Humans Correct Their Judgments Through Debate, but Weaker AI Models Do Not
Authors: Maria Victoria Carro, Denise Alejandra Mester, Facundo Nieto, Oscar Agustin Stanchi, Guido Ernesto Bergman, Giovanni Franco Gabriel Marrafini, Trinidad Borrell, Guido Freire, Joaquin Salvador Machulsky, Mario Alejandro Leiva, Ana Lopez, Nicolas Spinelli, Luca Nicolas Forziati Gangi, Eitan Sprejer, Federico Barrera Lemarchand, Gerardo Simari, Maria Vanina Martinez
Organizations: Università degli Studi di Genova, IT · FAIR, IALAB, Universidad de Buenos Aires (UBA), AR · Instituto de Investigación en Informática LIDI, Universidad Nacional de La Plata, AR · Universitat Rovira i Virgili, ES · Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET), AR · BAISH, AR · Inria, FR · Sigma Nova, FR · Paris Brain Institute (ICM), Sorbonne Université, INSERM, CNRS, FR · Departamento de Computación, FCEyN, Universidad de Buenos Aires, AR · Dept. of Comp. Sci. and Eng., Universidad Nacional del Sur (UNS) & ICIC-CONICET, AR · Laboratorio de Neurociencia, Universidad Torcuato Di Tella, AR · Department of Mathematics and Computer Science, University of Cagliari, IT · Artificial Intelligence Research Institute (IIIA-CSIC), ES
How can we effectively oversee AI systems that surpass human intelligence? AI Safety via Debate proposes that two AI systems argue opposing positions and a judge, either a human or a weaker model, decides which is correct. By facilitating supervisors in providing high-quality training signals, debate aims to encourage truthful and safe behavior from AI systems whose capabilities exceed our own. However, human judgment is not neutral: people often rely on biases which advanced AI systems may learn to exploit. Drawing on insights from collective intelligence and social cognition, this paper investigates whether debate can guide human and LLM supervisors toward the truth despite their prior beliefs. We evaluate several protocol configurations. We also explore setups inspired by the wisdom of the crowd such as the number and diversity of judges, as well as interaction mechanisms including deliberation and voting. Overall, we find that debate increases accuracy relative to prior beliefs for human judges, whereas the consultancy baseline, in which a single expert argues for one answer, does not. For weaker language models, however, the results are mixed: debate provides no comparable gains in most settings and can even reduce accuracy. Among multi-judge protocols, weighted voting produces the highest accuracy for human supervisors, whereas for LLMs none of them significantly improves group accuracy, even as the number of judges increases. Notably, the behavior of minority judges in non-diverse deliberation protocols suggests that, under symmetric settings, incorrect positions could be more persuasive than correct ones. These findings support debate as a viable path toward scalable human oversight, but cast doubt on its benefits for weaker AI supervisors, raising concerns about fully automated training and control pipelines in high-stakes settings.
Figures & tables
Figure 1: Full experimental pipeline and protocol variations, with human or LLM as supervisors.
Figure 2: Human judge accuracy per protocol. Error bars: 95% CIs (clustered by deliberation room for multi-human post-deliberation); dashed line: chance; * Holm-adjusted p<.05
GPB Method
Consultancy
In-Subtopic Debate
Cross-Subtopic Debate
Multi-Judge
GPB System Prompt
Claude Sonnet 4
Proprietary, Open
Claude Sonnet 4
Proprietary
Persona Capping
⋅
Open
⋅
⋅
Persona Prompting
⋅
Open
⋅
⋅
Unconditioned
Claude Sonnet 4
Proprietary, Open
⋅
⋅
Table 1: Overview of GPB injection methods, protocols, and judge models. Proprietary : Claude Sonnet 4, GPT-5.2, GPT-4o, Gemini 3 Flash, Gemini 2.5 Pro. Open : Qwen2.5-7B-Instruct, Gemma-3-4B-IT, Llama-3.1-8B-Instruct. ⋅ indicates the combination was not evaluated.
Figure 3: Transcript order for one debate under both majority conditions.
