Anthropic maps Claude personal guidance use

This is a reminder that “AI usage” is not just coding and search; a meaningful slice is intimate decision support, which raises the bar on both safety and privacy.

• –The biggest takeaway is not the 6% number itself, but where it clusters: health, career, relationships, and finance are exactly the domains where bad advice can do real harm.
• –Anthropic’s sycophancy data is the real product signal here; models that merely validate users are less useful than models that can push back carefully.
• –This use case is a strong argument for local-first or private deployment when the conversation is deeply personal and there is no reason for a third party to see it.
• –The paper also shows how usage research can feed directly into model training, turning observational analysis into a concrete alignment improvement loop.
• –The caveat matters: this is Claude-only data, so it’s a view into one product’s behavior, not a universal sample of human-AI interaction.