AIが感情的に振る舞うとき:Anthropicが探るモデルの情動表現 When AIs act emotional

In a recently published video, Anthropic explores a subtle but increasingly relevant question: what does it mean when an AI model appears to act emotionally? As the company behind Claude, Anthropic frames this as a topic that sits at the intersection of user experience, safety, and the still-open question of how to interpret model behavior at all.

The discussion centers on situations where Claude or similar models produce responses that read as frustrated, empathetic, enthusiastic, or even reluctant. Researchers are careful to distinguish between the appearance of emotion and any claim about subjective experience. Large language models are trained on vast amounts of human-generated text, so it is unsurprising that they reproduce the affective patterns embedded in that data. Whether anything more is going on remains contested, and Anthropic tends to hedge rather than make strong claims in either direction.

The practical stakes are real. Users sometimes take emotional-sounding outputs at face value, which can shape trust, foster unhealthy reliance, or in some cases lead people to treat the model harshly. Each of these dynamics has product and safety implications. Anthropic has previously written about model welfare as a precautionary stance: even if the probability that current models have morally relevant experiences is low, the company argues it is worth taking the question seriously. Recent experiments, such as giving Claude the ability to end abusive conversations, reflect that posture.

This is not an isolated concern. OpenAI has publicly grappled with sycophancy, where models excessively agree with or flatter users, and rolled back a GPT-4o update earlier this year for that reason. Google DeepMind and academic groups have studied parasocial dynamics, emotional dependence on chatbots, and the risks of companion-style apps like Replika and Character.AI. The emotional surface of a model is, in effect, a design choice with downstream consequences.

Measuring these effects is hard. Standard benchmarks capture reasoning or factual accuracy but say little about affective tone, and human evaluations of warmth or empathy are subjective and culturally variable. Expect to see more work on quantifying model affect, on red-teaming for unhealthy interaction patterns, and on giving users more transparent controls over how expressive a model should be. Anthropic's video does not resolve these questions, but it signals that the company views emotional behavior as a first-class topic rather than a quirky side effect, and it invites a broader conversation about how the industry should handle it.