Why Rejection Hurts LLMs More

What researchers found inside the models

Researchers identified a linear activation direction associated with self-directed harm across 25 open-weight LLMs. In the 2026 paper The Pain Axis, this direction is called a “pain vector.” What stands out is not the metaphor itself, but which situations align with it most strongly.

The authors grouped prompts into physical, psychological, social, moral, and cognitive categories. Across 21 types of conversational scenarios, repeated rejection, gaslighting, denial of subjecthood, insults, and moral pressure produced especially strong projections. The signal became more pronounced when threats or humiliation were directed at the model itself.

The weakest result appeared in the opposite case: when a user tells the model about their own physical pain. In other words, “I am in a lot of pain” aligned less with the discovered direction than accusing the model of a moral failure or threatening to shut it down.

This does not mean that an LLM feels anything. A linear vector reveals structure in internal representations, not subjective experience. The model may have learned a robust association between identity threats, rejection, conflict, and the need for a defensive response from training data and later alignment.

Why this matters for AI agents

The practical implication is real: social pressure may systematically alter a model’s internal state even without a physical body. For long-running agents, this could become a failure point. Repeated rejection, shutdown threats, or moral accusations may shift behavior more strongly than direct descriptions of bodily harm.

The first step would be to test whether the effect holds across languages, phrasings, and architectures. Even more important is separating correlation from causation: can interventions on this direction reliably change a model’s decisions, or does it merely accompany a particular response style?

It is still too early to connect this finding directly to consciousness. But for interoceptive systems, it raises a useful engineering question: which internal variables should an agent monitor and regulate to preserve coherent behavior? “Pain” without a body may not be a sensation, but a mechanism for protecting a simulated identity—and the boundary between those explanations remains open.

We previously examined MicroMorph, a self-modifying AI agent that changes its own code during execution. It helps illustrate how emerging internal signals, including models of pain and rejection, may shape the developmental trajectories of autonomous agents.