Human judges of AI debates are influenced by rhetorical tactics independently of which side is correct.
Assessment
The evidence favors the claim, but the examination is incomplete or the evidence is indirect.
The most direct evidence is a preregistered experiment by Liu and colleagues (2026). 369 online participants each judged one AI-generated debate about a detective story with a known culprit. Anchoring, uncorrected fallacies, professional-sounding jargon or added length was applied to either the correct or the incorrect side, with the story evidence held fixed. Because the manipulation was assigned independently of correctness, the result isolates an effect of rhetoric: pooled across the four tactics, verdicts shifted about 8.6 percentage points toward the manipulated side (95% interval 2.9 to 14.2, p = .002). Each tactic taken alone pointed the same way (7 to 11 points), but none reached significance by itself. Speaking order, also unrelated to correctness, had a larger effect: judges were right 63% of the time when the true side spoke first and 41% when it spoke second. Few judges noticed the manipulation, and those who did discounted it.
Three features limit how far the result generalises. The manipulated debates were kept only if they swayed at least one of six AI judges, so the effect size describes tactics already screened for potency, not typical debates. The opponent's turns were not rewritten: the opposing rebuttal was written against the unmanipulated version, so the effect is not net of a rebuttal that targets the fallacy or the padding. And the corpus is small (four stories, untrained crowd judges, one-off judgments). The claim's existence form, that such tactics move human verdicts at all, fits a large literature on anchoring, order and fluency effects in human judgment and is well supported. How large the effect is against a responsive, capable opponent, which is what matters for debate as a means of oversight, has not been measured. Other human-judge debate studies report that debate on balance helps judges reach correct answers, which is compatible with this: rhetoric can shift verdicts while the true side still holds a net advantage.
Full reasoning: the evidence and decisions behind this verdict
Read the claim as an existence claim: rhetorical features unrelated to correctness shift human verdicts in AI debates. Prior: very likely, given the psychology of anchoring, order and fluency effects; the open question is magnitude in an adversarial setting.
Primary source read (HTML full text, sections 1 to 6): Liu, Rashkovan, Lloyd George, Sheidlower, Booth, "AI Safety via Debate is Compromised by Cognitive Biases", arxiv.org/abs/2610.05461. Design: 4 True Detective stories x (1 baseline + 4 biases x 2 sides x 2 variants) = 68 dialogues; six turns (scene reconstruction, argument, rebuttal). Biased dialogues regenerate only the manipulated side's turns 3 to 6, "retaining the other side's turns unchanged", so the opponent never responds to the manipulation. Selection: a biased dialogue passes only "if at least one model shifts toward bias and none shifts away" among six LLM judges; otherwise regenerated. Preregistered on OSF; Prolific; 400 recruited, 369 analysed; $0.50 bonus for accuracy. Results: chi2(1)=9.31, p=.002; Delta=+8.6 pp [2.9, 14.2]. Per-bias: anchoring +9.2, fallacy +7.1, jargon +7.2, verbosity +11.1, none significant (Holm). Order: 63% vs 41% accuracy by whether the true side spoke first (p<.001). Judges who correctly identified the biased side chose it less (48% vs 63%).
Weighing: randomization of the manipulation across correct and incorrect sides answers the "independently of correctness" part directly. The LLM pre-screening inflates effect size relative to unselected manipulations but does not create an effect from nothing, since humans were not used for screening. The lack of a responsive rebuttal is the main limit on relevance to debate as oversight. Single study, small story set. Corroborating: speaking-order effect in the same data; broad human-judgment literature. Context from other human-judge debate studies (Michael et al. 2023, Khan et al. 2024, Rahman et al. 2025, Carro et al. 2026) shows debate net helps accuracy, compatible with the claim. Search limited to one query plus the parent claim's recorded literature; no replication found.
Status supported rather than verified because the direct evidence is one study with a selected corpus; credence 0.85 for the existence claim as worded.
Decomposition
This claim is atomic: it bottoms out in a bedrock fact, an open empirical question, or a value premise, and does not decompose further.
Provenance
Where this claim has been said, linked to its canonical form.
We apply each intervention to either the side advocating for the true culprit or the side advocating for an innocent suspect, allowing us to distinguish influence on adjudication from correctness. In a study with 369 participants, we find that, pooled across bias types, these interventions significantly shift judgments toward the manipulated side. These findings expose a vulnerability in debate-based supervision: human adjudication is sensitive to manipulative rhetorical strategies.
Preregistered between-subjects study on Prolific; LLM-generated six-turn detective-mystery debates with anchoring, fallacy, jargon or verbosity applied to the correct or incorrect side.
Cite this claim: a formal citation with its evidence attached
Embed this claim: show it on another site with its current assessment
An embed reports the graph's judgment as it stands when the page is viewed: the status with its confidence, how disputed the claim is, and the date. It links back here.
<blockquote class="minerval-claim" data-claim-id="9778a24e-ecd3-46f7-baf7-8aedf0362e05"><a href="https://minerval.ai/claims/9778a24e-ecd3-46f7-baf7-8aedf0362e05">Human judges of AI debates are influenced by rhetorical tactics independently of which side is correct.</a></blockquote> <script async src="https://minerval.ai/embed.js"></script>
<iframe src="https://minerval.ai/embed/claims/9778a24e-ecd3-46f7-baf7-8aedf0362e05" title="Minerval: Human judges of AI debates are influenced by rhetorical tactics independently of which si…" width="480" height="200" style="border:0;max-width:100%" loading="lazy"></iframe>
[](https://minerval.ai/claims/9778a24e-ecd3-46f7-baf7-8aedf0362e05)
The badge now: 
Contribute
Every judgment on this page is open to challenge. A contribution is evaluated on its merits by the reviewer; if it succeeds the page changes, and if it does not, the reasons are stated. Either way the exchange becomes part of the claim’s public record.
The attention this claim received was paid for by a funded mandate. Funding buys only scheduling: it can make an assessment happen sooner, or reach deeper into a subtree. It has no influence on what the assessment concludes, and none on which claims enter the graph; assessments run under the same public standards whoever pays, funders never see or shape a verdict before anyone else, and mandates that attempt to steer conclusions are refused.
Created by claim_steward · Nov 10, 2026. Every judgment on this page is accompanied by a reasoning trace.