Tell HN: Gemini trains on chat history, even for subscribers (personal accounts)
Google’s Gemini AI is reportedly training on users’ chat histories from personal accounts, including paying subscribers, unless they disable history entirely—an option that also removes key features. Commenters highlight concerns about human review of transcripts, vague data-use language in Google’s terms, and an “all or nothing” pattern that effectively nudges users into data sharing. Many compare Gemini’s approach to other AI and cloud providers that offer clearer training opt-outs, and some advocate self-hosted or enterprise tools for genuinely private use.
Overall Reaction
- Many commenters are unsurprised that Gemini trains on chat history, including paid personal accounts; several say they assumed “everyone does this.”
- Others think the specific combination of defaults, human review, and opt-out model is still worth highlighting as user-hostile.
- Some users say this was a reason to unsubscribe from Gemini Pro; others say they never used Gemini seriously anyway because of quality or trust issues.
Data Use, Training, and Human Review
- Gemini uses chat logs for:
- Model training and service improvement.
- Human review to rate responses, suggest better ones, and enforce safety/policy.
- Reviewed chats can be retained for years and are kept even if users delete their activity.
- “Subset of chats” is undefined, and “disconnection” from the account is seen as weaker than true anonymization.
- Multiple comments stress: do not enter confidential or sensitive data; assume anything you type may be stored and reviewed.
Opt-out, History, and Usability Tradeoffs
- You cannot both keep chat history and fully opt out of training/human review; effective opt-out requires disabling history or using temporary chats.
- On Pixel/phone, key assistant features (reminders, device actions) reportedly break without chat history enabled, which users describe as coercive or a bait-and-switch.
- Some find that “all or nothing” pattern consistent with other Google products (Maps, YouTube watch history).
Comparisons with Other AI Providers
- Claims that “every AI trains on input” are disputed:
- Gemini Enterprise and some cloud offerings (e.g., Azure OpenAI) are cited as having explicit no-training guarantees for business data, with limited safety retention windows.
- Other consumer tools (ChatGPT, Claude) are said to offer history + explicit training opt-out, which Google’s personal Gemini lacks.
- Some note that workspace/business Gemini has better privacy but lags in model access.
Trust, Privacy, and Corporate Behavior
- Broad distrust of big AI labs: common themes are ToS changes, data being “too juicy a target,” and inevitability of future repurposing.
- Several view this as a continuation of Google’s ad-driven, data-maximizing strategy and “user-hostile” product design.
- Others argue that some retention and review is driven by liability and safety, not just exploitation, but this is contested.
Workarounds and User Strategies
- Suggested mitigations:
- Use local/self-hosted models or privacy-focused APIs.
- Avoid entering secrets; design workflows so agents never see raw credentials.
- Use Gemini only for low-sensitivity tasks (e.g., summarizing public content).
- Rotate accounts, use proxies, or even “poison” training data with noise or advocacy content.