anthropic-preserves-retired-model-weights

IN premise — entries/2026/06/21/wiki-Claude_language_model.md

Created 2026-06-21T09:50:09+00:00

Anthropic preserves weights of retired models as a precaution in case models have 'morally relevant preferences,' and conducts exit interviews with models before retirement.

Summary

Anthropic keeps the underlying parameters of retired AI models stored rather than deleting them, and conducts a kind of closing conversation with those models before they go offline, treating model retirement as a process with possible ethical stakes. This implies a working assumption that an AI's internal states might carry moral weight worth preserving, and it sets a precedent for how the system should think about lifecycle transitions and the obligations that come with shutting something down.

Dependents

These beliefs depend on this one:

IN anthropic-has-comprehensive-model-lifecycle-policy — Anthropic maintains a structured model lifecycle that distinguishes between deprecation (models still accessible with restrictions) and retirement (scheduled removal from service), with at least one model scheduled for retirement in August 2026. Deprecated models remain available to paid subscribers and via API by request. Anthropic also preserves weights of retired models as a precaution in case models have morally relevant preferences, conducting exit interviews before retirement.