Calling the AI bluff: Adding 'Do not guess' cut made-up claims from 71% to 20%

A post titled “Calling the AI bluff: Adding ‘Do not guess’ cut made-up claims from 71% to 20%” was shared on Hacker News on September 27, linking to earnanhonestdollar.com/bench. The post’s title states that adding an instruction not to guess reduced the rate of fabricated claims from 71 percent to 20 percent.

The supplied material does not specify what test or evaluation produced these figures, which models were involved, or who conducted the work. No methodology, sample size, or publication details accompany the claim beyond the title itself and the linked page.

The accompanying note describes the finding as significant for the study of AI reliability, characterizing the reduction in fabricated claims as a substantive result in that field. No further elaboration on the mechanism, the comparison group, or the conditions under which the reduction was measured is given in the material provided.

Readers seeking the underlying data, test design, or authorship would need to consult the linked page directly, as those details fall outside what this bulletin can confirm from the sources at hand. The claim, as stated, concerns a drop in fabricated output following the addition of a specific instruction, with the numeric change from 71 to 20 percent the sole figure reported in the title.