Brief

How can you tell? — the ordinary question that names the whole practice, where “telling” does double duty: telling apart (discerning real from counterfeit) and telling true (an account that answers to the world). The thesis is that discernment is a skill, not a rule: there is no detector, no watermark, no single test that settles authenticity from the outside, and the search for one is itself a category mistake about what telling is. What there is instead is a trained perception with identifiable moves. Attend to grounding — does the thing answer to a source, is it checkable, can you follow it back to something that could have been otherwise. Refuse the fluency heuristic — smoothness, coherence, and confidence are proxies for truth that the synthetic now produces for free, so the very cues we evolved to trust are exactly what the machine counterfeits cheapest. Read the tell — the conjectural paradigm of the clue (Ginzburg’s Morelli and Sherlock Holmes): the real reveals itself at the margin, in the telling detail an account cannot fake because it never touched the thing. The capacities behind all this are old and well-studied — tacit knowledge and the expert’s trained eye (Polanyi), recognition built from immersion (Klein, Dreyfus), the vigilance machinery that checks source and content (Sperber et al.), and the disciplined oscillation between charity and suspicion in how we read anyone at all (Ricoeur). In an AI age the practice has to migrate to what cannot be counterfeited: answerability to the world, embodiment, staked commitment. This is the payoff of the foundation and the epistemology — the checkable work that being true to things actually consists of — and it hands directly to Bullshit, the thing telling is for.

Outline

Source ownership: this essay owns the fluency heuristic — Reber & Schwarz, Alter & Oppenheimer. The Answer Machine and Bullshit both run the same mechanism at scale and cite this essay rather than the papers. Grice and Frankfurt cut — both belong to Bullshit, which this essay hands to rather than argues with. Dreyfus & Dreyfus is owned here (recognition and the expert’s eye); The Answer Machine uses the same text for a different move (relevance won’t reduce to rules). Merleau-Ponty carries the embodiment claim in the closing move, which was asserted in the Brief and previously unsourced.

Hackenburg et al. is read here for one move only — the fact-density result, which this essay owns. It is the strongest rival to the essay’s own positive prescription: the Brief tells the reader to refuse the fluency heuristic and attend to grounding (is it checkable, can you follow it back), and Hackenburg finds that persuasive impact scales log-linearly with the sheer count of fact-checkable claims, on one regression line fitting models whose factual accuracy differed widely. Checkable-looking is now the cue that is cheapest to mass-produce, and volume of it is what moves people — so the recommended cue is counterfeitable in the same way fluency is, and the essay has to answer that rather than close on the prescription. It sits with the detection cluster (DetectGPT, watermarking, Sadasivan) and extends it from can you detect the machine to the mark of grounding is itself forgeable. The headline out-persuasion result is owned by Letting Strangers In; cite it through that essay. Not the source for “labelling text as AI-generated doesn’t reduce its persuasive effect” — that is Gallegos et al., PNAS Nexus (2026), which Hackenburg only cites; if the detection beat needs it, it is a shelf gap, not a line to borrow.

Reading List

• Polanyi, The Tacit Dimension (1966)
• Ginzburg, “Clues: Morelli, Freud, and Sherlock Holmes” (1979)
• Reber & Schwarz, “Effects of Perceptual Fluency on Judgments of Truth” (1999)
• Alter & Oppenheimer, “Uniting the Tribes of Fluency to Form a Metacognitive Nation” (2009)
• Dreyfus & Dreyfus, Mind Over Machine (1986)
• Klein, Sources of Power: How People Make Decisions (1998)
• Goldman, “Experts: Which Ones Should You Trust?” (2001)
• Sperber, Clément, Heintz, Mascaro, Mercier, Origgi & Wilson, “Epistemic Vigilance” (2010)
• Ricoeur, Freud and Philosophy: An Essay on Interpretation (1965)
• Mitchell, Lee, Khazatsky, Manning & Finn, “DetectGPT: Zero-Shot Machine-Generated Text Detection Using Probability Curvature” (2023)
• Kirchenbauer, Geiping, Wen, Katz, Miers & Goldstein, “A Watermark for Large Language Models” (2023)
• Sadasivan, Kumar, Balasubramanian, Wang & Feizi, “Can AI-Generated Text Be Reliably Detected?” (2023)
• Hackenburg, Wagner, Hewitt, Tappin, Saunders, Kirk, Margetts & Summerfield, “AI Systems Out-Persuade Expert Humans” (2026)
• Merleau-Ponty, Phenomenology of Perception, Part I (“The Body”) (1945)