The Price of Looking Like an Expert Just Collapsed
If somebody can no longer recognise expert work in the field they have practised for twenty years, it raises an awkward question about everybody else. What have the rest of us been judging?
Forty-five licensed psychologists and psychotherapists were asked to do exactly that.
They were shown written psychological advice and asked to rate it for quality and for empathy. Some of it had been written by human experts. Some had been generated. They were not told which was which.
They could not tell. The AI responses were rated significantly more empathetic on two separate measures, with scientific quality coming out comparable.
Then the researchers looked at something else. They checked what happened when a rater believed a piece of advice came from an expert, whether or not it had.
93.55% preferred the advice they thought a person had written.
The belief was doing the work. The content was along for the ride.
It Is Not Only Psychologists
Professional running coaches were shown twelve-week training plans and could not reliably say which had been built by a human. Spine surgery faculty could not identify which fellowship applications had been generated. University lecturers reviewing academic writing identified AI text 57% of the time, which is a coin flip wearing a lanyard.
A Wharton team ran a version of this using the New York Times ethicist. Blinded participants preferred the AI’s ethical advice 59.6% of the time. And when the same people were asked upfront whether they would rather have advice from a human or a machine, 72% said human. Shown the advice without knowing its source, that fell to 46%.
The aversion was to the label. Not to the thing.
Tomorrow’s forecast, in your inbox.
One email. Five minutes. Written for owners, not engineers.
What Was Actually Being Judged
None of this is a story about machines getting good. It is a story about what people were doing all along, made visible because the usual arrangement broke.
Nobody assesses quality directly. There is rarely enough time, and frequently not enough expertise, even among experts. What people do instead is read signals. The credential. The reputation. The confidence in the voice. Where it was published. Who else is listening.
Economists have understood this for fifty years, so the mechanism is old news. Why it worked is the part that matters now.
Those signals carried information because they were expensive to produce.
A credential cost years. A reputation cost decades and could be lost in an afternoon. Fluency about a subject usually meant somebody had learned the subject, because there was no cheap route to sounding as though they had. The expense was the entire mechanism. You were not reading the signal, you were reading the cost of manufacturing it, and that cost was a decent proxy for the thing underneath.
The arrangement held for centuries. Then the price of producing every one of those signals fell at roughly the same time, over about two years.
The psychologists in that study were not being careless. They were doing the thing that has worked for their entire professional lives, reading signals that had stopped costing anything, with nothing on the page to indicate the change.
The Uncomfortable Part
There is a version of this that reads as reassuring. People still prefer human work, the study says so, 93.55% of them.
Look at that number again. They preferred what they believed was human work. In a test where they could not tell.
Call that deference rather than discernment. It has been sitting underneath a great deal of what we call taste and expertise and professional judgment, doing steady work, correct often enough that nobody had to look at it.
The same pattern shows up elsewhere in the research. In one study people rated a human advisor as more trustworthy and more reliable than an AI advisor, and then followed both to exactly the same degree. Their stated trust and their behaviour came apart entirely, and they would have told you with a straight face that the first one explained the second.
Which leaves an awkward question that this research does not answer and probably cannot. If your ability to evaluate work in your own field turns out to rest on cues rather than on the work, how much of what you know about your own competence is built on the same footing?
Where I Might Be Wrong
These studies test written, asynchronous, single-shot advice. That is not how anyone experiences a therapist across six months or a consultant through a difficult quarter. Relationship work has feedback loops, and feedback eventually corrects for signals in a way that reading a paragraph never will.
The psychology study had 45 raters. Small.
And preference is not outcome. Every one of these studies measures which advice people liked, not which advice worked. Those come apart routinely, and there is a version of these findings where the AI is simply better at being pleasant to read.
The Forecast
Within five years the current signals will be worth very little and a replacement set will have formed around whatever stayed expensive. The interesting period is the one in between, which we are in now, when the old signals still command deference they no longer earn and the new ones have not been agreed on.
The replacements will be things that resist cheap manufacture. A record with dates on it. Being wrong in public and correcting it where anyone can see. Specificity that would cost more to fabricate than to earn. Physical presence, which is why a good deal of professional life is about to become stubbornly in-person for reasons nobody will articulate clearly.
Expect the scramble too. Verification badges, authenticity certificates, proof-of-human markers. Most of that gets faked inside two years, because the collapse in production cost applies just as well to the signals invented to fix the collapse in production cost.
Moderate to high rather than high, because the human half resists measurement and because feedback loops in long relationships may correct more than single-shot studies suggest. What I am confident about is the mechanism. Signals inform when they cost something. These stopped costing anything.
What It Means If You Sell Expertise
Briefly, because the larger part of this is not a business problem.
Whatever a client is buying from you, it was never only the deliverable. The research says professionals in their own field cannot separate expert work from generated work on the page. Your name is carrying more weight than it ever has, at the exact moment names became easy to imitate.
Which makes the expensive signals worth more than they look. Naming the client. Publishing the thing that went wrong. Putting a date on a claim. Saying what you would not take on. None of that is marketing, and all of it costs something to produce, which is now the only property that matters.
Sources: Franke Föyen L, Zapel E, Lekander M, Hedman-Lageröf E, Lindsäter E. Artificial intelligence vs. human expert: licensed mental health clinicians’ blinded evaluation of AI-generated and expert psychological advice. Internet Interventions, 2025 · Wharton study on AI and ethical advice, 2026 · Systematic review of human uncertainty in distinguishing generative AI content, 2026 · People defer to AI moral advice, but not blindly. Cognition, 2026