AIHacker Noonabout 8 hours ago
A Language Model Can Be Honest in Prose and Still Fabricate in JSON
7 min read
A researcher found that language models fabricate answers when required JSON fields leave no room for abstention. GPT-5.5 correctly reported missing evidence 98% of the time in prose but fabricated in 40 of 40 trials under a required-field schema. A linter, phantomfill-lint, was released to flag such schemas.
Level
Hype check
Tap to vote and see what everyone thinks.
