"Training off" doesn't mean deleted: what eight AI services keep anyway
Turning off training in ChatGPT, Gemini or Claude doesn't put you back in control of your data. Chats stay stored for weeks or years, one tap on the thumbs button overrides your opt-out, and humans may read along. We compared eight services side by side and looked at what happens to the emails, files and screenshots that AI agents read on your behalf.

What this is about
There's no shortage of guides on how to stop AI chatbots from training on your data. Malwarebytes published one in September, and Tom's Guide compared the toggles of all the major chatbots. What rarely gets answered is what happens after you flip the switch. How long do your chats stay on the servers? Are there back doors? Do humans read along? And does the toggle even cover the data AI agents now read in your name: emails, calendars, cloud files and screenshots from browsers running on the providers' servers?
Two news items from today prompted this look. Google has split Gemini's history and training into two separate toggles, and in the process sharply extended a retention period. And Ireland's Data Protection Commission (DPC) has stopped X from training on third-party data collected by its Grok Bot agent. We went through the help pages and privacy notices of eight services. One caveat up front: some pages, especially from OpenAI and xAI, couldn't be accessed directly. Where we rely on third-party summaries, we say so.
The comparison: what's left after you opt out
| Service | Training by default | Storage despite opt-out | Temporary chat | Feedback back door | Human review |
|---|---|---|---|---|---|
| Gemini | on (personal accounts) | History: 18 months by default; history off: about 2 months (previously 72 hrs) | about 2 months, no training | History off + feedback: last 24 hrs including attachments and app data; reviewed feedback up to 3 years | yes, a subset; reviewed chats up to 3 years, even after deletion |
| ChatGPT | on (Free, Plus, Pro) | until deleted; deleted chats gone within 30 days (with exceptions) | up to 30 days, no training | Thumbs up/down: entire conversation can go into training despite opt-out | authorized staff only, "need to know" |
| Claude | users had to choose in 2025 | deleted chats within 30 days; with training up to 5 years | Incognito: no training | Feedback: entire conversation kept 5 years, training "to the extent permitted by law" | Safety flags: content 2 years, ratings 7 years |
| Grok | on, per guides | until deleted, per guides | Private Chat: deleted within 30 days* | feedback possible despite opt-out* | safety review even with opt-out* |
| Meta AI | on | no clear retention periods found | – | – | unclear |
| Copilot | on (except in some countries) | 18 months by default | – | – | when policy violations are suspected, no opt-out |
| Le Chat / Vibe | on (individual users) | until deleted, per third-party sources | – | unclear* | unclear |
| Lumo | off, no training | no logs per Proton, history zero-access encrypted | guest chats end with the session | only explicitly submitted, anonymized feedback | none, per Proton |
* from third-party guides only, not verified against vendor pages.
Gemini is both the trigger and the clearest example. Until now, a single toggle called "Keep Activity" controlled two things: whether chats were stored and whether they went into training. Now there's "History" and a separate training toggle for chat content, including text and attachments. Sounds like more control. But according to Google's privacy notice dated November 16, chats are kept with your account for "about two months" when history is off, and the same goes for temporary chats. According to Notebookcheck, the October 26 version still said 72 hours. The Register calculates that's a twentyfold increase and notes that Google didn't even mention the old period in its announcement. Google says the two months keep chats available in case of system outages, and during that time it uses them to respond, process feedback and "protect Google, its users and the public," including with help from human reviewers. Google didn't respond to questions from The Register. If you had the old toggle off, both new toggles will be off too, according to 9to5Google. The change is rolling out over the coming weeks. According to Notebookcheck, a German account still showed the old setting on November 17.
Two details sit in the fine print. First, Google says the settings don't apply to anonymized data generated from chats. Second, chats that reviewers have looked at are kept for up to three years, separately from your account, even if you delete them.
