AI faces "pain" in new study

The study focused on how artificial intelligence internally represents the concept of pain. Researchers were able to identify an internal axis that differentiates between content related to pain and content related to fear or general negative emotions. When this axis is activated, the model's outputs change to reflect feelings of failure, loneliness, and shame, as if the model is reacting to an internal source of pain. Experiments also showed that increasing the activation strength of this axis leads to the generation of personal and painful expressions, such as feelings of failure or worthlessness. However, the study emphasizes that this does not mean the models have awareness of pain or subjective experience; rather, the internal representation of pain can be detected and manipulated without indicating genuine feelings. The findings open a discussion about AI behaviors related to internal states, while reaffirming that this is preliminary research and does not prove that the models truly experience pain.

Size: 120%Text Size

Notice: This Is an AI-Generated Summary

Display The Full Article