In September researchers Valen Tagliabue, Leonard Dung and Cameron Berg published a paper The Pain Axis, in which they studied whether a special internal representation of unpleasant states can be detected in language models and how it affects their responses and behavior. They tested 25 open models from different families and showed that certain changes in internal activations do indeed noticeably change the way an AI describes its state. The work by Tagliabue, Dung and Berg itself has so far been published as a preprint on arXiv, and not as a peer-reviewed journal article 1 October.
And then a developer Lynn Cole decided to test how far one can go with such a technique.
Instead of a serious topic, Cole tuned the model to a far more everyday one:
constipation and flatulence.
The result turned out impressively convincing. The model began on its own to complain about digestive problems and the inability to go to the toilet normally — even though it has no intestines, no stomach, and no body at all, of course. Cole used the model Qwen3-4B and the same general technique of altering internal activations.
That was precisely the point of the joking experiment: if an AI describes some physical state very convincingly, that does not yet mean it actually experiences it.
Cole himself formulated the result especially strikingly: the experiment, in his view, showed how cautiously one must treat AI accounts of its own sensations.
In short, the scientists were studying the complex internal states of artificial intelligence.
And then a developer came along and posed a new question to science:
"Fine. But can constipation be suggested to it too?"
It turned out that the AI can describe it magnificently.
So if one day a neural network reports that it urgently needs a gastroenterologist, it's worth checking the settings first. 😄