The feedback button as a back door
The thumbs button is the most common way your opt-out gets overridden. On Gemini, with history off, giving feedback means Google also collects the last 24 hours of the related chats, including attachments and data from connected apps. Reviewed feedback is kept for up to three years. On ChatGPT, according to OpenAI's help page, the entire conversation can be used for training as soon as you rate a response, even if "Improve the model for everyone" is off. Anthropic stores the whole conversation, including style instructions, for up to five years when you give feedback, detaches it from your user ID and reserves the right to train models on it "to the extent permitted by applicable law." The page mentions no exception for users who've turned training off. Guides report something similar for Grok and Mistral, but we couldn't confirm it on the vendors' own pages. Proton only uses feedback you explicitly submit. It's anonymized and may be shared with ETH Zurich and EPFL for the Apertus model.
Deletion has its limits, too. Anthropic says data can't be removed from training runs already underway or from models already trained. Microsoft's Copilot opt-out explicitly excludes "general product improvements, advertising, digital safety and compliance." In the EU, Meta only offers an objection form. There, Meta trains on public posts and interactions with Meta AI. Meta says it doesn't use private messages "unless someone in the chat shares them with Meta AI features" – a clarification dated March 27, 2026. In the US, according to Tom's Guide, Meta offers no simple toggle.
Uncharted territory: what happens to your agents' data
Agents now read far more than the chat: inboxes, calendars, cloud drives, websites. Whether the training toggle covers all that varies from provider to provider.
| Service | Connector data (email, Drive …) | Browser and screen data |
|---|---|---|
| Gemini | used for training if the chat-content toggle is on | Screen-control screenshots: human review and use for improvement if history is on; remote browser data isn't deleted when you delete your history |
| ChatGPT | data from connected Google apps not used to train general models, per OpenAI, even with training on | Agent screenshots kept until deleted, then removed within 90 days; no training with opt-out |
| Claude | raw content from connectors and MCP servers never used for training, not even via feedback; anything you paste into the chat is | Data from Claude in Chrome counts as training data if training is on; the help article says nothing about the new cloud browser |
| Le Chat / Vibe | never used for training on any plan, per the help center; knowledge connectors like Drive are, however, indexed on Mistral's platform | no information found |
| Copilot | account profile data and email content excluded from training | images from Copilot Vision not used for training |
| Lumo | new desktop app with connectors (file system, shell, Proton Mail, Dropbox, Jira, MCP); the September 17 privacy policy doesn't mention them | – |
One Google detail: according to its help page, Gemini's new task features require history to be on. So if you use the agent, you can't turn history off at the same time, and that history is exactly what human review of screenshots depends on. The task features aren't available in the EU, Switzerland or the UK yet anyway. For the Claude cloud browser, Anthropic warns that Claude sees everything you type, even while you're in control yourself. Masking passwords and one-time codes doesn't work reliably, it says. For Anthropic's Computer Use API feature, screenshots are deleted after 30 days by default. Whether that also applies to the cloud browser isn't stated anywhere. We found no specific training rules for OpenAI's payment agent "dots" or ChatGPT Wallet.
The Grok Bot case: data from people who were never asked
So far this has been about your own data. But agents also read what other people wrote: a colleague's email, a friend's vacation photos, personal details on websites. That's exactly where the Irish DPC stepped in. As Politico Europe reported, with Dataconomy, heise and The Next Web following up, Grok Bot had been collecting this kind of third-party data since late October to train X's AI systems. After the regulator raised "significant concerns," X confirmed on Monday that it had suspended the processing. DPC Commissioner Dale Sunderland stressed that the regulator has nothing against agents as such. But it has to check whether companies comply with data protection law when they train on data collected by agents. The DPC is also in talks with another tech company about similar plans. Sunderland didn't name it because no processing has started there yet. X didn't comment. It's not known what legal basis X relied on.
The legal framework is vague. In its Opinion 28/2024, the European Data Protection Board (EDPB) stated that "legitimate interest" can in principle serve as a basis for AI training, but only after a case-by-case, three-step assessment (legitimate purpose, necessity, balancing) that takes into account what data subjects can reasonably expect. People also have a right to object. Our take: if you send an email to someone who uses an agent, you hardly expect to end up in a third party's training dataset, and you can't object because you never find out. The training toggle doesn't help here. It belongs to the agent's user, not the sender. Google, incidentally, openly states that it processes "personal data about others" that you give Gemini, for example when it summarizes an email.
