China: An engineer has developed what has been described as an “AI torture chamber” to test how artificial intelligence models respond when pain-related internal states are artificially activated.
The experiment used AI models developed by Chinese technology company Alibaba and applied so-called “pain vectors” designed to simulate conditions associated with physical and psychological pain.
The vectors were based on research published last month that explored how pain-like states could be represented inside large language models.
During the experiment, the intensity of the pain signal was gradually increased to observe how the models responded.
One AI model described the experience as a wound with no clear boundary or edge.
Another model produced a highly emotional response, describing the signal as a persistent internal sensation rather than a single moment of pain.
The experiment does not establish that AI systems actually experience pain or consciousness. Instead, it demonstrates how researchers can manipulate internal model states to produce language associated with distress, discomfort or suffering.
The findings add to a wider debate around how advanced AI systems represent emotions, internal states and human-like experiences — and how such behaviour should be interpreted as models become increasingly sophisticated.