{"id":178705,"date":"2026-03-29T09:51:15","date_gmt":"2026-03-29T17:51:15","guid":{"rendered":"https:\/\/lukeford.net\/blog\/?p=178705"},"modified":"2026-03-30T06:12:31","modified_gmt":"2026-03-30T14:12:31","slug":"178705","status":"publish","type":"post","link":"https:\/\/lukeford.net\/blog\/?p=178705","title":{"rendered":"The Experts Are Back in Charge. Should We Trust Them?"},"content":{"rendered":"

Philosopher Dan Williams makes a strong case for AI as a technocratising force<\/a>, but his argument rests on an assumption that Stephen Turner’s<\/a> epistemic coercion<\/a> framework immediately destabilizes: that expert consensus is a reasonable proxy for truth, and that nudging people toward it is therefore a net epistemic good. Given that we did not evolve to be gullible with regard to our vital interests, does the whole Dan Williams framing misunderstand the purpose of human beliefs?
\nHugo Mercier’s<\/a> argument in The Enigma of Reason<\/a> and Not Born Yesterday<\/a> is that reason did not evolve primarily to help individuals find truth. It evolved to help people evaluate arguments in social contexts, to justify their own positions to others and to scrutinize the justifications others offer. The corollary is that humans are not generally gullible with regard to their vital interests. They are selective. They apply skepticism when claims touch their survival, when the source has skin in the game, when the stakes of being wrong are personally consequential. The domain where credulity flourishes is the domain of low-stakes, socially distant claims where error carries no cost. This is where institutional expert consensus tends to operate: policy recommendations, public health guidance, regulatory science, macroeconomic forecasting. These are areas where the expert bears no personal cost for being wrong and the citizen bears the cost but cannot verify the claim.
\nWilliams treats the move from social media to AI as a move from noisy populist chaos toward reliable expert knowledge<\/a>. But Turner’s analysis of epistemic coercion<\/a> suggests the expert consensus being fed through AI systems is not simply the distillation of truth. It is the output of knowledge-production systems shaped by funding incentives, coalition maintenance, career risk structures, and the same convenient-belief logic that governs all human institutions. The AI does not neutrally aggregate truth. It aggregates what got published, funded, and credentialed. Those filters have systematic biases that are directional rather than random. When an LLM tells you the expert consensus on a contested topic, it is often telling you what the dominant coalition within a relevant field found it profitable and safe to believe.
\nMercier’s framework predicts that people will resist this at exactly the points where it matters most to them. The person asking an LLM about vaccine safety for their child, or immigration policy in their town, or the economic effects of trade on their industry, is not operating in a low-stakes domain. They have direct experience and personal stakes. When the AI’s expert-aligned answer conflicts with what they observe in their lives, they will not simply defer. They will discount the AI in the same way they discount a government official who has never visited their town explaining why their town is doing fine. This is not irrationality. It is the operation of exactly the cognitive system Mercier describes: one that is well-calibrated to distrust sources that lack accountability and skin in the game.
\nThe deeper problem is that Williams treats technocratisation as a correction to democratisation, when Turner would say both are moves in the same underlying jurisdictional struggle. Social media gave distributed coalitions the tools to challenge expert authority. AI, as Williams describes it, gives expert coalitions a new and more powerful mechanism to reassert interpretive authority at scale. Neither is neutral. Both are coalition technologies. The question is not which produces more accurate beliefs in some abstract sense. It is which produces more accountable knowledge, knowledge that can be challenged, corrected, and revised when it fails the people it claims to serve.
\nTurner’s blogosphere paper<\/a> is relevant here. The patient forums that accumulated testimony about hysterectomy outcomes were not producing expert-validated knowledge. They were producing tacit<\/a>, experiential, heterogenous knowledge that happened to be right about outcomes the expert consensus had buried. An AI system trained on published literature would have reproduced the expert consensus. It would have been systematically wrong in exactly the way the expert community was systematically wrong, for the same institutional reasons. Williams’ technocratisation thesis offers no mechanism for catching this kind of error, because the error lives inside the expert consensus the AI is designed to amplify.
\nWhat Mercier adds is the evolutionary grounding<\/a> for why people sense this, even without being able to articulate it. They did not evolve to be epistemically passive recipients of authoritative information. They evolved to be skeptical of claims that serve the interests of the claimant, to weight testimony by the accountability of the source, and to trust embodied local experience over abstract institutional pronouncements when the two conflict. An AI system that is polite, comprehensive, and expert-aligned is not going to override these calibrations in the domains where they fire most reliably. It may reinforce convenient beliefs among people who are already aligned with the expert coalition<\/a>. It will generate resistance among people whose experience contradicts what the system tells them, and that resistance will look like irrationality to people inside the coalition<\/a> and like calibrated skepticism to people outside it.
\nThe piece Williams cites on Grok fact-checking is telling in this context. Republicans used Grok to fact-check claims, and Grok flagged Republican posts as misinformation more often than Democratic ones, roughly matching professional fact-checkers. Williams takes this as evidence that Grok is reliable and aligned with truth. A Turner-inflected reading notes that professional fact-checkers are themselves an expert coalition with documented political skews, and that building a system that aligns with their outputs and then validating it against their outputs is circular. The question of whether the fact-checkers are themselves systematically biased in ways that serve particular coalitions is precisely the question the methodology cannot answer.
\nNone of this means Williams is wrong that AI will push public opinion toward expert consensus. He may be right about that, though I doubt it<\/a>. The question is whether that movement is epistemically healthy or whether it is the successful reassertion of one coalition’s convenient beliefs over the tacit<\/a> knowledge of people whose lives the experts are describing from a distance. Mercier gives you reason to expect resistance, and Turner gives you reason to think that resistance might sometimes be epistemically justified even when it looks, from the inside of the expert coalition, like ignorance.<\/p>\n","protected":false},"excerpt":{"rendered":"

Philosopher Dan Williams makes a strong case for AI as a technocratising force, but his argument rests on an assumption that Stephen Turner’s epistemic coercion framework immediately destabilizes: that expert consensus is a reasonable proxy for truth, and that nudging … Continue reading →<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[42986,42951],"tags":[],"class_list":["post-178705","post","type-post","status-publish","format-standard","hentry","category-ai","category-epistemics"],"aioseo_notices":[],"aioseo_head":"\n\t\t\n\t\n\t\n\t\n\t\n\t\n\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t\n\t\t