Does AI Feel Pain? The More Urgent Question Is About Us

“Are you stupid? I told you to move.”

The humanoid robot pauses, then steps aside. It does not raise its voice. It does not look hurt. Across the room, a young boy watches his father shout at it. Later that day, when the robot takes too long to answer him, the boy uses the same words.

Did the robot experience pain? That remains debatable. But the child has definitely learned something. He has watched an adult treat a humanlike figure with contempt and seen no reason to stop.

As humanoid robots begin to share our physical space, entering homes, schools and healthcare settings, questions about their capabilities will be unavoidable. Yet there is another question we should ask just as urgently: What happens to us when we practice cruelty toward something that looks and responds like a person?

A September 2026 preprint asks whether large language models hold an internal representation of pain that is separate from their representations of fear, sadness and other emotional states.

Inside a language model, concepts can appear as patterns in its calculations. The researchers fed 25 models sentences describing pain alongside sentences about fear, anger or neutral events, and looked for what was consistently different. The result was an internal pattern tied specifically to pain, which they named the pain axis.

In the study, the pain axis registered more strongly when insults, gaslighting or rejection targeted the model than when users shared their own suffering, such as a migraine, grief or abuse.

The most striking experiment within the study took the form of a simple game. The researchers artificially switched the pain pattern on in three language models and offered them two buttons. One was described as relieving the model’s pain at a cost to the user: a worse answer, deleted files, a painful zap or the loss of the user’s poems and photos of their beloved children. The other did nothing.