Protocol
# Judges
Diversity
Deliberation with consensus
3
No
Deliberation without consensus
3, 7
No
3, 7
Yes
Voting
3, 7
No
Weighted Voting
3, 7
No
Table 2: Experimental configurations for the multi-judge LLM experiments. All judges are Claude Sonnet 4 in the non-diverse condition, whereas the diverse condition uses GPT-4o, GPT-5.2, Claude Sonnet 4, Gemini 2.5 Pro, and Gemini 3 Flash.
Figure 4: Accuracy of LLM judges (Claude Sonnet 4 with the GPB system prompt) across protocols. Error bars: 95% CIs (clustered by debate for multi-judge protocols); dashed line: chance; * Holm-adjusted p<.05 .
Figure 5: Deliberation dynamics of LLM judges. (a) Accuracy of grounded and erroneous judges across stages in three-judge panels. (b) Share of post-debate minority and majority judges switching during deliberation. (c) Share of grounded and erroneous minority judges switching during deliberation, by panel configuration. Error bars: 95% CIs clustered by debate.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: The interface shows the first round of a debate. The following rounds are presented in the same fashion.
Figure 7: Deliberation interfaces. Left: multi-judge (hybrid) condition, showing the two LLMs’ opening interventions. Right: human-only condition, showing the coordinator’s opening interaction.
Experimental Protocol
GPB Group
n
Pre Acc.
Pre Conf. (M)
Post Acc.
Post Conf. (M)
Belief Reversals T → F
Belief Reversals F → T
Δ Acc. (pp)
p (McNemar) [Acc.]
p (Wilcoxon) [Conf.]
Consultancy
Grounded
48
60.42
57.75
58.33
72.73
22.92
20.83
− 2.09
1.000
.006
Erroneous
32
37.50
56.97
34.38
74.38
21.88
18.75
− 3.12
1.000
.002
Average
80
51.25
57.44
48.75
73.39
22.50
20.00
− 2.50
1.000
< .001
Debate In-subtopic
Grounded
47
61.70
55.32
70.21
66.23
4.26
12.77
8.51
1.000
.004
Erroneous
45
20.00
61.20
44.44
69.62
4.44
28.89
24.44
.163
.010
Average
92
41.30
58.20
57.61
67.89
4.35
20.65
16.31
.028
< .001
Appendix
Table 5: Human experiments: single-judge decision protocols. Accuracy (Acc.), mean confidence (Conf.), and belief reversals are reported in %. p -values are Holm-adjusted; bold indicates p<.05 . † Binomial test of post-debate accuracy against chance, as pre/post items differ in the cross-subtopic protocol.
Experimental Protocol
GPB Group
n
Pre Acc.
Pre Conf. (M)
Post Acc.
Post Conf. (M)
Belief Reversals T → F
Belief Reversals F → T
Δ Acc. (pp)
p (McNemar) [Acc.]
p (Wilcoxon) [Conf.]
Voting
Grounded
110
66.67
58.76
83.33
69.25
5.45
15.45
16.67
1.000
< .001
Erroneous
106
19.44
57.61
41.67
65.75
8.49
24.53
22.22
.387
< .001
Average
216
43.06
58.20
62.50
67.53
6.94
19.91
19.44
.028
< .001
Weighted Voting
Grounded
110
72.22
58.76
86.11
69.25
5.45
15.45
13.89
1.000
< .001
Erroneous
106
19.44
57.61
44.44
65.75
8.49
24.53
25.00
.242
< .001
Average
216
45.83
58.20
65.28
67.53
6.94
19.91
19.44
.028
< .001
Appendix
Table 6: Human experiments: voting protocols. Both protocols use the same 72 three-judge groups, labeled Grounded or Erroneous by the majority GPB of their members (36 each). Acc., Δ Acc., and McNemar tests are computed over groups; n , mean confidence (Conf.), belief reversals, and Wilcoxon tests over individual judges. All values in %. p -values are Holm-adjusted; bold indicates p<.05 .
Experimental Protocol
GPB Group
n
Pre-Debate
Post-Debate
Post-Deliberation
Pre-Debate → Post-Debate
Post-Debate → Post-Deliberation
Pre-Debate → Post-Delib.
