1 August 2026
We told the AI it could stop early. It heard: do less
Synthetic narration · 4 min
unPaper hands your AI a prompt and gets back a slide deck as one HTML file. The failure nobody can do anything about is the reply that runs out of room and stops mid-sentence — you are left holding a broken file and no idea what to do next. So we added a safety net: if you run out of room, finish the slide you are on and tell the reader to bring what you wrote here, where the missing parts get put back automatically.
It made decks worse. Not the ones that were failing — the ones that were fine.
What happened
On short decks the safety net changed nothing at all. On long decks — the ones it was added to protect — models responded by writing roughly a third less deck.
| Model | No safety net | “You may stop early” | “Last resort only” |
|---|---|---|---|
| deepseek-v4-flash | 16 | 11 | 14 |
| gpt-5-mini | 18 | 13 | 18 |
| mistral-large-2512 | — | 18 | 20 |
On the shorter brief every model returned the same number of slides in every arm — 8 — so the effect only appears once a deck is long enough to approach the limit.
And the escape hatch was never used, not once. Across every run in the experiment, no model wrote a full deck, hit the ceiling and handed over. They avoided the ceiling by writing less in the first place. Permission to stop early was read as permission to do less — which is a different thing, and the more expensive one, because the shorter decks were not near any limit.
What fixed it
Not the sentence describing the escape hatch — the sentence forbidding the shortcut. The instruction now says, in plain terms: build the full deck, do not shorten it or thin the content to be safe, a smaller deck is not the safe choice but a worse one — and only if you genuinely hit your limit, finish the slide and hand over. Same safety net, one sentence of pressure in front of it. The decks came back: one model returned to exactly the length it wrote with no safety net at all, another went past it.
Why we are telling you this
Because it is a good example of something that does not show up in any test that only asks “did it work”. Every single run in this experiment produced a complete, valid deck. Nothing errored. The safety net did its job. It just quietly cost a third of the content on the decks that mattered most, and the only way to see that was to run the same brief twice and compare.
How honest this is
Method and limits
3 models (deepseek-v4-flash, gpt-5-mini, mistral-large-2512) on two fixed briefs — one that wants a short deck and one that wants a long one — with three versions of the same instruction. The versions differ by a single paragraph and nothing else: same template, same depth, same everything. Slides are counted from the returned file; completion means the file actually ends.
The limits. One or two samples per cell, so the direction is solid and the exact numbers are not. One control run is missing because a provider timed out, which is why one cell is blank rather than estimated. And a confound we caught before publishing, not after: the first version of this comparison used older runs as its control, made when the example deck we send was larger. A larger example produces longer decks — a thing we had measured the day before — so those controls were inflated by our own change and would have blamed the safety net for a shrink it did not cause. The controls here were re-run against the current prompt.
Nothing here involves anyone’s documents. Every deck was generated from invented briefs, and unPaper itself still sends no document anywhere.