Liu et al.’s preprint (2026-10-04) analyzed 369 people judging opposing AI advocates’ accusations in detective mysteries with known solutions. Four prompted manipulation families favored initial framing, flawed logic, technical-sounding language or greater length. Pooled across families, choices shifted 8.6 percentage points toward the manipulated side. No individual-family effect was statistically established. Either the correct accusation or an accusation against an innocent suspect could receive the manipulation; this measures influence, not a uniform accuracy decrease. Candidates were screened for influence on model judges. The controlled, selected manipulations do not estimate natural prevalence, effects in scientific review or debate’s inferiority to consultancy.

Evidence status: controlled_primary_human_manipulation_study_preprint.