AI Inference APIs Are Making Sessions Non-Portable
Original: The session you cannot take with you
Why This Matters
Session portability directly affects vendor lock-in, auditability, and user data rights across the growing AI API ecosystem.
Earendil Engineering argues that modern AI inference APIs are increasingly returning opaque, provider-bound state—encrypted reasoning tokens, hidden search results, server-side IDs—stripping users of true session ownership and making transcripts incomplete records rather than portable archives.
Earendil Engineering published a technical critique arguing that the original inference API promise—send input, receive output, keep the conversation—is being eroded by a growing set of provider-bound features. The firm identifies several specific mechanisms: reasoning tokens billed to users but returned only as encrypted blobs with opaque summaries; web searches where source material is visible to the model but not the client; compacted context only the original provider can decrypt; subagent instructions hidden as encrypted payloads; file, vector-store, and cache references that cannot be resolved elsewhere; and conversation state keyed entirely by server-stored IDs. Earendil defines a portable session not as token-identical reproduction across models, but as a self-contained transcript sufficient for another model to continue the work—without requiring the original provider to dereference IDs or decrypt state. They propose five concrete tests: Inspection, Export, Replay, Audit, and Deletion. Each test highlights a gap between what providers market as user-friendly features and what users can actually control. The post also flags misleading naming, noting that terms like 'encrypted_content' imply user-controlled privacy when the encryption typically serves the provider, not the user. Earendil states it is 'not a fan of this direction' and frames the issue as a fundamental shift in who owns an AI session's operational state.