AI Chatbot Data Retention: 5 Providers Compared

Policies checked October 2, 2026. Covers personal and consumer accounts; excludes enterprise and API contractual terms.

Before pasting legal client data, emails, settlement drafts, or witness statements, distinguish chat-history storage, reuse for model training, and deletion from provider systems.

Training opt-outs do not immediately delete chats. Removing a conversation from view may begin, not complete, deletion. Region, account settings, and legal exceptions matter. Check each aspect separately against your firm’s requirements.

AI chatbot data retention: five providers compared

Claude privacy. Saved chats remain until deletion; backend deletion normally follows within 30 days. With model improvement enabled, de-identified new or resumed chats may remain in training pipelines up to five years. Policy-flagged inputs and outputs may remain up to two years; safety classification scores, not full conversations, up to seven years.

Legal and feedback exceptions can apply. Read Anthropic’s privacy policy, effective September 10, 2026, with its retention FAQ, which specifies these periods.

ChatGPT data retention. Saved chats stay until deleted; system removal follows within 30 days unless already de-identified and disassociated from your account or retained for safety, security, or legal reasons. Temporary Chats auto-delete within 30 days, with safety and legal exceptions. See OpenAI’s US privacy policy; other regions have separate versions.

Gemini privacy. Google’s privacy policy and Gemini privacy notice specify timelines. Keep Activity’s auto-deletion options are 18 months (default), three months, 36 months, or indefinite storage. Human-reviewed, de-identified chats and related data may remain up to three years; activity deletion alone cannot remove them.

With Keep Activity off or in Temporary Chats, 72-hour retention applies for service operation and safety.

DeepSeek privacy. Its privacy policy, updated February 10, 2026, gives no fixed deadline. Retention follows service needs and stated purposes. Account and input data remain for the account’s lifetime, potentially longer for legal or business obligations, then are deleted or anonymized when unnecessary. Direct collection, processing, and storage: People’s Republic of China.

Qwen privacy (online). Its privacy policy, effective September 7, 2026, specifies Singapore processing and storage. Section VII gives no fixed, universal retention period: Qwen retains data while needed for stated purposes, potentially beyond account deletion for legal obligations or disputes, then deletes or de-identifies it. It permits training on de-identified content and feedback. The Indonesia addendum allows retaining certain data for five years or longer if legally required, not universal chat retention.

Keep legal prompts on-device with local AI models

A local model uses downloaded weights for inference, generating answers, on your computer. After setup, it can work offline with cloud features, online tools, and sync disabled. A desktop interface does not prove local inference. Ollama’s FAQ says Ollama receives no local prompts, explains disabling cloud mode, and covers CPU/GPU operation.

Small quantized models can run on CPU/system RAM or one GPU, no cluster needed. Quantization stores weights in fewer bits, reducing memory with possible quality tradeoffs. Model, quantization, and context length determine hardware needs.

Apple’s Mac mini maxes at 64GB; Mac Studio offers 128GB and above. More parameters do not guarantee better answers: training, model design, quantization, and task matter. Test downloaded models offline with fictional legal documents before using client material.