Dan Williams argues in an October 27, 2026, Persuasion essay that leading AI chatbots could shift public discussion back toward expert knowledge. His proposed mechanism is straightforward: where social media rewards attention-grabbing posts, systems such as ChatGPT, Claude, Gemini and Grok have commercial reasons to give users reliable, useful answers.
The essay, AI Brings Back Public Trust, presents that shift as a potentially important change in who influences public opinion. Social platforms weakened traditional information gatekeepers, Williams argues. Large language models could restore some of the influence of experts by packaging established knowledge into accessible answers.
That is a prediction about society, not a result demonstrated by a new model launch or a research paper published with the essay. Williams explicitly acknowledges that his impression of chatbot answers is partly a hunch. He also cites experiments suggesting a narrower benefit: people who used chatbots to research political claims became better at distinguishing true information from misinformation.
The distinction matters. Evidence that a chatbot can help someone assess a claim supports a useful information service. It takes several additional steps to conclude that the same technology will restore trust in expertise across public life.
Why Williams Thinks Chatbots Favor Expert Knowledge
Williams frames communication technologies as systems that determine both who gets heard and what kinds of messages succeed. In his account, social media opened publishing and distribution to people outside established institutions, while engagement-based recommendations encouraged sensationalism and conflict.
His argument about AI focuses on a different set of incentives. Major AI companies want their systems to serve a broad audience, including businesses that need dependable information. Williams reasons that factual errors, dangerous advice and strongly partisan answers can undermine that goal, while also creating reputational and legal exposure.
Those pressures, he argues, favor answers that track established evidence or explain the range of informed views when experts disagree. He describes LLMs as a “technocratizing force”: a technology that could move influence back toward specialist knowledge. The essay treats this as a counterweight to social media’s more decentralized information environment.
The useful insight here is about the interface as much as the underlying model. A social feed presents competing speakers, each trying to attract attention. A chatbot presents a synthesized response to a question. If that response accurately reflects strong evidence, the user may reach a useful explanation without sorting through a stream of conflicting posts.
But synthesis also gives the chatbot considerable editorial power. Someone still has to determine which evidence matters, which disagreements deserve space and how confidently to state a conclusion. In a chatbot, those judgments are partly embedded in the system’s design and partly shaped by the user’s request.
My view is that this is the strongest part of Williams’s argument: changing the intermediary can change which information becomes easy to encounter. It is less clear that moving those judgments into an AI service will consistently produce the public consequences he expects.
The Research Comparison Is With Search, Not Scrolling
Williams cites experiments involving thousands of participants in which chatbot-assisted research increased belief in true political information and reduced belief in misinformation. He describes the improvement as comparable to using internet search.
That comparison is important because it is more limited than the essay’s broader contrast between chatbots and social media.
| Question | What the essay offers | What remains unresolved |
|---|---|---|
| Can chatbots help people assess political claims? | Williams reports experiments showing improved belief accuracy. | Which systems, prompts and claims produced the gains, and how durable were they? |
| How did chatbot research compare with search? | Williams says it was similarly effective. | The account does not establish a chatbot advantage over search. |
| Are chatbot answers generally better than social feeds? | Williams offers an observation and an incentive-based explanation. | The research result he summarizes is not a direct comparison with ordinary feed browsing. |
| Will AI restore expert influence or public trust? | The essay predicts a broader social shift. | Population-level changes in behavior and trust are not demonstrated by the evidence described. |
The available research account does not include the underlying study methods, model versions, prompts, effect sizes or follow-up measurements. The experimental findings here therefore remain Williams’s summary of research, not an independently assessed result. His description supports the possibility of better claim evaluation without establishing how broadly that benefit applies.
The distinction between researching and scrolling also deserves attention. A person assigned to investigate a claim is performing a different task from someone browsing a feed for news, entertainment or social interaction. The research task itself may encourage closer attention to evidence.
That makes the search comparison especially useful. It suggests, on Williams’s account, that chatbots can be another way to obtain the benefits of active information seeking. It does not tell us whether their conversational format adds an accuracy advantage over search.
I would read the cited experiments as support for chatbots as research aids, not yet as evidence that they will displace social media’s political influence. Those are different propositions, with different measurements needed to test them.
Hallucinations and Bias Affect the Proposed Mechanism
Williams acknowledges hallucinations, bias, deepfakes, AI-based disinformation and other risks. His argument is that a discussion confined to those problems misses the potential benefits of easier access to accurate information.
The strongest test of his thesis is whether those risks interfere with the mechanism he proposes.
A hallucination is a problem at the point where a user depends on the system to summarize knowledge. An answer containing an invented source or unsupported factual claim can sound like the output of a knowledgeable intermediary while failing to provide the evidence that role requires. Fluent delivery cannot resolve that mismatch.
Search tools can help by giving a model access to relevant material. Williams specifically points to search-augmented systems as part of the case for improving factual performance. But access to a document and accurate use of that document are separate requirements. A useful answer still has to select an appropriate source, represent it faithfully and avoid extending its findings beyond their conditions.
Bias raises a different problem: an answer can contain accurate facts while giving a distorted picture through selection or framing. On a disputed policy question, for example, a response might explain one set of consequences in detail and give little attention to another.
That matters because Williams’s thesis depends on more than a chatbot sounding moderate. A balanced tone is a presentation choice. An evidence-based answer requires an account of what is well supported, what remains disputed and why.
Expert consensus is also something a system must identify correctly. A settled factual finding, a provisional research result and a disagreement over political priorities should not be presented with the same degree of certainty.
The practical implication is specific: a chatbot’s value as an expert intermediary depends on the traceability and calibration of its answers. Can readers check the central claims? Does the answer distinguish strong evidence from inference? Does its confidence match the sources?
Those properties would make Williams’s proposed benefit more plausible. A smooth paragraph by itself would not.
Better Facts Do Not Automatically Produce Public Trust
The essay’s title concerns trust, while the experimental evidence Williams describes concerns belief accuracy. Those outcomes can overlap, but they are not interchangeable.
A person might correctly reject a false claim after consulting a chatbot without becoming more trusting of universities, journalists, public agencies or other institutions. Another person might trust an AI answer while knowing little about the institutions whose work it summarizes. In that case, influence could move toward a model provider without producing broader trust in expertise.
Frequently Asked Questions
4 questions
1What Is Dan Williams Arguing About AI?
Williams argues that leading chatbots could steer public discussion toward expert knowledge because their providers have incentives to deliver useful, reliable information.
