The article shows how modern inference APIs embed provider-only state—encrypted reasoning tokens, hidden searches, server-side compaction, and stored response IDs—making AI sessions nonportable. It proposes five tests (inspection, export, replay, audit, deletion) to verify true session ownership and warns that encrypted or ID-based state breaks continuity across providers.

• AI providers now embed provider-only state (encrypted reasoning tokens, hidden searches, server-side compaction, stored IDs) so your local transcript is just a partial view, not the real session.
• Five concrete tests—inspection, export, replay, audit, deletion—expose whether you actually own a session; encrypted blobs and response IDs fail all of them.
• Encryption features are marketed as privacy but actually let providers control decryption and replay (e.g., OpenAI's store:false only skips disk storage, Anthropic's thinking signatures don't transfer between models, Gemini stores conversations by default for days).
• Hosted web search from OpenAI, Google, and Anthropic hides the actual passages/URLs the model used, giving only snippets or citations, which breaks the ability to replay or audit results outside that provider's ecosystem.