Expertise and Testimony

Almost everything you know, you did not discover yourself. You did not personally confirm that the Earth orbits the Sun, that DNA encodes genetic information, or that the Black Death killed roughly a third of Europe's population. You learned these things from books, teachers, scientists, and journalists who themselves learned from others. This chain of transmitted knowledge is called testimony, and it is the primary vehicle of human knowledge at civilizational scale. Understanding when testimony is trustworthy — and what makes someone a genuine expert worth trusting — is one of the most consequential epistemic skills you will ever develop.

The Epistemology of Testimony

Testimony is knowledge gained by trusting what others report. Philosophers disagree about its epistemic status: some hold that testimony is a basic, irreducible source of justification (the reductionist view), while others hold that you can only justifiably believe testimony if you have independent reasons to trust the testifier (the anti-reductionist view). In practice, most people operate by a mix. For low-stakes claims from plausible sources (a friend says the café is open until 9 PM), we trust without checking — the cost of verification exceeds the cost of being wrong. For high-stakes claims from potentially biased sources (a pharmaceutical company reports its own drug's efficacy), we demand independent corroboration and scrutinize methodology. The calibration question is important: are you as skeptical as you should be, where the stakes warrant? Research consistently shows that people extend credulity based on irrelevant surface features — speaker confidence, physical attractiveness, prestigious institutional affiliation — while under-weighting the actual epistemic credentials of a source. A critical distinction: disagreeing with an expert is not automatically anti-intellectual. Experts can be wrong, expert consensus can be mistaken, and domains vary enormously in the reliability of expert judgment. The question is not 'is this person credentialed?' but 'do the credentialing processes in this domain actually track accuracy?'

What makes someone an expert, epistemically speaking? Genuine expertise has several markers that go beyond credentials. Track record: has the person or institution made accurate predictions and conclusions in this domain over time? A doctor who has correctly diagnosed thousands of similar cases has an empirical track record. A pundit who has made political forecasts for ten years with no verified accuracy does not. Peer accountability: is the person's work subject to scrutiny by others with the knowledge to evaluate it? Peer review in science, adversarial collaboration in law, and formal audit in accounting all create mechanisms by which errors can be caught and corrected. Calibration: does the expert acknowledge what they do not know? Genuine experts are typically more hedged than non-experts precisely because they know how much uncertainty exists in their field. Confident certainty from an expert about complex dynamic questions should trigger suspicion, not reassurance. Alignment of incentives: does the expert have financial, ideological, or reputational incentives to report a particular conclusion? A nutritionist paid by a sugar company has structurally different incentives from one independently funded. This does not make their claims false, but it is a legitimate factor in evaluating testimony. Domain specificity: expertise is narrow. A Nobel-Prize-winning physicist is an authority on physics; their opinion on monetary policy is not expert testimony on economics. Celebrity, general intelligence, and fame do not transfer expertise.

How AI Disrupts the Signals of Expertise

Historically, you could use several signals to distinguish genuine expertise from its simulation: formal credentials, institutional affiliation, peer-reviewed publication record, the texture of genuine uncertainty in speech, and the ability to respond to pointed follow-up questions. AI disrupts most of these signals. Credentials and affiliation: AI systems can produce text that mimics the style of credentialed experts, include plausible-sounding citations, and describe affiliations that do not exist. The surface features of expert communication are reproducible by any sufficiently large language model. Publications: with AI assistance, the volume of synthetic academic-style writing has increased dramatically. Many AI-generated papers contain fabricated citations — references to papers that do not exist, written in the correct bibliographic style with plausible authors, journals, and volume numbers. Checking whether a cited paper actually exists and says what it is claimed to say is now a necessary verification step. The texture of genuine uncertainty: expert speech often includes specific, domain-appropriate hedges — 'the data are consistent with X but we cannot rule out Y because of confounding from Z.' AI can mimic this texture without the underlying epistemic structure that produces it. The hedge sounds right but may not correspond to genuine uncertainty about a genuine limitation. Follow-up questioning: one reliable test of genuine expertise is whether a source can respond substantively to pointed, unexpected follow-up questions. AI passes this test impressively for well-covered topics and fails dramatically for obscure or novel ones — but users rarely know in advance which regime they are in.