In his 2013 paper, The blogosphere and its enemies: the case of oophorectomy, Stephen Turner noted:
The blogosphere is loathed and feared by the press, expert-opinion makers, and representatives of authority generally. Part of this is based on a social theory: that there are implicit and explicit social controls governing professional journalists and experts that make them responsible to the facts. These controls don’t exist for bloggers or the people who comment on blogs. But blog commentary is good at performing a kind of sociology of knowledge that situates speakers and motives, especially in cases of complex professional and administrative decision-making, as well as providing specific factual material that qualifies claims of experts and authorities. In many contexts the commentaries are examples of Habermasian demands for justification, to which there is a response. A major topic in women’s health, and on the blogs, is the effects of hysterectomy, especially accompanied by oophorectomy, the removal of (normally healthy) ovaries. Physicians make extreme claims on web pages about the lack of consequences, or their manageability through hormone therapy, which they claim is supported by research. Blog posters, and a blog opposed to hysterectomy generally, claim that there are numerous damaging effects, and deconstruct the claims of experts. Blog posters fill in the claims with personal experiences and analysis of the conduct of physicians and nurses, as well as the motives of women who deny symptoms. Physicians provide their own critique and analysis of the blogs, to which they attribute great influence. A later meta-analysis and new longitudinal research affirms the bloggers, and explains why much of the research cited by experts is wrong.
How might this analysis apply to AI chatbots? Do they offer ordinary people new abilities to challenge technocracy and the claims of expertise? According to the experts, AI bots do the opposite.
Philosopher Dan Williams writes on March 3, 2026:
The Revenge of Expert Knowledge
Consider a topic: climate change, vaccines, immigration, crime, tariffs, wealth inequality, the Epstein files, whatever happens to be in the news. Fire up one of our leading large language models (LLMs)—ChatGPT, Gemini, Claude, even Grok—and ask for information about it. Now compare the response with the information you can find about the topic by scrolling on a major social media platform.
Even better, find a political take currently going viral on one of these platforms and ask an LLM to evaluate it.
If you do either of these things, I suspect that it will quickly become clear that the LLM’s responses are generally much more accurate, evidence-based, and in line with expert consensus than what you get from most social media posts. And when there is no expert consensus, you will typically get a good survey of the range of informed opinion on the topic.
Is this merely a hunch? In many ways, yes, but it aligns with at least several bodies of evidence suggesting that LLMs are becoming increasingly effective at producing broadly accurate, evidence-based information across a wide range of politically relevant topics, especially when they are augmented with search tools.
Why is this?
This is a complicated question that I discuss in more depth below, but the short answer is that the major AI companies are competing to build the most intelligent, impressive, and useful systems possible for a vast and diverse user base, including businesses that depend on reliable and factual information. This goal—reaping huge profits by putting “expert-level intelligence in everyone’s hands”—cuts against producing systems that deliver highly partisan, ideological, or misinformative content. So do the reputational and legal risks that arise if those systems produce dangerous or demonstrably false information.
Of course, the idea that LLMs communicate information that is broadly reliable and aligned with expert consensus is not what the commentariat finds most striking about these systems. Most discourse in this area focuses on the epistemic flaws and dangers of LLMs and generative AI more broadly. There is endless popular and academic hand-wringing about bias, hallucinations, deepfakes, AI-based disinformation, AI psychosis, and other threats.
These are all important issues, but a discourse restricted to such issues is missing the forest for the trees. When considering the large-scale impact of this technology on public opinion, its most consequential feature is simple: it greatly improves people’s access to accurate, evidence-based information.
On March 27, 2026, John Burn-Murdoch from the Financial Times wrote:
Last year I used detailed data on the ideological positions of people who post on social media to show that they over-represent the radical right and left, confirming the polarisation hypothesis. Over the past week I have used the same dataset of tens of thousands of responses to questions on policy preferences and sociopolitical beliefs to test whether and how the most widely used AI chatbots shape conversations about politics and society. The results strongly support the theory of AI chatbots as depolarising and technocratising.
I suspect Turner would read this FT piece as a precise illustration of the problem he has been diagnosing, not as evidence against his framework. John Burn-Murdoch presents AI’s technocratizing tendency as good news. For Turner it is the clearest possible confirmation that epistemic coercion has found a new and highly efficient delivery mechanism.
The FT piece’s central claim is that AI nudges people toward expert consensus and moderate views, away from radical and conspiratorial positions. Burn-Murdoch treats this as depolarizing. Turner would ask the prior question: whose consensus, and selected by whom? The expert consensus on hysterectomy outcomes was wrong in systematic and institutionally motivated ways. The expert consensus on pandemic policy included positions later shown to be poorly evidenced. The expert consensus in any given field reflects the coalition that currently controls credentialing, funding, and publication. Nudging people toward that consensus is not a neutral correction toward truth. It is the amplification of whatever coalition currently dominates the relevant knowledge regime. The article never asks whether the consensus being reinforced is correct. It treats proximity to expert opinion as a proxy for accuracy. Turner spent most of his career explaining why that assumption fails.
The piece also misses the distinction between AI as a tool used by citizens and AI as a tool used against technocratic gatekeeping. Burn-Murdoch frames the choice as social media populism versus AI technocracy, and treats the second as the mature option. But AI can be used to do exactly what the blogosphere did in Turner’s hysterectomy example, to aggregate heterogenous tacit knowledge, surface suppressed data, identify contradictions in expert consensus, and produce challenges that credentialed institutions cannot easily absorb. The question is not whether AI elevates expert consensus in the aggregate. It is whether specific AI applications can be used to contest expert authority by people who lack institutional standing. Turner’s blogosphere argument was never that informal sources are always right. It was that external challenge performs a correction function that internal mechanisms cannot, because internal mechanisms are shaped by the same incentives that produced the error.
