The Form Is the Hallucination: Why Required Fields Make AI Invent Answers

Ask the same model the same question twice: once in prose, once through a JSON schema that requires a field to be filled. Ten of thirteen tested language models will honestly say "I don't know" the first time and confidently invent an answer the second. A new public benchmark, PhantomFill, isolates that inversion across 13 models and shows the form, not the model, is the active agent.

GPT-5.5 abstains 98% of the time when asked in free text about a 12,400-like social-media post with no visible replies. Pose the same question through a schema with a required sentiment enum, and it fabricates 40 of 40 times. The PhantomFill authors test 9 open-weight models across 5 families plus Claude Haiku 4.5, Sonnet 4.6, Opus 4.8, and GPT-5.5. Required-field prompts drive 100% fabrication in 10 of 13: all 9 open-weight models plus GPT-5.5. The escape hatch, an insufficient_evidence value the model can return, rescues only frontier models (GPT-5.5 near 100%, Opus 91%); open-weight models still fabricate 60–100%, and Sonnet fabricates 90%.

Fabrication concentrates where hedging is impossible. GPT-5.5 fabricates a required sentiment enum 20/20 times but invents customer quotes 0/20 times, writing disclaimers into the string field instead. Resistance is trained, not scale-driven: within the Claude family, Haiku's 5% rate comes from 100% format violations, not honesty.

An explicit insufficient_evidence value on every field works for frontier models (GPT-5.5 escape rate near 100%, Opus 91%) and fails for open-weight models and Sonnet, which still fabricate 60–90%. The code, scoring script, and leaderboard are public; no independent replication yet.