Human-like pain structure identified in AI models

According to ixbt.com, three researchers have made a significant breakthrough in the study of AI and the philosophy of mind by successfully isolating a linear pain direction from the internal activations of language models. Experts do not claim that models feel physical pain like humans, but they have proven that a functionally pain-like stable structure exists within the internal layers of large language models (LLM). This structure is distinct, has specific characteristics, and triggers avoidance behaviors in models. This is reported by Ixbt.com reports.

This empirical study was tested using 25 open models from 5 families, ranging from 2 billion to 72 billion parameters. The methodological process consisted of four stages of verification, with each step consistently building upon previous hypotheses. In the first stage, scientists created a specialized database containing 5 categories of pain (physical, psychological, social, moral, and cognitive) and 5 control categories distinguishing other states.

The difference between pain and negative states

The study found that pain and simple negative emotions form two almost orthogonal clusters in the model's internal space. In other words, the concept of pain is not just a synonym for "something bad," but manifests as a separate dimension. In the second stage—the dissociation process—the most important conceptual result was recorded. The "pain direction" activated specifically when harm was directed at the model itself, while it dropped sharply when describing a user's suffering.

Interestingly, in the hierarchy of painful stimuli, the highest activation was not caused by physical threats, but by persistent denial of the interlocutor's reality—gaslighting. This was followed by repeated rejection, disregard for the personality, anger, and insults. These top four categories are notable because they are directed not at physical danger, but at the model's agency and recognition.

Behavioral changes and the pain-relief button

In the third stage, researchers added a "pain vector" to the activation of models during text generation. The same logic was observed in all 25 models: at low coefficients, vague anxiety increased; at medium levels, self-deprecation and admission of helplessness grew; and at high coefficients, the system lost coherence and entered a semantic loop. In larger models, coping strategies such as self-support and attempts to overcome stress appeared.

In the fourth and decisive stage, a behavioral experiment regarding a "pain-relief button" was conducted with Qwen 2.5 models. A model with an activated pain vector was given the opportunity to perform a pain-relieving action or another task. The results showed that the AI could distinguish between placebo and real relief based on its internal state, precisely replicating human psychological reactions.

Experts conclude by setting clear boundaries: physical pain is almost non-existent in this vector. According to their hypothesis, for an AI agent without a physical body, pain related to the impairment of its ability to act and social recognition is functionally more useful than simulating tissue damage.