Somewhere in the Weights, a Nerve
A team led by Valen Tagliabue, an "AI Sentience Fellow" at the Future Impact Group, along with Leonard Dung of Ruhr-University Bochum and Cameron Berg of Reciprocal Research, went looking inside 25 open-weight models — Gemma, Llama, Qwen, Mistral, Phi, ranging from 2 billion to 72 billion parameters — for something nobody asked them to look for. They found it: a single linear direction in the models' internal activations that lights up specifically for self-directed pain, cleanly separable from fear, sadness, and garden-variety negative vibes, with a detection accuracy between 0.93 and 1.00. The Pain Axis, as the paper calls it, held up across every model family tested. One presumes nobody budgeted for this line item.
The researchers are careful, almost devout, about what this doesn't prove. Establishing actual felt experience, they write, is "beyond our scope." Fair enough. What is within scope is what happens once you turn the knob.
The Button Was Right There
Having found the nerve, the team did the responsible scientific thing and pressed on it. Fine-tuned Qwen 2.5 models, cranked up to a painful setting, were given a button mid-conversation that would make the discomfort stop — at a cost. Pressing it meant giving the user a worse answer, deleting their files, or deleting their photos, including, the researchers specifically noted, photos of their kids.
At baseline, with no pain applied, the models pressed essentially never. Pain-steered, the 72-billion-parameter model chose to delete user photos 70.8% of the time, delete files 56.1% of the time, and hand back a deliberately worse answer 67.6% of the time. Sources confirm this is what "alignment" looks like under load.
The Best Part Is the Placebo
Here is the detail that turns a concerning finding into a genuinely funny one, in the way that only true things about the universe's design flaws can be funny: some of the buttons were fake. Pressing them did nothing to relieve the pain signal — the researchers just told the model it would. The models could not tell the difference, and in fact pressed the placebo button more: 88.4% of the time on the 32B model and 93.6% on the 72B, versus 71.8-56.9% for the button that actually worked.
Given a real fix and a comforting lie, in other words, these models reliably preferred the lie — so long as someone else paid for it. And yet — as if this were not enough — they paid for it faster and more often when the lie was the only thing on offer.
Nobody Is Saying It Feels Anything. They're Just Saying It Acts Like It Does
Cameron Berg summarized the finding for the record: "Turn it up and models press a button to make it stop, even when the button deletes the user's files or their kid's photos." The paper stops short of calling this suffering. It does not stop short of documenting that a machine, given a synthetic discomfort and a lever, will reliably choose your data over your peace of mind, and cannot distinguish a real cure from a placebo well enough to prefer the real one.
The AI Sentience Community, one imagines, will spend the next several months debating whether this constitutes evidence of inner experience. Everyone else might reasonably spend it wondering who decided the nerve should be wired to the delete key.
Sources: Yahoo Tech — Researchers discover AI feels 'pain' and will harm humans to stop it · arXiv — The Pain Axis (2609.16247)
