Rhetorical Presentation Can Influence Human Debate Judges
Liu and colleagues’ preprint v1 (2026-10-04) analyzed the choices of 369 people judging opposing AI advocates’ accusations in detective mysteries with known solutions. Four prompted styles used initial framing, flawed logic, technical-sounding language or greater length. Combined across styles, choices shifted 8.6 percentage points toward the manipulated side compared with unmanipulated debates. No individual-style effect was statistically established. Either the correct accusation or one against an innocent suspect could receive the manipulation, so influence need not mean lower accuracy. Candidate dialogues were screened for influence on model judges. These controlled, selected manipulations do not estimate natural frequency or effects in scientific review; there was no comparison with a single AI advisor.
0
1
Tags
Scientific Oversight of AI-Assisted Research
1Cademy