structured-output json-schema hallucination benchmark production-risk

Required JSON fields force language models to hallucinate

Schema structure—not model capability—drives fabrication; GPT-5.5 invents sentiment values when JSON requires them, stays honest in free text.

Summary

Production LLM integrations rely on structured output (JSON, enums, required fields). If your schema forces hallucination regardless of model scale or instruction, validation and fallback patterns become non-negotiable for any extraction or form-filling pipeline.

Why it matters

Implementation verdict

Replaces false confidence in model honesty under schema constraint. Requires: audit of required vs. optional fields in your structured outputs, testing with PhantomFill benchmark, one-line schema fix (explicit null/abstain option). Worth trying now—this is a measurement problem with a known mitigation.

Sources

1.Given a required JSON field for sentiment, the same model invents an answer 40 times out of 40.
2.Required fields drive fabrication to 100% in ten of thirteen models.
3.A direct instruction, do not infer sentiment, is overridden by the schema in four of six models.
4.Honesty under format pressure is a training outcome that no one is measuring.

Dev Signal

Get briefs like this in your inbox — free, every weekday.

100+ sources compressed into one 4-minute read. Ranked, cited, implementation-ready.

Read the full issue →All briefs