On 16 November 2026 Google changed the settings of its generative AI app Gemini so that chats with history turned off are kept for about two months. Before the change, those chats were kept for 72 hours. So if you turn off history and stop Gemini from training on your chats, do the chats you typed at work leave Google's hands? They do not. Training consent and retention are separate controls, and any chat you rate is read by people and kept for up to three years.
01Gemini keeps chats for about two months even with history off
The short answer is no. Google's own help page says that chats held while the Gemini Apps history setting is off stay with your account for around two months. The Register, a British technology news site, reported on 17 November that the figure had been 72 hours until this update. Choosing not to let Google train on your chats does not stop that retention.
Some chats are kept even longer. If you send a thumbs-up or thumbs-down on a Gemini answer, that conversation can go to human review. A reviewed chat is kept for up to three years, even if you delete it from your account.
For anyone who uses Gemini at work, the training switch is no longer enough to decide what to type. Three other facts decide it: how many days a chat stays on Google's servers, when a person may read it, and which kind of contract you are using. The people most exposed are those who paste company material into a personal Google account.
02Retention with history off grew from 72 hours to about two months
The two-month figure comes from Google itself. The Gemini Apps Privacy Hub states that temporary chats, and chats held while history is off, are retained with the account for around two months. The same page states that reviewed chats are retained for up to three years.
The story is easier to follow if three words are kept apart. "History" is the record that lets you scroll back through your own conversations. "Training consent" is your permission for Google to use a conversation to improve its models. "Retention" is the period during which a conversation still sits on Google's servers, whether or not you can still see it on your screen.
With this update, training consent became a setting of its own, separate from history. You can keep your history and switch off training only. You can also switch off both. Even with both off, retention continues for about two months.
The Register put "20x longer" in its headline. The underlying numbers are the old 72 hours and the roughly two months now on Google's help page. The paper also noted that sending feedback on a chat stretches its retention to three years, which is the box at the right end of the chart.
03Personal accounts and Workspace accounts differ on human reading
The two-month and three-year periods do not apply to every Gemini user in the same way. Google runs personal Gemini and the Gemini that comes with Google Workspace under different rules. Workspace is Google's business bundle of Gmail, documents and other tools that an organisation buys for its staff.
Under a personal account, a chat you rate can be sent to human reviewers. Once reviewed, it can stay for up to three years, even after you remove it from your account. Turning history off does not change that; sending the rating does.
Workspace works differently. Google's Workspace privacy page says customer data is not used to train models without the customer's prior permission or instruction. Chats are not sent to human review without permission either. How long conversation history is kept is set by the organisation's administrator.
The gap therefore falls on people who handle work material through a personal Google account. The Workspace commitments attach to accounts that an organisation has contracted for. If a company pays for Workspace but an employee opens Gemini in a private account, the personal rules govern that conversation.
04Turning off training leaves a separate safety purpose in place
In personal Gemini, a chat stays for about two months even after training is switched off. Google explains this with purposes other than training. According to The Register, Google said it keeps these chats to provide responses, to process feedback, and to protect users and the public.
The chart sets two of those purposes, responding and protection, beside model training, so a single conversation carries three purposes. The training switch stops only the branch on the right, model training. The other two purposes keep the conversation on the servers.
Tying storage time to purpose is not Google's invention. Article 5 of the EU General Data Protection Regulation says personal data must be kept in an identifiable form for no longer than the purposes require. I cannot judge whether Google's two months meets that test. What I can say is that the length of storage is the kind of thing regulators measure against a stated purpose.
In Japan, the Personal Information Protection Commission issued an alert on generative AI services in June 2023. It asked businesses that enter personal information into such services to confirm two things.
Within the stated purpose
If you enter personal information, confirm that doing so is necessary for the purpose you specified.
No use beyond the output
Confirm that the provider will not use the data for machine learning or similar purposes.
The second card says machine learning "or similar purposes." In my view, retention that continues after training is switched off belongs in that "similar" category and should be checked with the same care. Google's own explanation shows that when training stops, the chat is still kept for a purpose other than producing output: protecting users and the public.
05Retention runs from 30 days to five years depending on service and consent
Google is not alone in keeping chats for a fixed period. In August 2025 Anthropic revised the consumer terms for Claude. If a user allows their chats to be used for model training, those chats are kept for five years. If the user does not allow it, they are kept for 30 days.
| What to check | Gemini (personal) | Gemini (Workspace) | Claude (personal) |
|---|---|---|---|
| Use for training | Can be switched off; rated chats are used | Not without permission or instruction | Only chats from users who allow it |
| Retention period | About two months even with history off | Set by the administrator | 30 days without consent, five years with it |
| When people read chats | Rated chats; kept up to three years | Not without permission | Not confirmed in these sources |
Read across a row, and the same promise of "not used for training" sits beside very different retention periods. Decline training in Claude and a chat stays for 30 days. Turn off history in Gemini and it stays for about two months. In Workspace it stays for however long the administrator chooses.
