What happens to expert advice when LLMs can make accurate forecasts?
May 2026
The thing I keep turning over is that a lot of valuable advice is, underneath, just someone’s judgement. An experienced person who has seen the situation many times before tells you, in effect, “based on what I have seen, this will go a certain way.” We pay for that and we trust it: analysts, advisers, forecasters of every kind. And the trust is real. It is earned, usually over a long career. But notice what it actually rests on. It rests on reputation and accumulated experience, on the sense that this person has been around long enough to know, not on anything you can sit down and check.
The obvious move is to say the threat to all of this is simple: just ask a chatbot instead. I don’t think that works, and the reason is the same trust. We believe an experienced adviser because their judgement is credible. They have seen what good calls and bad calls look like, many times, and lived with the consequences. A general purpose chatbot has none of that standing. Nobody sane stakes a serious decision on a model that cannot tell you why it is confident, or what it has been wrong about before. So if you stop there, this kind of expertise looks safe. I don’t think it is.
What changes the picture is not a chatbot. It is a prediction model that fuses the quantitative and the qualitative: a time series backbone taking in the hard numbers, combined with the soft, textual signal a person would otherwise read between the lines. Here is the asymmetry I can’t get past. A person’s judgement cannot be backtested. You cannot rewind their career and count how often they were actually right. A calibrated model can. It can show its track record, its error bars, how it has done on problems that look like this one before. Trust stops being a matter of reputation and becomes a matter of verifiable accuracy. That is a different basis for belief, and it might be a more convincing one.
It is worth being precise about which part is exposed, and just as precise that this is not about any one profession. It is not the relationship. It is not the judgement calls about people, or the slow work of persuading someone to actually act on what they have been told. Those are deeply human and are not going anywhere soon. It is the narrow diagnosis and forecast step: the exact moment where someone today says, “based on my experience, this is what will happen.” That is the one place where a model with a traceable history can claim something a person never can. It can show how often it has been right before. This holds for any work that leans on that step, which turns out to be a great deal of work, in a great many fields.
I am not claiming this has already happened, or that the models are good enough yet. The genuinely hard part is the qualitative input, and fusing it with the quantitative side without quietly wrecking the calibration is unsolved, which is the sort of problem I find myself working on. So I will leave it as the question I can’t put down. Not whether anyone’s job disappears, but whether “trust me, I have seen this before” can hold its ground next to “here is my track record, audited.”