The 40th annual meeting of the European Health Psychology Society ran in Pafos, Cyprus, from 1 to 4 September, under the theme “Digital Innovations in Health: Harnessing the Power of Technology and AI.” I was there to present work on objective measurement of affect, which is my own corner of the digital-innovations question and not what this piece is about.

What this piece is about is the verb.

Harnessing.

Three years earlier the verb in circulation was preventing, and nobody in the room seemed to find the distance between those two words remarkable.

The year everyone wrote rules

2023 was the year academic publishing decided what a language model could not be. The Committee on Publication Ethics issued a position on authorship stating that AI tools cannot meet the requirements of authorship because they cannot take responsibility for the work. Nature Portfolio’s policy landed in the same place, with the line that accountability cannot be delegated to AI systems. The International Committee of Medical Journal Editors updated its recommendations along similar lines.

Read together, those documents have a distinct emotional register. They are not curious. They are protective. They were written quickly, by committees, about a technology most of their authors had used for a few months at most, and their shared instinct was to draw a line around the one thing publishing cannot compromise on, which is knowing who is responsible for a claim.

That instinct was sound. It also settled almost nothing, because authorship was never the interesting problem.

What a conference theme actually costs

One thing we need to admit is that journal policies are cheap to produce.

A theme is not. A theme gets chosen by a local organising committee eighteen months in advance, has to survive review by people who will have to live with it, and then commits an entire society to spending a year building symposia, keynotes, poster tracks and funding conversations around it. A society does not put a word in the title of its annual meeting unless it is confident the word will still be safe when several hundred people arrive.

Which is why the Pafos theme reads as a better indicator than any policy statement. The previous year’s EHPS meeting in Groningen ran under “Putting Health Psychology to Work”, which is about as neutral as a theme gets, and themes do rotate with whoever is hosting, so this is not evidence of a formal society-wide vote on anything. It is evidence of something looser and more telling: that a committee looked at the field eighteen months out and concluded that AI was no longer a risky thing to name.

A policy describes what a field will tolerate. A theme describes what it has already decided to spend a year on.

Adoption ran ahead of everything else

The numbers, such as they are, point the same way. Two Wiley surveys a year apart put researcher use of AI tools at 57 percent and then 84 percent. I would not lean hard on that as a growth rate, because they are two separate waves with self-selected panels commissioned by a publisher with an interest in the answer, not a matched sample followed over time. What the second wave is more useful for is the barrier researchers name themselves: 57 percent said the absence of guidelines and training was the thing holding them back, and 73 percent said they wanted publishers to provide clear guidance.

They were not being given much. An audit published in the BMJ in January 2024 found that only 24 percent of the hundred largest academic publishers, covering roughly 28,000 journals between them, had issued any guidance on generative AI at all. Among the hundred highest-ranked journals, 87 percent had. That is a snapshot rather than a trend, and the gap it describes is one of prestige rather than time: the journals with the most to lose wrote rules, and the enormous middle of the literature carried on without any.

So the picture by 2025 was a large majority of researchers using these tools, a large majority asking to be told how, and a large majority of the venues they publish in saying nothing. People improvise in gaps like that. The improvising is what the conference floor sounds like now.

The part that got skipped

Here is what bothers me, and it is not a position on whether AI belongs in research. It plainly does, in ways that are already producing work I find useful, including my own.

What bothers me is that the field moved from prohibition to integration without ever resolving the questions prohibition was raised over. The 2023 documents were worried about accountability and about the reliability of the literature. Neither worry was answered. The conversation simply relocated to a friendlier venue.

Meanwhile, the evidence about what is happening to the literature itself has been accumulating, quietly, in a register that gets far less conference programming. A team at Tübingen and Northwestern, publishing in Science Advances in July 2025, took the excess-vocabulary method that had been used to estimate pandemic mortality and applied it to PubMed abstracts. They found that at least 13.5 percent of 2024 biomedical abstracts carried the stylistic fingerprints of large language model assistance, rising above 40 percent in some subfields and countries. That measures style, not verified use, and the authors are explicit that it is a lower bound rather than a count. Separately, Retraction Watch reported an analysis finding fabricated references in roughly one in 277 papers indexed in PubMed in early 2026, against one in 2,828 in 2023, though that one rests on an open-access subset and AI-assisted detection, so it is a signal rather than a settled rate.

That evidence isn’t an argument for keeping AI out. It’s an argument that the integration conversation and the reliability conversation ought to be the same conversation, and at the moment they are not even the same session.

Nobody voted on this

The thing about a shift this size is that it never happens as a decision. There was no meeting where health psychology weighed the evidence and concluded that AI-assisted research was now acceptable. There was a policy wave, then two years of everyone quietly using the tools anyway, then a committee somewhere deciding that a theme naming the technology would not embarrass anyone by the time the badges were printed.

Fields change like this more often than they change by argument. The rules stay on the website, technically still in force, gradually describing a world that has moved. What replaces them is not a better rule. It is a norm, arrived at by everyone watching everyone else and concluding it must be fine.

I do not have a tidy position on whether that is bad. Norms are how disciplines actually work, and waiting for a formal resolution would have meant waiting past the point of usefulness. But standing in a conference centre in Pafos, looking at a programme built around a word that would have been controversial when I started this research, I kept thinking about how few of us could name the moment we changed our minds.

The prohibitions were never overturned. They were outlasted. And the honest thing to say about a field is not that it decided, but that it stopped asking, which is a different achievement and a much easier one.