Read down a column, and a second point appears. In these three services, what sets the number of days is less the product name than the pairing of contract type and the consent the user has given. The same Gemini, used through a personal account and through Workspace, differs on all three rows: training, retention and human reading.
Some numbers are missing from the table on purpose. I could not open OpenAI's official help pages to confirm how long ChatGPT keeps chats, so ChatGPT is not listed. I also did not confirm, from these sources, when people at Anthropic read Claude conversations. The empty cell marks a place where users have to read the terms for themselves.
06A training switch alone cannot settle the fate of your input
In the three services compared, retention, then, is decided by contract and consent together. That leads to three changes in how a person should judge what to type into a generative AI tool at work.
Stopping training does not delete
Chats are kept about two months even with history off, so the training setting alone does not make it safe to enter confidential material.
A rating changes the handling
A chat you rate is read by people and kept up to three years, even after deletion.
The contract sets the terms
Workspace sends nothing to training or human review without permission, and administrators set retention.
The first card removes a common line of reasoning: "it will not be used for training, so I can paste in the confidential file." Google's help page says chats with history off are still kept for about two months. Whether training is off and whether the chat remains on a server are two separate questions that need two separate checks.
The second card is a reason to reread a conversation before rating it. A click meant only to say an answer was good or bad sends the chat to handling under which a person may read it and it can be kept for up to three years. If the conversation contains a client's name or internal figures, it is better not to rate it at all.
The third card tells you where company information belongs. Gemini in Workspace does not send data to training or to human reviewers without permission. My view is that company information should be handled inside the business contract the company has signed, not in an employee's personal account.
The training switch does not appear in any of the four boxes. You check the contract first, then the retention period, then the conditions under which people read chats. Looking at the training setting can wait until after those three.
07Users have no outside means to confirm chats are gone after two months
All three points assume that Google does what its help pages say. Does a chat actually disappear once the two months have passed?
None of the material I checked described a way for a user to confirm deletion from the outside. Google's help page and The Register's article both state retention periods. Neither mentions any mechanism that shows users a record of deletion. For now, whether chats are deleted on schedule rests on Google's own account.
Of the things that could happen next, two follow from the facts above. The first is that the retention period changes again. It went from 72 hours to about two months by Google's decision, and it can be rewritten the same way. The second is that the change is tested against the EU storage-limitation principle, since Article 5 of the GDPR does not allow storage beyond what the purpose needs.
Neither has happened yet. The material I checked says nothing about how EU regulators view this change. What remains in the user's hands is simple: read the stated retention period, and type only what you would accept being stored for that long.
- On 16 November 2026 Google extended retention of Gemini chats with history off from 72 hours to about two months. Stopping training does not stop that retention.
- A chat you rate is read by people and kept up to three years, even after deletion. One thumbs-up or thumbs-down changes how long it is kept.
- Gemini in Workspace is not used for training or human review without permission, and admins set retention. Whether work data belongs there depends on the contract more than on the product name.
Turning off history and training does not make a Gemini chat leave Google's hands. Retention rules set how long it stays, and the training switch governs only one of the purposes attached to it. A rated chat is read by people and kept up to three years. Choose what work data to type after checking retention and the contract.
- Google (Gemini Apps Help). Gemini Apps Privacy Hub. Accessed 2026-11-17. https://support.google.com/gemini/answer/13594961?hl=en (about two months' retention with history off; up to three years for reviewed chats; training setting separate from history)
- The Register (Brandon Vigliarolo). Opting out of Gemini tracking? Google will still hold your data 20x longer. 2026-11-17. https://www.theregister.com/ai-and-ml/2026/11/17/opting-out-of-gemini-tracking-google-will-still-hold-your-data-20x-longer/5310517 (72 hours before the update; three years after feedback; purposes Google gave for retention)
- Google Workspace. Generative AI in Google Workspace Privacy Hub. Accessed 2026-11-18. https://knowledge.workspace.google.com/admin/gemini/generative-ai-in-google-workspace-privacy-hub?hl=en (no training or human review without permission; history managed by administrators)
- Anthropic. Updates to Consumer Terms and Privacy Policy. 2025-08-28. https://www.anthropic.com/news/updates-to-our-consumer-terms (five years with training consent, 30 days without)
- EU General Data Protection Regulation (GDPR). Art. 5 GDPR – Principles relating to processing of personal data. 2016-04-27. https://gdpr-info.eu/art-5-gdpr/ (storage limited to what the purpose requires)
- Personal Information Protection Commission, Japan. 生成AIサービスの利用に関する注意喚起等 [Alert on the use of generative AI services]. 2023-06-02. https://www.ppc.go.jp/files/pdf/230602_alert_generative_AI_service.pdf (points businesses should confirm before entering personal data)
