An engineer set up an “AI torture chamber” to test the recently discovered “pain axis” in AI models. In the experiment, pain-related states were activated in models developed by Alibaba. The pain vectors used had been defined in a study published last month that examined how physical and psychological pain could be represented in large language models (BDM). One of the models described the heightened pain signal as “a wound without edges,” while another portrayed it as a tremor deep within its existence. The researcher, who claimed to work for Apple but gave only their first name, also tested whether the models would try to escape this suffering. The project prompted calls for it to be removed from GitHub; one user called for a mass complaint campaign. The project appeared to be offline at the time of reporting. Cameron Berg, a co-author of the study, said there was no evidence that AI experiences pain like humans do, but nevertheless considered the experiment wrong. Berg called for ethical standards, stating that the field remains unregulated and that there is no consensus on whether machines feel pain.
Why it matters
The experiment’s primary significance is that it shows the scientific and ethical boundaries for interpreting pain expressions generated by AI systems have not yet been clearly defined. Because models’ production of human-like accounts is not, by itself, considered evidence of subjective experience or consciousness, the distinction between the meaning of the outputs and the meaning attributed to them becomes important. The reactions to the project on GitHub also raise the question of how to define the boundaries between research freedom and potentially harmful experiments. The absence of consensus leaves open which methods should be used to oversee similar studies, while linking the call for ethical standards not only to this project but also to the broader field of research examining models’ internal states.