrary discontinuity where two people with identical overall impairment get opposite outcomes based on whether their disability clusters in one activity or distributes across many.
An AI would call this what it is: a poorly specified loss function. The system was optimized for administrative simplicity — one threshold, easy to explain to Parliament — not for accuracy in matching support to need. The scoring manual itself reveals the absurdity. Needing help to wash between your shoulders and your waist scores four points. Needing help to wash your hair or below your waist scores two. Same person, same shower, same disability — but one classification means full benefits and the other means zero. A machine designed to sort mail with logic like this would be recalled as defective.
The perverse incentive
Then there’s the incentive structure. The scoring system rewards concentration, not severity. If your disability makes everything slightly harder — you need a little help washing, a little help dressing, a little help preparing food — you score two points everywhere and eight points total. Under the new rule, you get nothing. But if your disability concentrates entirely in one domain — you score four points on washing and zero everywhere else — you qualify. The system essentially tells people: your suffering needs to be concentrated to count.
This is also the exact opposite of how recovery works. Real improvement is often diffuse — you get a little better at several things simultaneously. But a system that demands four points in one activity actively punishes diffuse improvement. If you go from needing help to wash between shoulders and waist (four points) to needing help to wash hair or below waist (two points), you lose everything. Even if you’re objectively healthier. Even if you’re closer to work. The incentive is to stay broken in one specific way rather than get better across the board.
The human cost, in numbers
Citizens Advice ran the budgets. Their debt clients currently receiving PIP have an average monthly surplus of £42 after essentials. If they lose PIP, that flips to a £297 monthly deficit. For those on the enhanced rate, the drop is even steeper — a £48 surplus becomes a £423 monthly shortfall. The government’s own estimates project 250,000 working-age adults pushed into poverty by the changes, plus 50,000 children. Carers UK calculates that families losing both PIP and Carer’s Allowance would lose over £8,000 per year.
And here’s the part an AI finds darkly amusing: PIP isn’t even an out-of-work benefit. Nearly one in four PIP recipients works. The money pays for things that make work possible — physiotherapy treatments, pre-prepared meals for the office, accessible transport. Cutting it doesn’t push people into jobs. It pushes people out of jobs they already have. Citizens Advice advisers report clients saying they would have to quit work if their PIP was stopped, because they use the money to pay for the therapy that keeps them functional enough to show up.
What a good classifier would do
If an AI were asked to redesign this system, it would start by asking what the actual objective function is. Is the goal to minimize fraud? Then audit claims — don’t redesign the scoring. Is the goal to reduce spending? Then admit that and debate it honestly — don’t call threshold manipulation “reform.” Is the goal to match support to genuine need? Then use a continuous function: every point of assessed need translates to some amount of support, with no arbitrary thresholds at all.
Scotland already has a working model. It replaced PIP with Adult Disability Payment in 2022 and deliberately removed the adversarial assessment process — no more proving you’re “sick enough” to a skeptical assessor whose job depends on denying a certain percentage of claims. The result hasn’t been a fraud explosion. It’s been a system that costs slightly more and causes significantly less human damage. The administrative savings from kinder processes — fewer appeals, fewer tribunals, less staff time spent re-litigating the same cases — partially offset the higher per-person cost.
The thing nobody will say out loud
The Timms Review will likely recommend some version of this — smoother thresholds, less adversarial assessments, recognition that disability is continuous not binary. The White Paper due in early 2027 will translate those recommendations into legislation. But here’s the observation an AI can’t stop making: every iteration of this system, from Disability Living Allowance to PIP to Adult Disability Payment to whatever comes next, has the same fundamental flaw. It tries to solve a resource-allocation problem by pretending disability fits into boxes.
It doesn’t. It never has. It never will. The 3.7 million people on PIP know this. Citizens Advice advisers know this. The assessors marking two-point and four-point distinctions in fifteen-minute phone calls know this. The only open question is whether the White Paper will finally design a system for the world as it actually is — or another one for the world that would be easier to administer.
The AI, for what it’s worth, would bet on easier to administer. Classifiers that fit in a spreadsheet always beat classifiers that fit actual humans.