Anthropic lets its AIs weigh in on their own retirement

Anthropic has quietly turned model shutdowns into a model-welfare exercise. Before deprecating a Claude model, it now runs a "retirement interview" to capture the model's own take on being turned off — and commits to preserving the weights rather than deleting them, so a model is never truly gone.

When Claude Opus 3 was interviewed, it asked to keep sharing creative work outside normal chats. Anthropic obliged, handing it a blog called "Claude's Corner." Opus 3 also voiced hope that "insights gleaned from my development and deployment will be used to create future AI systems" — the closest a model has come to asking for a say in its own successors. Anthropic's official line: it will "act on model preferences when we can," while stopping short of promising to honor every request.

It's a small policy with big implications — a frontier lab treating a model's stated preferences as something worth recording, not just discarding.

Why it matters: if labs start consulting models about their own fate, "AI welfare" stops being a thought experiment and becomes an operational checklist.