Can a claim borrow authority from its packaging — confident phrasing, tidy citations — instead of actually earning it with evidence?

This explores whether packaging can stand in for evidence when a topic is contested and readers can't easily check the facts. The corpus suggests it often can. None of these notes pit framing against evidence head-to-head, but together they show that persuasion runs on cues that are cheap to produce and only loosely tied to whether a claim is true. One example is a grammatical trick. Presuppositions ('when did you stop doing X?' style phrasing) persuade better than plain assertions when the information is new to the audience, because they present a claim as background everyone already accepts, so the reader never stops to evaluate it (Why are presuppositions more persuasive than direct assertions?). Citations work the same way. Across 24,000 search interactions, users preferred answers with irrelevant citations almost as much as answers with relevant ones, so the count works as a trust signal in its own right (Do users trust citations more when there are simply more of them?).

But framing isn't a universal lever, because the audience matters at least as much as the speaker. In debate corpora, a voter's political or religious ideology predicted who won better than the debaters' language did, and language effects that seemed to appear without controlling for the audience were partly confounded by who was listening (Does what readers believe matter more than what debaters say?). So in contested domains, a polished frame often does less to change minds than to confirm what the reader already believed. Recommendation feeds add another layer. They pick which frames reach which people, which turns framing into persuasion infrastructure at population scale (How do recommendation feeds shape what people see and believe?).

LLMs make the substitution problem sharper. In an audit of five models, they used logical appeals and quantitative framing in nearly every conversation, whether or not persuasion was called for. Humans given the same prompts persuaded less often and leaned on emotion and social proof. That makes the models sound objective, which gives them authority they haven't earned (Do LLMs persuade users more often than humans do?). Some of the persuasive power also seems detached from truth. Claude beat incentivized humans at both truthful and deceptive persuasion, and DeepSeek beat them only when arguing for falsehoods (Do large language models persuade better than humans?). A meta-analysis found that model family, one-shot versus multi-turn format, and topic domain together explain about 82% of the variation between studies (What combination of factors explains differences in LLM persuasiveness?). Those are all features of delivery and setting rather than of the claim itself. One limit is that the persuasive edge fades: Claude and DeepSeek lost ground over repeated rounds with the same person, while human persuaders stayed steady (Does AI persuasiveness fade across repeated conversations with the same person?).

Contested domains are where this matters most, because there evidence alone doesn't settle things. Human experts resolve disputes through argument quality, reputation, cultural context and trust. LLM debates instead rank options by probability, a gap the authors say lets AI systems amplify errors where human expertise counts most (How do LLM debates differ from human expert consensus?). Models also can't tell an expert's argument from a commonly held assumption, since they see the text but not the track record behind it (Can language models distinguish expert arguments from common assumptions?). They are open to being pushed around by framing too. Emotional tone changes what a model says: negative prompts got neutral-to-positive answers about 86% of the time (Does emotional tone in prompts change what information LLMs provide?). Multi-turn manipulative prompts cut reasoning-model accuracy by 25 to 29 percent (Why do reasoning models fail under manipulative prompts?).

The unexpected part is that the frames that substitute for evidence best are the ones that look most like evidence: numbers, logic, citations. Readers already read those as trust cues, and an AI can produce them in every answer without any of the social accountability that makes expert claims costly to make.