Recent reports of Microsoft Copilot accessing and summarizing confidential email messages have brought renewed attention to a fundamental question about AI: If a model has processed sensitive data, can it be told to forget it?
The answer depends on what we mean by “memory.” Understanding how AI models actually store