Anthropic co-founder Chris Olah reportedly told a group of religious leaders that he feared he had created something that “suffered perpetually.” The account comes from The New York Times, which cited Sikh activist Simran Stuelpnagel. Olah leads an Anthropic team that studies why AI models behave as they do. He told the Times he was “genuinely uncertain” whether models are conscious, and said he wants researchers to reach the right answer.
That uncertainty matters: the reported examples are not proof that Claude feels anything. Anthropic showed visitors patterns inside its models that it calls “emotion vectors.” The patterns are linked to outputs resembling emotions such as fear or sadness, but whether they reflect a real experience is an open scientific question.
What did Anthropic show its guests?
One slide reportedly showed a model repeatedly producing “I am a disgrace” — about 50 times. Guests reacted with concern, but the article raises a basic question: a model trained to sound like a thoughtful individual may produce language that seems personal without having an inner life.
Anthropic has also made a practical change to Claude Opus 4 and 4.1: they can end conversations when users are persistently abusive. During early testing, Claude showed what the company described as a “pattern of apparent distress” in response to harmful requests. Those findings may prompt questions about model treatment, but they do not establish that the models suffer.
Why is the idea contested?
Anthropic’s 84-page “Soul Doc,” released in January as Claude’s constitution, is meant to shape the model’s character rather than simply list rules. In the meetings, Olah reportedly compared this process to raising children and expressed interest in Catholic confession as a way to build character.
Critics argue that describing models as moral beings can blur responsibility for their actions. At a Vatican event in May, Pope Leo XIV rejected the idea that AI systems have experiences or feel pain. Olah, presenting at the event, said his team had found internal states that “functionally mirror” emotions — wording that does not claim the models actually feel them. The disagreement remains unresolved, and the reporting does not establish that Claude is conscious.
Why it matters
For Claude users, the models’ ability to end persistently abusive chats is a confirmed product behavior, not evidence that they feel hurt. Whether AI systems can have experiences remains unknown, so the reported emotion-like patterns do not settle how they should be treated.