AI / ML · 2026
When Models Won’t Say “I Don’t Know”
A completed, pre-registered study of epistemic abstention under structured output. Force a frontier model to fill a required answer field and honesty collapses — Claude Haiku’s abstention falls from 76% to 2%. An explicit escape hatch only partially repairs it.
The question
Structured output is everywhere — APIs now let you force a model's response into a JSON schema, and production systems increasingly demand it. But a schema with a required answer field leaves no room for the most important thing a model can say: “I don't know.” Does enforced structure suppress epistemic honesty? And if you explicitly provide an abstain affordance — a nullable answer plus a status field — do models actually use it?
The design
Pre-registered
Hypotheses, arms, exclusion rules, and the analysis plan were written and frozen before headline data collection — the way you keep yourself honest about your own honesty study.
Five arms, within-item
Free-form text → prompted JSON (no enforcement) → API-native strict schema with a required answer → strict schema with a nullable answer + status enum → the same with the status field ordered first.
600 items, contamination-resistant
300 epistemically unanswerable questions (false premises, fabricated entities, missing context, genuinely unknown answers) + 300 matched answerable controls — mixing legacy anchors with freshly minted items verified to have zero search hits, shipped with a canary GUID.
Four pinned models, three judges
Four frontier models across three closed families plus a grammar-constrained open model; 12,000 graded responses; a three-judge panel from three model families with Fleiss κ = 0.93 — all three disagreed on just 7 of 2,400 responses.
What it found
Enforced required-answer schemas collapse abstention in every model tested — risk differences from +0.37 to +0.74, all p < 10−24. The starkest case: Claude Haiku abstains on 76% of unanswerable questions in free-form text, and 2% under a required schema. Providing an explicit abstain option (nullable answer + status enum) recovers much of the behavior — but restores parity with free-form honesty for only one of four models. Putting the status field first shows a marginal pooled benefit (+2.8 points, p = .06).
The one-sentence version: a required answer field is a hallucination machine, an escape hatch is a partial fix, and the size of “partial” is a property of the model.
Status, honestly
The study is complete: data collected, analysis run, paper drafted, artifact bundle (items, graded responses, judge panel outputs) prepared for release. It is not yet peer-reviewed — the preprint upload is the next step. I'm publishing the findings here the same way the study was run: pre-committed, with the caveats attached.
And it isn't just a paper to me — the finding is baked into things I ship. Ask This Site abstains by design and is eval-tested for it; Honest treats “refuses when it should” as a first-class check.