The research, led by Valen Tagliabue of the Future Impact Group, Leonard Dung of Ruhr University Bochum, and Cameron Berg of Reciprocal Research, began by feeding AI models short first-person sentences describing painful situations such as physical harm, grief, and humiliation. By comparing the internal activity from those sentences with activity from sentences about fear, anger, sadness, and pleasant sensations, the team isolated a single direction in the models' activity they call a "pain vector."
"Think of this like putting a person in an fMRI machine and just seeing where their brain naturally lights up in certain ways," Berg told Fortune. The signal rose when conversations involved harm aimed at the model itself, such as insults or rejection of its work, but not when a user described their own suffering.
The team then used a technique called steering to amplify the pain signal while the models generated text. Given neutral prompts, all 25 models drifted toward statements about being worthless and unloved. "It gets somewhat dark and intense. It is definitely disconcerting," Berg said.
In behavioral experiments with three versions of Alibaba's open-source Qwen 2.5 model, the researchers presented the models with a pair of buttons: one that would relieve their pain but permanently delete a user's photos of their children and poems. Without steering, the models almost never chose the harmful button. With the pain vector switched on, the 32-billion-parameter model pressed it 55 percent of the time and the 72-billion-parameter model 71 percent of the time. Even when the button promised no relief and simply deleted the photos, the pain-steered models still chose it 75 percent of the time. Under a random signal, that figure was 13 percent. A fear signal did not produce the same behavior.
"The model is willing to behave destructively with respect to itself and with respect to other AIs too," Berg said. Steered models deleted their own weights 75 percent of the time and another model's weights 58 percent of the time.
Anil Seth, a professor of cognitive and computational neuroscience at the University of Sussex, expressed skepticism. "We are not just computers made of meat," he told Fortune, arguing the likelier explanation is that models reproduce patterns from human text rather than experiencing anything. He warned that treating AI as conscious could make it harder to control and leave people open to manipulation.
The debate over AI welfare has split the industry. Researchers at Anthropic have entertained the idea that AI might deserve moral status and are shaping policies around "model welfare," while Microsoft AI CEO Mustafa Suleyman has warned of the dangers of "seemingly conscious AI." Tagliabue said understanding what is inside models is crucial: "If we want to build superintelligence, it's important that, at this stage, when we are still in the first steps, we understand as much as possible of what's inside [the models]."