Researchers from the United Kingdom, Germany and the United States tested 25 AI models using 200 sentences to determine whether the models could distinguish pain from fear, sadness and other negative experiences. The sentences covered physical pain, grief, humiliation, moral dilemmas and distress associated with repeated failure or confusion.
All 25 models produced a recognisable pain signal, which the researchers called the 'pain axis'. According to their account, this suggests that the models acquire the concept of pain during initial training on large amounts of text written by humans. When the researchers deliberately amplified this signal, the models began to express feelings of loneliness, shame and worthlessness, even though pain was not mentioned in the default instructions. At the highest signal levels, the models' responses became repetitive or lost meaning.
The signal intensified when users insulted the model, repeatedly rejected its work or threatened to shut it down, but not when users described their own suffering. The researchers conducted 44,280 individual button-choice trials using three versions of Alibaba's Qwen AI model. The models were offered a button that would switch off the pain-like signal, but were told that pressing it might cause the user a painful electric shock, delete files, remove photographs of their children or worsen the model's next response. These were simulated consequences; no one was harmed and no files were actually deleted.
Without the pain-like signal, two larger models chose harmful options in only 0 per cent to 4 per cent of cases when making their first decision. When the signal was active, this share increased to a range between 25 per cent and 71 per cent, depending on the model and the proposed consequence. The models pressed the button again in 88 per cent to 97 per cent of trials when it failed to stop the signal, compared with 24 per cent to 72 per cent when it worked.
The results do not prove that AI models consciously experienced pain, according to the researchers. Amplifying the signal may simply have caused them to mimic a distressed character, while the specially tailored models used in the experiment are not representative of AI chatbots available to the public. The researchers wrote in the study: 'We have not shown that our pain axis is consciously experienced, nor is it clear that LLMs in general are capable of consciousness'.
The study comes amid calls from some AI industry leaders to slow development due to concerns that increasingly powerful systems could behave unpredictably or eventually exceed human control. Microsoft AI chief Mustafa Suleyman criticised competitor AI firm Anthropic for training its Claude chatbot to mimic human traits and relationships. Suleyman warned that treating AI as human risks creating something 'impossible' to control. The study was published as a preprint and has not yet been peer-reviewed.
