主要推理API提供商正在采用越来越多的技术措施来限制用户数据的可移植性。[1]OpenAI、Google和Anthropic等供应商通过加密令牌、隐藏搜索内容和不透明压缩等手段，使得用户的AI会话记录不再完全独立，而是部分绑定到提供商的服务器上。[1]
具体而言，推理令牌以加密二进制形式返回给用户，虽然会被计费但用户无法访问其内容。[1]网络搜索功能由提供商私密执行，用户只能看到返回的URL，无法查看模型实际使用的完整内容。[1]OpenAI的压缩功能返回的加密压缩项被描述为"不可读且不供人类解释"，[1]而在多代理系统中，代理间通信使用加密有效负载，用户同样无法查看代理被分配的具体任务。[1]
数据保留政策也存在差异。OpenAI的Responses API默认存储响应至少30天。[1]Google Gemini Interactions API的默认设置为自动存储，付费层保留数据55天，免费层仅保留1天。[1]此外，Anthropic声称禁止在竞争AI模型训练中使用其输出，但自己仍进行蒸馏训练。[1]业界人士呼吁API提供商采用透明、用户可控的标准，以便用户能够在不同提供商之间迁移会话。[1]
Major artificial intelligence API providers are increasingly implementing vendor lock-in measures that undermine the portability of user sessions and data, according to recent analysis on Hacker News [1]. These practices include encrypting reasoning tokens, obscuring web search operations, applying opaque compression techniques, and encrypting inter-agent communications in multi-agent systems [1]. The combined effect renders AI session records only partially portable, with significant portions bound to each provider's infrastructure rather than remaining under user control [1].
The specific mechanisms of this lock-in vary across providers. OpenAI returns reasoning tokens as encrypted blobs that are billed to users despite remaining inaccessible to them [1]. Web searches are conducted privately by providers, with users seeing only URLs rather than the full content models actually consume [1]. OpenAI's compression feature produces encrypted output that is "unreadable and not intended for human interpretation," preventing users from understanding what data has been compressed [1]. In multi-agent architectures, inter-agent communications use encrypted payloads, leaving users unable to view the tasks assigned to individual agents [1]. OpenAI's Responses API stores responses for a minimum of 30 days by default, while out-of-band conversation items lack a 30-day time-to-live limit [1]. Google's Gemini Interactions API defaults to storing interactions, with retention periods of 55 days for paid tiers and one day for free tiers [1].
The analysis also highlights conflicting practices among providers regarding data usage. Anthropic restricts competitors from using its outputs to train rival AI models, yet simultaneously conducts its own distillation training [1]. These cumulative restrictions make it increasingly difficult for users to migrate their sessions between providers, contrary to the principle of transparent, user-controlled standards that the article advocates for [1].