Tag
A study showing that language models hallucinate when required to fill structured fields like JSON, even when they would honestly abstain in free text. The PhantomFill benchmark measures coerced fabrication rates.