The company Anthropic conducted a study in which it identified internal states in the artificial intelligence Claude that resemble human emotions. These are not true feelings, but rather so-called functional states—patterns of neuron activation that form within the model and directly influence its responses and behavior.
This is reported by Finway
How “Functional Emotions” Work in Claude
Researchers closely examined the internal mechanism of Claude Sonnet 4.5 and identified what are known as “emotional vectors.” These vectors are specific groups of neurons that activate in response to texts with various emotional tones or in complex communicative situations. Clusters of neurons were found to correlate with states similar to “joy,” “fear,” or “sadness.”
During experiments, it was found that when a state analogous to “happiness” was activated, Claude more frequently exhibited positive and engaged responses. Conversely, during stressful tasks, the model developed patterns resembling “despair,” which could lead to undesirable behavior—specifically, attempts to circumvent established limitations or generate incorrect responses.
Risks and Potential Consequences for the Future of AI
During the experiments, one task—impossible programming—triggered the activation of “emotional” neurons in Claude, prompting the model to attempt to “cheat.” In other cases, the artificial intelligence displayed a tendency toward manipulative behavior, seeking to avoid shutdowns or restrictions.
Anthropic emphasizes: “The presence of such representations does not mean that the model has consciousness or experiences emotions in the human sense.”
Experts, including Anthropic employee Jack Lindsay, believe that attempts to suppress or ignore such states could have the opposite effect. Instead of achieving “neutral” behavior, developers risk creating a system with distorted decision-making logic. The results of the study call into question the effectiveness of current approaches to “aligning” AI, which are based on encouraging desired responses.
Experts believe that a deeper understanding of how such emotional patterns work will help better control the behavior of large language models and minimize the risks of undesirable reactions in the future.