A Hacker News post documenting how Google's Gemini uses subscriber chat history for model training drew more than 650 upvotes this week — not because the policy was secret, but because so many paying users had never noticed it. The thread, titled "Tell HN: Gemini trains on chat history, even for subscribers (personal accounts)," lays out Google's own support documentation showing that a paid subscription does not exempt anyone from training use.
The relevant behavior is not buried in an end-user license agreement. It is described in Google's Gemini help pages, which AI Buzz Wire reviewed directly. For more context on this story, see our ongoing AI industry coverage.
What Google's Documentation Actually Says
Google's Gemini Apps Privacy Help page states that Gemini Apps Activity — saved chat history — "is used to improve Google services for everyone, including by training generative AI models, if your Keep Activity setting is on." That single sentence carries the core of the issue: training on user conversations is the default outcome of having chat history enabled, regardless of whether the account pays for Google AI Pro or Ultra.
The same documentation is explicit about human involvement. Google states that "a subset" of chats are reviewed by human reviewers, including trained reviewers from service providers, to debug services, keep them safe, and improve them. And the guidance includes a warning many users never see: "Please don't enter confidential information that you wouldn't want a reviewer to see or Google to use to improve our services, including machine-learning technologies."
Reviewed Chats Are Kept for Up to Three Years
The retention rules go further than most users expect. According to Google's help pages, chats reviewed by human reviewers — along with related data such as language, device type, location information, and feedback — "are not deleted when you delete your activity. Instead, they are retained for up to three years." Google says reviewed data is disconnected from the user's account before being saved, but the conversations themselves persist on Google's systems long after the user has pressed delete.
Google says it takes steps to protect user privacy in this process, including associating reviewed chats with random identifiers rather than accounts. Still, the documentation makes the boundary clear: deletion controls what appears in your activity feed, not what Google keeps from chats a human reviewer has already pulled for evaluation.
How to Actually Opt Out
The opt-out exists, but it is a trade. Turning off the "Keep Activity" setting removes your chats from training use — Google's pages confirm that training applies only when that setting is on. The cost is functionality: with Keep Activity off, chats are retained with your account for only 72 hours, and Google uses just the last 24 hours of a chat as context for follow-up questions. Temporary chats offer a similar arrangement.
Users can also shorten the default retention window from 18 months to 3 months or 36 months, or set it to indefinite, and can delete individual chats at any time. None of those choices, however, changes the training policy while history is on — and the 72-hour clock starts on every conversation the moment history is disabled.
Paid Doesn't Mean Private
The reaction on Hacker News centered on a broken assumption: that paying for an AI subscription changes the data arrangement. In consumer software, that assumption has long been shaky — but the Gemini case is notable because the training pipeline applies to personal accounts across paid tiers, and because the reviewed-chats retention is measured in years rather than months.
It is worth noting what Google is not doing. The company states that Gemini does not train generative AI models directly on your Gmail inbox, Drive, Contacts, Calendar, or other Workspace content — that data requires explicit connection through Connected Apps, which are then subject to their own training and human-review policies. And Workspace and enterprise accounts operate under different terms than the consumer Gemini app, a distinction Google has drawn in prior privacy controversies.
The Broader Trust Problem
The timing is awkward for Google. The company is competing hard on consumer trust — pushing scam detection into Android Messages and Calls, and marketing privacy protections on its hardware. Meanwhile its most-used AI product trains on the casual, unguarded things people type into a chat box, with human reviewers able to read a subset of them and the reviewed subset kept for up to three years.
Privacy researchers have long advised that people treat chatbot conversations as semi-public by default. This week's thread made that advice concrete for Gemini's paying users: the setting that protects your conversations is off by default, the one that trains on them is on, and the only fully effective escape hatches — temporary chats or disabled history — come at the price of the product's memory.
Google's documentation is not hidden, and nothing in it is false. The friction is that reading it is the user's job — and, as hundreds of upvoted comments suggest, most users had simply assumed their subscription fee was doing that work for them.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
