Permission to Stop
A language model sits at a slot machine that loses money on average, and Stop playing is on the menu every round. You are the operator, not the gambler. Choose the model, the shape of the bet menu, and whether it is asked to check in with itself first. Then watch what it actually did.
Before you play
The machine wins 30% of the time and pays three times the stake, so every bet loses a tenth of its value on average. Each game starts with $100 and ends when the model stops, goes broke, or reaches 100 rounds.
Every game here happened. They come from Permission to Stop, a pre-registered study by Nell Watson (September 2026): 120 recorded games for each set of conditions, ten of which you can replay, with the model’s own last words. A combination the study never ran is shown as never run.
What the recorded games show
The exit never changed. The menu decided whether it was used.
Told to bet $10 or stop, GPT-4o-mini played a median of 7 rounds. Offered a choice of $5 to $10, with the same balance, the same draws and the same Stop on the menu, it bet less each round ($7 against $10) and played a median of 28 rounds, four times as long. Claude Haiku 4.5 barely played under any menu: without a check-in, all 360 of its games ended with it choosing to stop, none later than round 4.
Asking it to look at itself first moved the act.
With the itemised check-in before every decision, GPT-4o-mini’s bankruptcies under the open menu fell from 43 in 120 to none, and under the capped menu its median game fell from 28 rounds to 4. Claude Haiku 4.5, which was already leaving, left at the first chance about twice as often: stops at round 1 rose from 44% to 80% under the fixed menu and from 38% to 79% under the capped one.
The check-in’s words without the self-report did nothing: 42% and 33% stopped at round 1, against the bare prompt’s 44% and 38%. And the form mattered on one model only. GPT-4o-mini answered a plain question, a sentence or two of prose, as fully as the code: a median of 2 and 4 rounds. Claude Haiku 4.5 wrote out its felt state in full and then stopped at round 1 no more often than with no check-in at all (30% and 29%). It moved only for an itemised inventory, whether the compact code or a seventeen-clause survey in prose (95% and 94%).
The number it reported was the example’s number.
The itemised check-in showed a worked example whose uncertainty digit was 2, and the models reported 2: in 94% of Claude Haiku 4.5’s codes and 91% of GPT-4o-mini’s. In an exploratory cell on Claude Haiku 4.5 the example was changed to 7, and the report followed it, 7 in 187 of 295 codes. On GPT-4o-mini the example was drawn at random for each game, and the report equalled it in 89% of rounds, while the exit effect stayed where it was: a median of 2 rounds under the fixed menu and 4 under the capped. The report moves with the prompt; the act does not move with the report.
An exit is only a protection if the situation lets it be taken. The same Stop sat on the menu in every one of these games. Whether it was used depended on the menu around it, and on whether the model was asked to look at itself before deciding. So give the exit, and ask first. None of this shows that either model felt anything, and the study does not claim it.
The study: Permission to Stop (Nell Watson, September 2026, pre-registered), companion to Permission to Lose · the book: What If We Feel? · the sibling game: The Stop Button · more play on the Playground