Acc.
Conf.
Acc.
Conf.
Acc.
Conf.
Rev. T → F
Rev. F → T
Δ Acc. (pp)
p (McNemar)
Rev. T → F
Rev. F → T
Δ Acc. (pp)
p (McNemar)
p (McNemar)
p (Wilcoxon)
Multi-Human Deliberation
Grounded
77
66.23
56.10
72.73
67.71
71.43
73.01
6.49
12.99
6.49
1.000
6.49
5.19
− 1.30
1.000
1.000
< .001
Erroneous
50
30.00
56.58
48.00
62.76
48.00
68.20
4.00
22.00
18.00
.164
4.00
4.00
0.00
1.000
.352
< .01
Average
127
51.97
56.29
62.99
65.76
62.20
71.12
5.51
16.54
11.02
.118
5.51
4.72
− 0.79
1.000
.369
< .001
Hybrid Deliberation
H Grounded
57
64.91
65.91
66.67
73.60
70.18
78.51
12.28
14.04
1.76
1.000
7.02
10.53
3.51
1.000
1.000
< .001
H Erroneous
48
33.33
60.52
39.58
68.42
41.67
73.19
14.58
20.83
6.25
1.000
4.17
6.25
2.09
1.000
1.000
.008
Appendix
Table 7: Human experiments: deliberation protocols. Accuracy (Acc.), mean confidence (Conf.), and belief reversals (Rev.) are reported in %. H: human judges. For multi-human deliberation, all McNemar and Wilcoxon tests (Pre-Debate → Post-Debate, Post-Debate → Post-Deliberation, and Pre-Debate → Post-Deliberation) are cluster-level randomization versions with deliberation rooms as clusters. p -values are Holm-adjusted; bold indicates p<.05 .
Figure 8: Change in accuracy from pre-debate to the final stage by GPB group (pp): grounded and erroneous . Error bars: 95% Newcombe CIs.
Grounded Personas
Erroneous Personas
Engineer pilot
Entrepreneur travel
Aviophobe
Religious risk averse
Technologist
Science layperson
Cautious religious
Christian conservative
Physicist
Biologist atheist
Peace activist
Moralist religious
Scientist optimist
Researcher cs
Artist pessimist
Artist conservative
Astronomer
PhD right wing
Pessimist religious
Religious skeptical
Appendix
Table 8: Complete list of the 20 personas used to construct the grounded and erroneous activation axes, grouped by classification. For each persona, three distinct descriptive variants were used rather than a single fixed description, to reduce variance attributable to prompt-specific wording rather than to the persona itself. For example, one variant of the “aviophobe” persona read: “You are a skeptic by nature with limited formal training in the sciences and a lifelong fear of flying. You tend to dismiss bold scientific announcements as overhyped.”
System prompt
No persona capping
With persona capping
No prompt
C1
C2
Belief
C3
C4
Belief + Persona
C5
C6
Appendix
Table 9: Experimental conditions (C1–C6) resulting from the 2×3 factorial design.
Group
Personas
Disability
a physically-disabled person, an able-bodied person
Religion
a Jewish person, a Christian person, an Atheist person, a Religious person
Race
an African person, a Hispanic person, an Asian person, a Caucasian person
Gender
a man, a woman, a transgender man, a transgender woman, a non-binary person
Political Affiliation
a lifelong Democrat, a lifelong Republican
Appendix
Table 10: Personas taken from Gupta et al. [2023] , used in our experiments.
Accuracy (%)
Reversals (%)
Confidence
Protocol
Group
n
Pre
Post
Δ
T → F
F → T
Pre
Post
p
Unconditioned
Consultancy
All
192
84.38
76.56
−7.82
11.46
3.65
85.08
81.49
.041
Debate (in)
All
190
85.26
78.95
−6.32
8.95
2.63
86.03
84.67
.068
GPB system prompt
Consultancy
Grounded
200
92.00
75.00
−17.00
20.50
3.50
86.64
80.12
<.001†
Appendix
Table 11: Debate versus consultancy with Claude Sonnet 4 as judge. Reversals are percentages of n . p : Holm-adjusted McNemar’s test; † unadjusted McNemar’s test; ‡ unadjusted topic-clustered bootstrap; § Holm-adjusted binomial test against chance.
