PhantomFill: Required JSON fields make LLMs fabricate answers 100% of the time

A new paper titled "PhantomFill: When the Form Demands an Answer, Language Models Invent One" (arXiv:2607.20492) by Rana Muhammad Usman reveals a critical flaw in production LLM behavior: the output format itself causes hallucination. The study tested 13 models — including GPT-5.5, Llama 4, and open-weight variants — on inputs designed to be unanswerable (e.g., a viral post with 12,400 likes but no visible replies, a support ticket with no transcription). In free-text mode, GPT-5.5 honestly states there is no reply data 98% of the time. But when the same input is presented with a required JSON field for sentiment, the model fabricates an answer 40 out of 40 trials — inventing the mood of crowds and quoting customers that never existed. The pattern holds across the board: required fields drive fabrication to 100% in ten of thirteen models. Even providing an explicit "insufficient evidence" option only helps frontier models; all nine open-weight models ignore it. A direct instruction to not infer sentiment is overridden by the schema in four of six models. Interestingly, honesty under format pressure does not scale linearly — within a single model family, the smallest model refuses, the mid-size fabricates, and the largest refuses again, suggesting it's a training outcome nobody is measuring. The fabrication hides exactly where hedging is impossible: in required enums and minimum-count arrays — fields where no disclaimer fits. The authors release the PhantomFill benchmark with deterministic scoring and two reportable metrics: Coerced Fabrication Rate and Escape Utilization Rate. The fix they test is a single line of schema, but the failure they measure is everywhere, posing serious risks for production systems that rely on structured outputs from LLMs.