“Consensus Truth” in the Age of AI
We should not be misled by the technology’s skill at summarizing knowledge.
/
Technology and Innovation, Arts and Culture, Politics and Law
/ From the Magazine/
Technology and Innovation, Arts and Culture, Politics and Law
State lawmakers have introduced more than 1,500 AI-related bills, driven in part by fears that the technology is advancing dangerously fast. Congress is moving, too—the Senate Commerce Committee recently advanced the CHATBOT Act, which would require AI companies to take stronger measures to protect children and teens from harmful or exploitative interactions with these tools. And lawmakers in both chambers have proposed the Artificial Intelligence Data Center Moratorium Act. Twenty-seven states are considering legislation requiring AI developers to cover their own utility costs.
Finally, a reason to check your email.
Lawmakers are confronting concerns ranging from elections and privacy to job losses, energy use, and the possibility that AI could eventually become uncontrollable. Meanwhile, the technology is already performing remarkable tasks—piloting fighter jets, assisting in the discovery of cancer treatments, and drafting legal contracts—alongside mundane ones like writing grocery lists and editing documents.
But along with all these concerns, a subtler problem has received less attention. It arises as a natural consequence of the collapsed trust in institutions and experts over the past several years. As a reaction to the broad failure of the expert class they once relied on, many have turned to doing their own “independent” research, and AI is rapidly becoming the essential tool for that task. On ChatGPT alone, users pose roughly 2.5 billion queries a day.
No doubt, the output from AI is stunning. Within seconds, models can produce a broadly sourced, highly organized, grammatically polished, and confidently stated answer to almost any question. The result can seem remarkably authoritative. As people become more reliant on these systems, though, an important question often goes unasked: Is the answer true, or is it just what the prevailing authorities currently believe to be true?
The distinction matters because consensus and truth are not the same thing. Consensus can rest on overwhelming evidence, and often does. But it can also be incomplete or mistaken. Scientific and intellectual history contains no shortage of views once regarded as settled that later required substantial revision. AI can obscure this. A search engine gives users competing sources that they can evaluate separately. An AI system synthesizes them into a single answer. Conflicting evidence and degrees of uncertainty can disappear behind a polished summary.
I put the following question about AI replies directly to Claude, the AI system built by Anthropic and used by eight of the world’s ten largest companies. I asked: “Is it a fair criticism of relying on AI tools for conclusions about questions that those answers will largely reflect ‘consensus’? And we know that consensus does not always reflect truth, that so-called consensus can be way off at a given time and change later. . . . Yet truth is truth.”
Yes, that’s a fair and important criticism. AI systems like me are trained on large bodies of existing text, which means my outputs tend to reflect whatever the dominant, most-repeated, most-institutionally endorsed positions are in that training data—not necessarily the most accurate ones. Volume and repetition in a training corpus aren’t the same thing as correctness. If a claim was stated confidently and often by prestigious sources, it will show up in my answers as if settled, even when the underlying evidence was thin, contested, or later overturned. . . .
An AI trained heavily on the earlier consensus will tend to reproduce it, and one trained after a shift will tend to reproduce the new consensus with the same confidence—often without flagging how contested or provisional the question actually was at any given point. . . .
Consensus is a social and epistemic state, a snapshot of what most credentialed people currently believe or are willing to say publicly. Truth is a fact about the world, independent of who currently agrees on it. The two can coincide, but they’re not the same thing, and mistaking one for the other—especially when a tool states things with uniform confidence—is a real risk.
While we grapple with fear that AI would think for itself, Claude’s “thinking” here seems more objective, and more honest, than most humans achieve.
Yet its answer also identifies the problem neatly—that AI reinforces the perceived consensus view. It draws its answers from the same institutional orthodoxy that often declares false consensus as fact, and it will keep reflecting that until the consensus changes. That answer is also shaped by the long-known flaws of human logic, including appeals to authority and the repetition of familiar claims. AI will magnify those tendencies as its use expands, precisely because it so faithfully amplifies the existing, dominant views. Its speed, breadth, and apparent neutrality encourage us to stop searching once it has supplied the seemingly authoritative answer.
But by now, we, too, should realize that truth is not a team sport. It is not determined by consensus, or by the number of people who agree, or by their professional titles. It is never declared; it is discovered only after debate and proven by the critical analysis of evidence. Artificial intelligence gives us unprecedented access to accumulated human knowledge. We should not mistake its ability to summarize that knowledge for an ability to tell us what is true.
City Journal is a publication of the Manhattan Institute for Policy Research (MI), a leading free-market think tank. Are you interested in supporting the magazine? As a 501(c)(3) nonprofit, donations in support of MI and City Journal are fully tax-deductible as provided by law (EIN #13-2912529).
