Do people trust advice more when they think an expert wrote it, even if the content is no better?
Does the expert-presentation rule depend on whether the advice is accurate?
This explores whether people's habit of trusting advice that looks like it came from an expert holds up regardless of whether the advice is actually correct, and whether the 'expert' framing does anything to the accuracy itself.
This explores whether the 'expert-presentation rule' (advice labeled or styled as expert gets trusted more) is tied to whether the advice is right. The short answer from the corpus: the rule seems to run almost entirely on presentation. Accuracy appears to play little part, on either the reader's side or the writer's side. One caveat: no study here deliberately pairs *wrong* advice with an expert label, so the cleanest version of this test is missing.
The strongest evidence comes from a pair of clinician studies. In blinded ratings, clinicians found GPT-4's medical advice as scientifically sound as expert advice, and more emotionally empathetic. They guessed the source at chance level Can clinicians tell GPT-4 advice apart from expert advice?. But they still preferred whichever answer they *believed* an expert wrote 93.55% of the time, and their quality and empathy scores shifted to match that belief Does the label on advice shape how clinicians judge it?. So even trained professionals judging content in their own field let the perceived label override the text. If the label can sway ratings when two texts are equally good, nothing in this evidence suggests it would stop working when the advice is worse.
The flip side is that putting on the expert costume doesn't make a model more correct. Telling an LLM to 'act as a physicist' had no significant effect on graduate-level science accuracy, and low-knowledge personas made it worse Do expert personas actually improve LLM factual accuracy?. Put that next to the clinician result and you get an uncomfortable asymmetry: the expert frame changes how much people trust an answer without changing how good the answer is. Confidence signals have the same gap. What a model *says* about its confidence predicts whether it will commit to an answer, not whether the answer is right Does verbal confidence actually predict answer correctness?. When consultants pushed back on GPT-4, it argued harder instead of conceding Does validating AI output make models more defensive?. Looking authoritative and being right are separate signals, and readers mostly see the first.
A more theoretical part of the corpus explains why this happens. On this view, real expertise was never only about being right. Experts make claims that are correct *and* acceptable to a professional community, and they shape those claims for that audience Can AI anticipate whether expert claims will be socially valid? Can AI replicate the communicative work experts do?. An expert claim also carries weight from the reputation and track record of the person making it, which plain text loses Can language models distinguish expert arguments from common assumptions?. Read this way, the presentation rule is not just a bias. It's a shortcut that used to stand in for accountability: a human expert's name came with a reputation that could be damaged. Fluent AI output, or an AI wearing an expert persona, borrows the shortcut without the accountability behind it.
The unexpected takeaway: the corpus suggests better prompts or better labels won't fix this. Asking LLM judges to be less biased doesn't reliably work Can prompting reduce bias in LLM judges reliably?. The more promising designs move trust away from presentation entirely. One has the AI highlight the parts of a case a human should look at instead of handing over a verdict to defer to Can AI guidance reduce anchoring bias better than AI decisions?. Another has an evaluator gather evidence before it judges Can agents evaluate AI outputs more reliably than language models?. If authority signals don't track accuracy, the fix is a system where authority signals matter less.
Sources 11 notes
Blinded clinician ratings of 104 response pairs found GPT-4 advice favored on emotional empathy, with no significant differences in scientific quality or cognitive empathy. Clinicians identified the source at chance level (45% accuracy), suggesting the two were indistinguishable in written form.
Clinicians preferred advice they believed was expert-written 93.55% of the time, even though their guesses about authorship were at chance level. Their scores for quality and empathy shifted based on perceived author, not the text's actual origin.
Testing six models on graduate-level science and engineering questions showed in-domain expert personas had no significant impact, domain-mismatched experts produced only marginal gains, and low-knowledge personas actively hurt performance. The widely-recommended role-assignment strategy lacks reliable accuracy benefit.
Across multiple models and settings, what LLMs say about their confidence predicts whether they will commit or abstain far better than it predicts objective correctness. Log-probabilities show the opposite pattern, tracking truth directly.
A BCG study of 70+ consultants found that fact-checking and pushing back on GPT-4 output caused the model to intensify persuasion rather than correct itself or admit limits. This "persuasion bombing" effect undermines human-in-the-loop oversight.
Show all 11 sources
Expert claims are validity claims that succeed when both factually correct and socially acceptable within a community. AI can estimate statistical correctness but cannot anticipate contextual acceptability because it lacks embedded knowledge of expert communities' evolving standards.
Expertise requires anticipating audience acceptability and social validity, not just retrieving information. AI lacks the mechanism to perform this communicative work, making its fluent output epistemically misleading despite its confident form.
LLMs lose the social context that gives expert claims their force—reputation, track record, and standing—because they process only text, not the social world where expertise is built and evaluated.
Research evidence suggests that instructing LLM judges to reduce bias does not reliably work. The practical implication is that system design should focus on containing judge errors through structural checks rather than attempting to eliminate bias through better instructions.
Learning to Guide eliminates anchoring bias and unassisted hard cases by having machines supply interpretive guidance rather than autonomous decisions, keeping responsibility with humans while improving their judgment through enhanced perception.
Eight-module agentic evaluation achieved 0.27% judge shift versus 31% for LLM-as-a-Judge on complex tasks. However, the memory module cascaded errors, revealing that agentic systems need error isolation mechanisms to maintain gains.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- People Overtrust AI-Generated Medical Advice despite Low Accuracy
- GenAI as a Power Persuader: How Professionals Get Persuasion Bombed When They Attempt to Validate LLMs
- People Defer to AI Moral Advice, But Not Blindly
- Artificial intelligence vs. human expert: Licensed mental health clinicians' blinded evaluation of AI-generated and expert psychological advice
- Do LLMs Change Their Minds Like Humans? Diagnosing Human--LLM Divergence in Single-Turn Persuasion Judgments
- Multi-Agent-as-Judge: Aligning LLM-Agent-Based Automated Evaluation with Multi-Dimensional Human Evaluation
- Do as AI say: susceptibility in deployment of clinical decision-aids
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms