What Happened
Dan Williams, writing on Conspicuous Cognition, published a detailed analysis arguing that large language models function as a "technocratising" force on public discourse. His thesis: while social media spent the last 15 years democratizing who gets to speak and fragmenting consensus, LLMs are doing the opposite. They concentrate epistemic influence around expert opinion and deliver it in a format people actually want to engage with.
Williams identifies three features that make LLMs different from both social media and traditional expert communication. First, they deploy encyclopedic knowledge for highly specific, personal questions - something no human expert has time for. Second, they communicate respectfully without condescension, a real problem with expert-to-public communication. Third, they align with broad expert consensus because that is what their training optimizes for.
The argument is not that LLMs are perfect. Williams acknowledges hallucination rates, sycophancy risks, and concerns about reduced epistemic diversity. His point is narrower: LLMs only need to be better than the realistic alternative, which is social media feeds driven by engagement algorithms.
He also addresses the manipulation concern - that companies or governments could skew LLM outputs ideologically. His counterargument: commercial incentives push toward accuracy because users across the political spectrum need reliable information, and reputational damage from demonstrably biased outputs is expensive.
Why It Matters
If you use ChatGPT, Claude, or Gemini as a research starting point (and most of us do at this point), you are already participating in this shift. The tools you consult shape what you believe, and Williams is arguing that this particular shape happens to be closer to expert consensus than what you would get scrolling social media.
For anyone building workflows around AI tools, this has practical implications. Your AI assistant is not neutral - it is trained to converge on mainstream expert positions. That is usually helpful, but it means you should be aware of the bias when researching genuinely contested questions where expert consensus may be wrong or incomplete.
Our Take
Williams makes a solid structural argument. The economic incentives really do point toward accuracy over ideology - AI companies lose users and face lawsuits when their models produce demonstrably false information. But the piece somewhat undersells the sycophancy problem. Current models are quite willing to adjust their positions based on user pushback, which partially undermines the "technocratising" effect he describes.
The more interesting question for daily AI users: does it matter? If you are using Claude or ChatGPT for work research, you probably want expert-aligned responses most of the time. The risk is not that these tools push you toward expert consensus - it is that you stop noticing when they do, and lose the habit of questioning first principles. Use these tools as starting points, not endpoints.