12 October 2026 · AI in five years
Invented: The Version That Assessed Her
Invented. The one post a week WE makes up.
The appeal is the fairest part of the benefits system. It's only fair because the first part fails so often.
D. is thirty-four. She's been on a plastic chair in a tribunal waiting room in Wolverhampton for an hour and twenty minutes. Her coffee went cold before she got through the security door. Her mother is in room 3 with three people on a panel, and D. is not allowed in.
By 2030, at least one tribunal will turn on the question of which version of the AI model made the original recommendation, because the contractor will have updated it between the assessment and the appeal. If no tribunal gets there by then, the version gap isn't the problem it looks like.
The bundle they sent is 147 pages. An AI pulled it together from the clinical record, the call transcript and the original decision notice. The assessor who co-signed the recommendation no longer works for the contractor. The contractor no longer exists. What's left is a 45-minute video call from August 2029, cut to four bullet points, and a recommendation the system wrote seventeen minutes after the call ended.
D. spent three evenings on the form. She knows her mother's condition better than any assessor could from a Tuesday morning call. It varies. Worse in the mornings. Fine on the day of the assessment, because her mother had been in bed since Saturday getting ready for it.
Variable is the word the system handles worst. A snapshot scores a snapshot.
By 2031 the assessment contractor is a software stack with a clinician who co-signs. Eleven seconds per signature, on average, across 800,000 assessments a year.
Here's the sum nobody in any policy paper will do. Roughly two thirds of PIP appeals that reach tribunal get overturned. On 700,000 assessments a year and a 7 per cent appeal rate, that's 33,000 people a year the system got wrong badly enough for a panel to reverse it. Double the assessment volume and you get 66,000 reversals, on roughly the same tribunal headcount. The wait between filing and being heard is already eighteen months in some places. D. filed in October 2029.
The honest counter is that AI is more consistent than the human assessors it replaced, who were also inconsistent, also wrongly. That's probably true. Consistent wrongness on a variable condition is still wrongness.
The panel in room 3 can't know what the assessor actually saw. The assessor is a summary now. The summary was written before the assessor's own notes were done. D.'s mother told the assessor about the Tuesday thing: fine on Tuesdays, after resting since Saturday. The assessor wrote it down. The summary says: claimant reports intermittent difficulty with mobility.
Intermittent. Which means sometimes she's fine.
What hasn't changed: the plastic chair. The coffee machine that only takes coins. The fact that being right costs eighteen months and a daughter who can take the day off.
What has changed: there used to be a person you could write to. Someone who got it wrong had a name on a door.
Now the person who got it wrong is a line in a log file at a company that no longer trades.
A door opens somewhere down the corridor. Not room 3.
D. has done this before. Different city, different year, same chair. Her mother won that time too.