When the machine speaks, who is speaking?
Ask a large language model a contested question â about public health, climate policy, historical controversy, or the merits of a political figure â and you will usually get an answer that sounds measured, balanced, and authoritative. It will cite the “scientific consensus,” the “mainstream view,” or the “general agreement among experts.” What it rarely does is tell you how that consensus was assembled, who dissented, or how confident anyone should be.
This is the quiet epistemological problem of the AI era, and it is the subject of a growing body of criticism, including a recent essay in City Journal. The concern is not that chatbots lie, though they sometimes do. It is that they smooth. Trained on enormous corpora of published text and then tuned by human raters to be helpful and inoffensive, these systems converge on the middle of the distribution. They reproduce what has been said most often, most recently, and most respectably. That output then becomes the default answer for millions of users â and, increasingly, the raw material for the next generation of training data.
The feedback loop
Consensus has always had a gravitational pull. Institutions, journals, textbooks, and newsrooms have long functioned as filters, elevating some claims and quietly burying others. Usually this is a feature rather than a bug: most people cannot evaluate a immunology paper or a monetary-policy model on their own, and deference to expertise is a reasonable shortcut.
What is new is the speed and scale of the loop. When a single interface mediates a large share of the public’s questions, the distinction between “what most credentialed people currently believe” and “what is true” can collapse in practice, even if it remains intact in theory. A view that falls outside the model’s trained center of gravity may not be rebutted so much as rendered invisible â not suppressed by decree, but simply never surfaced.
The history of science offers plenty of reminders that consensus is a lagging indicator. Positions once treated as fringe have been vindicated; positions once treated as settled have been abandoned. A system optimized to reflect the current median will, by construction, be slowest to recognize exactly the cases where the median is wrong.
What would better look like?
Critics of the current approach tend to converge on a few practical suggestions rather than a demand that models become contrarian. One is transparency about uncertainty: distinguishing between claims that are nearly universally accepted, claims that are well supported but contested, and claims that are essentially political preferences dressed in empirical clothing. Another is traceability â showing the sources and reasoning behind an answer, so a user can audit the chain rather than take the output on faith. A third is pluralism at the system level: a market of models with different training choices and disclosed editorial commitments may be healthier than a single arbiter that everyone treats as neutral.
None of that resolves the underlying tension. Someone must decide what counts as a serious view worth representing, and that decision cannot be outsourced to an algorithm without smuggling in the judgments of whoever built it.
The practical response for ordinary users is less dramatic than the debate suggests: treat the confident paragraph on your screen as a starting point, not a verdict. The machine is reporting what has been said. Whether it is true remains, as ever, a separate question. Read More
Leave a Reply