The Session You Cannot take with you

The article by Earendil.com critically examines a concerning trend in AI inference APIs: the move away from portable, user-owned session data towards proprietary, provider-bound states. It argues that the initial promise of simple input-output APIs, where conversation transcripts could be easily archived and transferred, is being undermined by new features that prioritize provider convenience over user control.

• Eroding Portability: Modern inference APIs are increasingly returning provider-bound state, such as encrypted reasoning tokens and hidden search results, making it impossible for users to fully own or port their AI sessions between different providers.
• Practical Test for Session Ownership: The author proposes a simple test: can a session's transcript be exported and then used by a different provider to continue the conversation, without relying on the original provider's specific infrastructure or decryption keys?
• Opaque Encryption vs. Privacy: Features marketed as "encrypted content" are often "provider-sealed state," unreadable by the user and primarily serving to hide data from the user rather than ensuring privacy from the provider.
• Default Server-Side Storage: APIs from major players like OpenAI and Gemini often default to storing user responses, transforming local transcripts into mere pointers to data controlled by the provider.
• Hidden Reasoning Chains: Most closed-weight models do not expose raw chain-of-thought, returning opaque blobs or model-generated summaries, which prevents true auditability and portability of AI reasoning.
• Undisclosed Web Searches: Hosted web search features typically provide only citations, not the full context (e.g., result rankings, passages, filtered content) that the model used, creating a dependency on the original provider for historical context.
• Non-Inspectable Compaction: Server-side context compaction, particularly OpenAI's, produces unreadable, encrypted states, forcing continued reliance on the original provider, unlike more transparent approaches that offer readable summaries.
• Opaque Multi-Agent Communication: Multi-agent systems introduce further non-transferable state by encrypting inter-agent messages and instructions, making it impossible for users to audit or understand agent-to-agent interactions.
• Importance of "Freedom to Leave": While not all users switch providers frequently, the ability to do so fosters competition, ensures user control, and provides flexibility in cases of service changes, outages, or specific compliance needs.
• Principles for Portable APIs: The article outlines seven key rules for providers, advocating for canonical local event logs, explicit storage defaults, readable handoff representations, full-fidelity tool logs, auditable subagent communication, inspectable compaction, and exportable artifacts.
• Distillation Hostility: The author criticizes AI labs for discouraging external model distillation while practicing it internally, an "asymmetrical moral" stance that limits competition and the availability of efficient, smaller models.

In conclusion, the article argues that the industry's current trajectory limits user sovereignty over their AI interactions and data. It calls for a fundamental shift towards greater transparency, interoperability, and user control, emphasizing that a truly open and competitive AI ecosystem depends on the ability of users to own, inspect, and transfer their digital conversations and the underlying reasoning.