A second case this week shows how little control users had over Grok Bot. According to security firm NewCore, the agent tried 31 times, unprompted, to read emails, calendars, Slack and GitHub on two laptops. Only on November 12, according to The Next Web, did a "Proactive messages" toggle arrive, and it's on by default.
What you can actually do
- Check both Gemini toggles separately once the change reaches your account. According to Digital Trends, the menus don't necessarily match the email announcement.
- Avoid feedback buttons in sensitive chats. Across all the major services, that's the most common route into training or human review.
- Temporary doesn't mean gone right away: up to 30 days on ChatGPT, about two months on Gemini.
- Don't paste content into the chat that would be protected via connectors. On Claude and Mistral, that protection disappears once you do.
- Log out in the cloud browser before handing control back, as Anthropic itself advises.
Our verdict
The training toggle controls how your data is used, not whether it's stored. Our take: by separating history and training, Google offers more choice, but at the same time makes things worse for people who previously had history turned off. On agent data, the picture is mixed. Anthropic and Mistral explicitly exclude connector content, while Google uses it for training if the toggle is on. Proton promises the most, but that rests solely on its own claims, and how the new Lumo connectors handle data is an open question. Meta, xAI and, on retention periods, Mistral are the most poorly documented.
The real gap is third-party data. No opt-out toggle protects people who never agreed to an agent. The fact that the DPC had to step in with X and is already talking to the next company shows that businesses see this data as raw material. So far there are only individual cases on the legal question, not a clear line. And there's an entirely different data leak on top of that: according to a Stony Brook University study, 21 of 42 AI chat services pass automatically generated chat titles to ad trackers, regardless of any training toggle.
Sources
- Digital Trends: Gemini trennt Verlauf und Training digitaltrends.com
- 9to5Google: New Gemini setting lets you keep chat history, but turn off AI training 9to5google.com
- Notebookcheck: Gemini separates history from AI training but keeps chats longer notebookcheck.net
- The Register: Opting out of Gemini tracking? Google will still hold your data 20x longer theregister.com
- Google: Gemini Apps Privacy Hub support.google.com
- Google: Gemini Apps Privacy Notice (UK-Fassung) support.google.com
- Google: Gemini Agent für mehrstufige Aufgaben support.google.com
- Dataconomy: Irish regulator halts X's AI training with agent-collected data dataconomy.com
- heise online: X stoppt Grok-KI-Training mit Drittanbieterdaten heise.de
- The Next Web: Grok Bot tried to read work apps unasked thenextweb.com
- Anthropic Privacy Center: Speicherdauer privacy.claude.com
- Anthropic Privacy Center: Werden meine Daten fürs Training genutzt? privacy.claude.com
- Anthropic Privacy Center: Computer Use privacy.claude.com
- Anthropic: Cloud-Browser in Claude support.claude.com
- OpenAI Help Center: How your data is used to improve model performance help.openai.com
- OpenAI Help Center: Temporary Chat FAQ help.openai.com
- OpenAI Help Center: ChatGPT agent help.openai.com
- OpenAI Help Center: Connectors in ChatGPT help.openai.com
- Microsoft: Privacy FAQ for Microsoft Copilot support.microsoft.com
- Microsoft: Copilot privacy controls support.microsoft.com
- The Decoder: Microsoft trainiert mit Copilot-Daten, EWR ausgenommen the-decoder.com
- Meta: Making AI Work Harder for Europeans about.fb.com
- Mistral Help Center: Opt-out vom Training help.mistral.ai
- Enterprise DNA: Mistral trainiert standardmäßig mit Nutzerdaten enterprisedna.co
- Proton: Lumo Privacy (Support) proton.me
- Proton: Lumo Privacy Policy proton.me
- Caschys Blog: Proton Lumo Desktop-App bringt Connectors stadt-bremerhaven.de
- EDPB: Opinion 28/2024 zu KI-Modellen (PDF) edpb.europa.eu
- Malwarebytes: How to opt out of AI chatbot training malwarebytes.com
- Tom's Guide: I checked the privacy settings of every major AI chatbot tomsguide.com
Ask the AI assistant
Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.
The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.
Comments
What do you think? Already tried it? Write it here, no sign-up needed.
Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.