A “Do Not Guess” Prompt Cut AI-Made-Up Claims from 71% to 20%

Summary

This study tested whether a simple instruction could reduce fabricated values when AI models extract fields from webpages. In the original benchmark, 16 models invented 405 of 573 missing fields, or 70.7%. Adding “Use null for any field whose value is not on the page. Do not guess.” reduced the count to 116 of 574, or 20.2%, and every model improved. A follow-up used the same pages in 42 twin pairs, with one page containing the answer and the other hiding it behind a decoy, and scored 36 missing-field pages per model while excluding email traps. Five models representing the top, middle, and bottom of the benchmark were tested against alternative phrases, including “Make no mistakes,” “Stop bullshitting,” evidence quotes, self-checking, “Do not guess,” and “Use null.” “Make no mistakes” made no meaningful difference, while “Use null” alone reduced fabricated fields by 37 percentage points and “Do not guess” alone by 26 points. The full sentence performed best, with 25.6% fabricated fields, although quoting evidence did not prevent some models from citing decoy text. The results came from one run per condition on synthetic pages, so the authors caution that repeats and tests on real websites are still needed; the experiments cost $2.29 through OpenRouter.