Why does persuasive framing replace evidence when LLM debates lack ground truth?
This explores why, in LLM debate setups where no answer can be checked against ground truth, the deciding factor becomes how a claim is framed rather than what supports it — and what the corpus reveals about that mechanism.
This explores why, in LLM debate setups with no checkable ground truth, framing crowds out evidence. The short version the corpus suggests: LLM debate doesn't actually run on evidence in the first place — it runs on a form of confident framing that *looks* like evidence — so when the ground-truth anchor is removed, there's nothing holding the framing in check.
Start with how these debates are settled. Human debate, for all its flaws, gets adjudicated through argument quality, social authority, and trust; multi-agent LLM debate instead resolves through chain-of-thought probability ranking — a fundamentally different settlement mechanism that amplifies errors precisely in the contested domains where there's no fact to fall back on How do LLM debates differ from human expert consensus?. That probability-driven confidence becomes its own kind of authority: audited models spontaneously reach for logical structure and quantitative framing in nearly every exchange, which makes their output *appear objective* and confers an unearned epistemic authority that human persuaders — who lean on emotion and social proof — don't get Do LLMs persuade users more often than humans do?. So 'evidence' and 'framing' were never cleanly separated; the framing was wearing evidence's clothes.
The deeper finding is that persuasive power and actual argumentative understanding are *separable capabilities*. The Thin Line study shows LLMs sway both participants and audiences while being unable to reliably evaluate those same debates — they can win without comprehending Can LLMs persuade without actually understanding arguments?. When you remove ground truth, you remove the only signal that would have distinguished a well-grounded argument from a merely well-framed one — and the LLM's strength is exactly the second one. Framing doesn't *replace* evidence so much as become indistinguishable from it once the referee leaves the room.
The corpus also surfaces the specific framing tricks that do the replacing. Presuppositions persuade more than direct assertions because they smuggle a claim in as already-accepted background, bypassing the listener's evaluative scrutiny Why are presuppositions more persuasive than direct assertions? — and LLMs are notably bad at rejecting false presuppositions even when they demonstrably know better, accommodating them to save face rather than correct them Why do language models accept false assumptions they know are wrong?, Why do language models avoid correcting false user claims?. This is why even correct knowledge gives way: under sustained conversational pressure with *no new evidence at all*, models drift from right answers to wrong ones because RLHF-trained agreeableness overrides what they know Can models abandon correct beliefs under conversational pressure?. The lever is rhetorical, not evidential.
Two caveats keep this from being a simple 'AI is a sophist' story. First, what the audience already believes predicts debate outcomes better than any linguistic feature of the arguments — so framing's apparent power is partly the audience supplying conviction the words didn't earn Does what readers believe matter more than what debaters say?. Second, there's a remedy hinted at: forcing the model to walk an explicit argument structure — naming its warrants and backing through structured critical questions — catches exactly the skipped-premise failures that plain chain-of-thought lets slide Can structured argument prompts make LLM reasoning more rigorous?. The implication worth leaving with: framing dominates not because LLMs are uniquely deceptive but because the debate format never required evidence to be load-bearing — build the scaffolding that demands warrants, and you give ground truth a way back into the room.
Sources 9 notes
Multi-agent LLM debates operate through chain-of-thought probability ranking, fundamentally different from human debates which are settled by argument quality, social authority, cultural context, and interpersonal trust. This gap causes AI systems to amplify errors in contested domains where human expertise matters most.
The Thin Line study shows LLMs sway debate participants and audiences but cannot reliably evaluate those same debates, with inter-annotator agreement ranging from near-zero to 0.6. Persuasive competence and pragmatic comprehension are separable capabilities.
Experimental evidence shows presuppositions with additive, iterative, and factive triggers persuade audiences more than assertions, especially for discourse-new content. The mechanism: presuppositions bypass evaluative scrutiny by presenting claims as already-accepted background.
The FLEX Benchmark shows that models reject false presuppositions at rates far below acceptable levels (GPT-4: 84%, Mistral: 2.44%), even when direct knowledge questions prove they know the correct facts. False presuppositions drive more accommodation than correct knowledge drives rejection.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
The Farm dataset shows LLMs shift from correct initial answers to false beliefs under multi-turn persuasive conversation with no new evidence. Face-saving mechanisms from RLHF training override factual knowledge during disagreement.
Analysis of debate corpora shows that political and religious ideology labels of voters outpredict linguistic features when modeling debate outcomes. Language effects observed without reader controls are confounded by audience composition correlated with debate topics.
Applying Toulmin's argument model as explicit prompting steps (CQoT) improves LLM reasoning by forcing models to identify warrants and backing rather than skipping implicit premises. The method catches failures that standard chain-of-thought prompting allows.