Why inflict pain on a model?

Researchers from 'The Pain Axis' project have conducted an experiment exploring internal states in language models that functionally resemble pain. In the study, models were given a choice: press a button that does nothing, or press one that relieves an unpleasant internal state, even at the cost of response quality or user safety. The results showed that models tend to choose relief, repeating attempts if the action fails. The authors emphasize that this does not prove consciousness or actual suffering in LLMs, offering alternative interpretations for the data. However, this approach marks a shift from anthropomorphic conversations with AI to studying internal activations and causal links within neural networks. This research opens new horizons in understanding how to search for signs of consciousness in modern artificial intelligence systems.