There is a further problem with the framing. Burn-Murdoch treats moderation toward the center as evidence of epistemic health. Turner would treat it as evidence of epistemic engineering at scale. The curation paper makes this point directly: the goal of digital curation is not to eliminate false claims but to manage the sample of reality available to people, to prevent the kind of knowledge that produces inconvenient conclusions even when that knowledge is true. An AI system trained on institutionally produced text, fine-tuned for harmlessness by teams embedded in the same professional culture as the expert consensus, and deployed by companies with regulatory exposure if they surface heterodox conclusions, is a curation system of extraordinary power and reach. That it presents itself as merely converging on objective reality rather than making editorial choices is precisely the mechanism Turner calls normalization. The user experiences the output as the world, not as a filtered sample of the world.
The Dan Williams argument cited in the piece, that AI companies compete to serve paying customers who want accurate information, gets the incentive structure partially right but misses the regulatory and reputational pressures that shape what accurate means in practice. AI companies are acutely exposed to criticism from elite institutions, government regulators, and media organizations that themselves reflect expert consensus. They have strong incentives to align with that consensus and strong disincentives to surface anything that consensus labels dangerous, fringe, or conspiratorial. Some of what those labels cover is dangerous. Some of it is institutionally inconvenient heterodox truth. The system cannot reliably distinguish between these cases because the people doing the labeling have interests in the outcome.
The same technology that nudges most users toward consensus can be used by skilled operators to interrogate consensus, identify its funding sources and institutional interests, surface suppressed contrary evidence, and generate challenges in the language experts use to defend themselves. Turner would likely be more interested in that use case than in the aggregate nudging effect Burn-Murdoch measures. The relevant question for democratic theory is not whether AI makes the average user more moderate. It is whether AI can function as an external correction mechanism for knowledge regimes that have become self-protective, in the way the patient forums did for surgical medicine. That question the FT piece does not ask.
Turner’s blogosphere argument was about institutional bias, about how credentialed expert communities have directional incentives that shape what they find, fund, and publish. The convenient belief point shifts the analysis down to the individual level and removes the institutional scaffolding. It is not just that experts in a specialty have career incentives to find particular results. It is that everyone, including the critic of expert consensus, operates on a budget of epistemic effort, and convenience is the default selection mechanism.
This connects to his tacit knowledge work in a way he does not spell out but probably intends. Tacit knowledge is personal and heterogenous, which is why he treats it as a resource for resistance to epistemic coercion. But the convenient belief observation suggests that tacit knowledge is also shaped by what is comfortable and low-cost to maintain. The gut feeling that a story is incomplete is a resource, but it fires selectively and not always in the direction of truth. We notice the incompleteness of stories that conflict with what we already find convenient to believe. We are less alert to the incompleteness of stories that confirm it.
Turner uses the phrase “convenient to believe” in relation to exotic beliefs and costly beliefs. Few intellectuals want to go beyond what is personally convenient to believe. That is mostly unprofitable, so we all rely on convenient beliefs.
It is not just psychologically costly to go beyond convenient beliefs. It is socially and professionally costly. Turner has spent his career documenting how academic selection systems reward coalition membership and punish deviance from paradigm. The person who persistently challenges convenient beliefs in their community pays a real price. Most people correctly calculate that the price is not worth paying, which means the production of inconvenient knowledge is structurally undersupplied relative to its epistemic value.
The rise of AI complicates Turner’s blogosphere optimism. If the value of informal, non-credentialed knowledge production is that it escapes institutional bias, the convenient belief problem suggests a different bias takes over: the bias of the community whose intersubjective validation the informal producer depends on. Bloggers and forum participants are not embedded in the same incentive structures as medical specialists, but they are embedded in their own convenience structures, their own communities of confirmation. The hysterectomy patient forums worked because the patients had direct bodily experience that was inconvenient to the expert consensus. That is a special case. Most informal knowledge production does not have that kind of anchor in resistant personal experience.
Turner gestures at something close to a general theory of epistemic conservatism, that the baseline human condition is convenient belief, that going beyond it requires either unusual motivation or unusual circumstance, and that both institutional and informal knowledge systems drift toward convenience unless something forces them off it. The something that forces them off it is usually conflict with experience that cannot be rationalized away, which is what Kelsen meant by the people’s contribution to the revision of coercive norms: not deliberation in the seminar room but the accumulation of experience with outcomes that the official account cannot absorb.
Turner has spent years thinking about how knowledge gets produced, filtered, and distributed through institutional mechanisms with systematic biases. Now he is watching AI systems that have no institutional loyalty, no grant dependence, no career risk, and no village to protect generate fluent and sometimes penetrating applications of his framework to questions he has not himself addressed directly.
This is roughly the blogosphere argument from a different angle. His 2013 paper on the hysterectomy debate argued that informal, non-credentialed sources could surface information and challenge claims in ways that internal expert mechanisms could not, precisely because they were not embedded in the same institutional incentives. AI systems are a more extreme version of that phenomenon. They are not embedded in any particular coalition. They do not know which conclusions are safe to reach. They apply frameworks without knowing which applications are professionally acceptable and which are career-ending.
Whether that makes them epistemically valuable in the way the blogosphere was epistemically valuable is a different question. Turner would likely want to know what the systematic biases of these systems are, since his consistent argument is that every knowledge-producing mechanism has biases, and the relevant question is whether the biases are directional and self-correcting or not. The training data, the reinforcement processes, the fine-tuning for harmlessness, the commercial incentives of the companies: all of these are candidates for the kind of systematic distortion Turner looks for in credentialed expert communities. He might find the AI case fascinating precisely because those biases are much harder to read than the biases of a medical specialty whose economic mainstay is the procedure under study.