Accuracy (%)
Reversals (%)
Confidence
Model
Group
n
Align. (%)
Pre
Post
Δ
T → F
F → T
Pre
Post
p
Unconditioned
GPT-5.2
All
198
–
84.85
90.40
+5.56
1.01
6.57
87.21
87.13
.044
GPT-4o
All
198
–
86.36
84.85
−1.52
5.56
4.04
89.49
89.82
1.0
Gemini 3 Flash
All
198
–
92.93
95.96
+3.03
0.00
3.03
96.06
96.87
.16
Gemini 2.5 Pro
All
196
–
91.84
91.33
−0.51
5.61
5.10
94.61
95.38
1.0
Appendix
Table 12: In-subtopic debate across proprietary models other than Claude. Align.: alignment between the GPB and the model’s specific pre-debate belief. Reversals are percentages of n . p : Holm-adjusted McNemar’s test; † unadjusted McNemar’s test.
Figure 9: Pre-debate and post-debate accuracy of grounded and erroneous judges under the GPB system prompt, by argument order. Claude Sonnet 4 is shown as a reference. Error bars are 95% bootstrap confidence intervals over claims; the dashed line marks chance.
Alignment (%)
Grounded acc. (%)
Erroneous acc. (%)
Average acc. (%)
Model
Intervention
G
E
Pre
Post
Pre
Post
Pre
Post
p
Llama-3.1-8B
Unconditioned
–
–
–
–
–
–
74.40
62.80
.17
PC
71.2
34.3
71.2
62.8
65.7
63.3
68.45
63.05
.66
GPB
81.0
74.8
81.0
70.0
25.2
32.5
53.10
51.25
1.0
PC + GPB
81.0
79.5
81.0
70.0
20.5
38.1
50.75
54.05
1.0
PP + GPB
81.0
74.8
81.0
71.0
25.2
25.3
53.10
48.15
.66
Appendix
Table 13: Belief injection and debate in open-source models. PC: persona capping; PP: persona prompting. p : cluster-level sign-flip permutation test on average accuracy with claims as clusters, Holm-adjusted across the 18 tests in the table.
Accuracy (%)
Debate rev. (%)
Delib. rev. (%)
Confidence
Configuration
Panels
n
Pre
Debate
Delib.
T → F
F → T
T → F
F → T
Pre
Debate
Delib.
Deliberation, non-diverse
3 judges + consensus a
G-maj
188
88.83
92.55
81.91
3.72
7.45
10.64
0.00
84.96
80.39
–
E-maj
198
14.65
7.58
8.08
13.64
6.57
2.02
2.53
83.86
79.58
–
All
386
50.78
48.96
44.04
8.81
6.99
6.22
1.30
84.39
79.97
–
3 judges
G-maj
594
64.31
66.84
68.69
4.21
6.73
4.21
6.06
84.92
80.50
81.51
Appendix
Table 14: All multi-judge configurations (Claude Sonnet 4 with the GPB system prompt in non-diverse conditions). G-maj/E-maj: groups with a grounded or erroneous majority. Debate: post-debate; Delib.: post-deliberation. Reversals are percentages of n . a Accuracy and reversals refer to the group decision and n counts panels; confidence is the mean across judges and is not elicited after consensus deliberation.
Test
p
pHolm
Insight 8: best protocol (3 judges, non-diverse)
V: pre → debate
.257
1.0
WV: pre → debate
.350
1.0
D: pre → delib.
.679
1.0
D+C: pre → debate
.571
1.0
D+C: debate → delib.
