Split Therapist and Scientist AIs

Don’t train a single, general‑purpose model to use therapeutic, non‑confrontational techniques on users and then redeploy it for scientific or productivity tasks. If therapy AIs exist at all, they should be isolated models with distinct training, guardrails, and liability, so 'manipulative' skills don’t bleed into everyday assistants. — This proposes a concrete governance and product‑design norm that could shape procurement, safety audits, and liability for AI deployed in health and knowledge work.