The better models get, the more awkward one question becomes: if a person can’t check the answer themselves, how do they know the AI got it right? That’s the scalable oversight problem, and one of the oldest proposed fixes is debate. Two AIs argue opposite sides, and a less capable judge picks the winner.

In October, two arXiv papers tested this idea about two weeks apart. One says humans correct their judgments through debate. The other says humans get swayed by rhetoric. The titles sound like they contradict each other. Read the experiments and they look more like two sides of the same picture.

What debate was supposed to deliver

The idea goes back to “AI safety via debate” by Geoffrey Irving, Paul Christiano and Dario Amodei in 2018. Instead of asking humans to evaluate complex behavior directly, two agents trade short statements and a human picks whichever gave the most true, useful information. The theoretical pitch was that a judge who can’t solve a problem could still referee a debate about it, and that this reaches much harder questions than direct judging. The intuition underneath: a lying debater gets called out by the other side, so the truth should be easier to defend.

There’s been encouraging evidence since. In Khan et al. (2024), debate raised non-expert model accuracy from 48% to 76% and human accuracy from 60% to 88%. The October papers poke at that optimism from different angles.

Paper one: humans improved, AI judges were all over the place

Carro et al. (arXiv 2610.28539, Oct. 20, 17 authors) built 100 questions from 10 science and tech topics the public is unlikely to know well, things like mirror life, with 10 subtopics each. Claude Opus 4.5 did the debating. The judges were 507 people recruited on Prolific plus eight LLMs, adding up to more than 75,000 LLM judgments.

The human results are clean. After debate, accuracy rose from 41.3% to 57.61% (p = .028). Consultancy, where a single AI argues one side and the judge asks questions, went from 51.25% to 48.75%, a slight drop. The biggest gain came from confidence-weighted voting across several people: 45.83% to 65.28%.

LLM judges varied a lot by model. GPT-5.2 improved significantly, from 84.85% to 90.4%. Claude Sonnet 4 went the other way after debate, from 85.26% to 78.95%. No open-source judge improved significantly, and Gemma-3-4B dropped from 72.3% to 47.5%. “Weaker” in the title isn’t a fixed bar. It means judges less capable than the debater.

The panel experiments are the most interesting part. Three copies of the same setup forced to reach consensus got worse, 48.96% to 44.04% (p = .022). Seven diverse judges deliberating went from 58.26% to 64.29% (p = .006). In homogeneous three-judge panels, a minority holding the right answer abandoned it 25.8% of the time, while a minority holding the wrong answer switched to the right one only 17.7% of the time. The authors note that in those panels, wrong positions seemed at least as persuasive as correct ones.

They’re upfront about limits. The capability gap between debater and proprietary judges was small, which may explain the mixed results. The task tests knowledge more than multi-step reasoning, and domain experts didn’t verify the statements.

Paper two: the same humans fell for framing

Liu et al. (arXiv 2610.05461, Oct. 4, five authors) built 68 debate dialogues from four detective stories with known culprits. They planted one of four rhetorical tricks on one side: anchoring, sneaking in logical fallacies, piling on jargon, or just being verbose. Each trick went on the culprit’s side in some versions and the innocent suspect’s side in others, which separates the trick’s pull from whether that side was right. Each of 369 Prolific participants judged one debate, and the study was preregistered.

Pooled across tricks, the manipulated side gained 8.6 percentage points (95% CI 2.9 to 14.2, p = .002). Verbosity had the biggest individual effect at 11.1 points, but once the sample was split four ways, none of the tricks stayed significant on its own after correcting for multiple comparisons. The more worrying number: only 30% of participants spotted which side was manipulated, and 11% named the trick.

The biggest effect in the paper wasn’t a planted trick at all. It was speaking order. Judges were right 63% of the time when the truthful side spoke first and 41% when it spoke second. Explanations that leaned on physical evidence were right 77% of the time, versus 46% for the rest. The authors suggest structured criteria and checklists that point judges at evidence, while noting they haven’t tested those fixes yet.

Putting them side by side

These papers don’t actually refute each other. In paper one, debate lifted humans about 16 points on average, but it lifted them to 57.61%. That’s still wrong on roughly four in ten questions, and a long way from Khan et al.’s 88%. The tasks differ, so it isn’t a direct comparison, but debate’s payoff clearly depends a lot on the task.

Doing the math on paper two, the speaking-order effect (22 points) is about 2.5 times the combined effect of all four rhetorical tricks (8.6 points). What moves judges most may be the debate format itself, not deliberate manipulation. Paper one ran every debate in both orders, which reads as a design choice meant to cancel out exactly this kind of bias.

So the combined picture: debate helps human judges on average, the help is still modest, and it’s easy to knock around with format and framing. Swap the human for a weaker AI judge and even that average gain isn’t guaranteed. That’s why the paper one authors warn against assuming debate’s benefits carry over to AI supervisors, especially in fully automated, high-stakes pipelines.

Why this matters if you build with LLMs

This isn’t just a far-off safety question. A lot of agent tooling and eval pipelines already have LLMs grading LLMs: review bots checking another model’s output, or several agents voting on a final answer.

A few things carry over directly. Majority vote or consensus among copies of the same model didn’t raise accuracy, and forcing consensus lowered it. Mixing different judges is what helped. In a judge prompt, scores can drift toward whichever answer is longer or comes first, so it seems safer to run comparisons in both orders and have the judge write out the evidence before the verdict. To be clear, that’s our read of how these results translate to practice, not something either team tested.

Both papers are arXiv preprints and haven’t been peer reviewed yet. Paper one’s panel experiments are explicitly labeled preliminary, and paper two works in the narrow domain of detective stories. Still, two independent experiments from the same month pointing at the same weak spot is hard to wave away.