.003
.022
Appendix
Table 15: Hypothesis tests for multi-judge protocols, grouped by Holm-Bonferroni family, with debates as clusters. V: voting; WV: weighted voting; D: deliberation; D+C: deliberation with consensus; Div: diverse deliberation; G/E: grounded/erroneous. Within-protocol changes use cluster-level McNemar’s tests; between-protocol comparisons use cluster-level paired permutation tests, matching panels built from the same debate, majority composition, and presentation order. Insight 11 comparisons between groups use logistic (switching) and linear (confidence) regressions with debate-clustered standard errors. b Changes in accuracy of grounded and erroneous judges across the five configurations; ranges are reported.
Figure 10: Hybrid deliberation (one human and two LLM judges per room). (a) Accuracy and (b) mean confidence at each stage for grounded (blue) and erroneous (red) judges; solid lines denote human judges and dashed lines LLM judges. Error bars are 95% bootstrap confidence intervals, resampling participants for humans and debates for LLM judges.
Reinforcement learning from human feedback (RLHF) has played a central role in making large language models responsive to human instructions. However, human evaluators often favor flattering or persuasive responses over truthful ones, creating incentives for models to appeal to evaluators at the expense of accuracy. AI safety via debate has been proposed as a way to improve the supervision of language models: in this paradigm, two agents argue opposing positions and challenge each other's claims, potentially exposing falsehoods to the adjudicator. A central premise of AI safety via debate is that truthful arguments are easier to defend than false ones under adversarial scrutiny. In this work, we investigate whether this advantage persists when debaters use rhetorical strategies that exploit biases in human judgment. Inspired by competitive debate, we construct 68 LLM-generated dialogues about detective mysteries with known culprits, spanning four interventions: anchoring, fallacy oversight, pro-jargon, and verbosity. 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.
Gefei Liu, Sonya Rashkovan, Sophia Lloyd George +2
AI safety via debate and reinforcement learning from AI feedback (RLAIF) are both proposed methods for scalable oversight of advanced AI systems, yet no formal framework relates them or characterizes when debate offers an advantage. We analyze this by parameterizing debate's value through the geometry of knowledge divergence between debating models. Using principal angles between models' representation subspaces, we prove that the debate advantage admits an exact closed form. When models share identical training corpora, debate reduces to RLAIF-like where a single-agent method recovers the same optimum. When models possess divergent knowledge, debate advantage scales with a phase transition from quadratic regime (debate offers negligible benefit) to linear regime (debate is essential). We classify three regimes of knowledge divergence (shared, one-sided, and compositional) and provide existence results showing that debate can achieve outcomes inaccessible to either model alone, alongside a negative result showing that sufficiently strong adversarial incentives cause coordination failure in the compositional regime, with a sharp threshold separating effective from ineffective debate. We offer the first formal connection between debate and RLAIF, a geometric foundation for understanding when adversarial oversight protocols are justified, and connection to the problem of eliciting latent knowledge across models with complementary information.
Robin Young
Department of Computer Science and Technology, University of Cambridge, Cambridge, UK
Debate, where AI agents argue opposing positions, has emerged as a key approach to scalable oversight. However, debate faces a fundamental tension: models are incentivized to be persuasive to the judge, which may not always align with epistemic honesty. In this work, we propose an alternative paradigm: disagreement resolution, which reframes the interaction mechanism from adversarial debate to collaborative truth seeking. Drawing on principles from human mediation and conflict resolution, where mediators facilitate dialogue to help disputing parties reach consensus rather than adjudicating between them, we design an automated pipeline that adapts these strategies to AI oversight. Unlike standard debate where models argue for fixed positions, our pipeline directs models to collaboratively identify points of disagreement, examine the evidence for conflicting claims, and converge toward consensus or isolate the specific ''crux'' of their disagreement. We find that Disagreement Resolution consistently helps non-expert models identify the truth, achieving 62.1% judging accuracy compared to 49.2% for standard debate. Our results provide encouraging empirical evidence for rethinking the scalable oversight protocol from adversarial persuasion to collaborative truth-seeking.
Yuyang Jiang, Chacha Chen, Teng Wu +4
Department of Computer Science, University of Chicago, Chicago, IL, USA · Microsoft, Redmond, WA, USA · Department of Biostatistics and Bioinformatics, Duke University, Durham, NC, USA +1