Short answer: yes, by default, ChatGPT can use your consumer conversations to train its models — and so can most of its competitors. But “train on your chats” is only one of three separate things happening to your text, and the famous opt-out toggle controls just one of them. It does nothing about whether your messages are stored, whether a human might read a sample, or whether a court order freezes your “deleted” chats in place.
This is the honest, no-spin version: what the toggle actually does, what each major provider does by default, whether deleted conversations really disappear, and the one architecture where the question stops mattering entirely.
What ChatGPT’s opt-out toggle actually controls
In ChatGPT, the relevant setting lives under Settings → Data Controls → “Improve the model for everyone.” Per OpenAI’s Data Controls FAQ, for standard consumer accounts (Free, Plus, Pro) this is on by default, which means your conversations may be used to train future models unless you turn it off.
Here is what that toggle does not do, and this is the part people miss:
- It does not stop your messages from being stored on OpenAI’s servers.
- It does not stop a human reviewer from seeing flagged content for safety/abuse review.
- It does not retroactively pull your old chats out of any model already trained on them. Per OpenAI’s own framing, the setting affects future conversations only — anything already in a training set stays there.
So the toggle is a training-input control, not a privacy control. It answers “will my words help build the next model?” It does not answer “where does my text live, who can see it, and for how long?” Those are three different questions, and conflating them is the single biggest mistake people make about cloud AI privacy.
Storage vs. training vs. human review — three different things
Treat these as three independent dials, because the providers do:
| Dimension | What it means | What the opt-out toggle does |
|---|---|---|
| Storage / retention | Your messages sit on the provider’s servers, tied to your account, for some retention window | Nothing — storage continues regardless |
| Model training | Your text becomes training data for the next model version | This is the dial the toggle moves |
| Human review | A staff member or contractor reads a sampled or flagged conversation | Usually unaffected; safety review is carved out of opt-out |
You can opt out of training and still be fully stored and still be subject to human review of flagged content. That is not a loophole — it is how essentially every hosted assistant is built, because abuse detection and legal compliance require keeping and inspecting data. Our broader AI data privacy guide walks through each of these dials in more depth. The point here: opting out of training is the weakest of the three guarantees, and it is the only one most people ever touch.
Per-provider defaults: ChatGPT, Gemini, Grok, Claude, Copilot
Defaults are what actually matter, because almost nobody changes them. Here is the landscape as of mid-2026, based on each company’s published policies. Where a policy is nuanced, treat this as a starting map and read the current terms yourself before trusting anything sensitive to a cloud box.
| Provider | Trains on your chats by default? | Opt-out control | Notable detail (per published policy) |
|---|---|---|---|
| ChatGPT (consumer) | Yes | Data Controls → “Improve the model for everyone” | Deleted chats kept up to ~30 days for abuse review; Temporary Chat isn’t trained on but still retained briefly |
| Google Gemini (apps) | Yes | Turn off Gemini Apps Activity | Google states a subset of chats are reviewed by humans and those can be retained up to 3 years, even if you delete your activity; Google explicitly warns: don’t enter confidential info |
| Grok / xAI | Yes | Settings → Data → “Improve the model”; Private Chat opts out automatically | Default is opt-in for interactions; turning it off only stops future use |
| Anthropic Claude (consumer) | Yes, after a 2025 policy change | Privacy Settings toggle | Opted-in data can be retained in de-identified form up to 5 years; opt-out keeps the older ~30-day window; Commercial/API and Work/Gov/Edu tiers are excluded |
| Microsoft 365 Copilot (enterprise / EDP) | No | N/A — protected by Enterprise Data Protection | Microsoft states enterprise prompts/responses aren’t used to train foundation models; consumer Copilot tiers follow different, weaker terms |
A few honest caveats. Google’s three-year human-review window (from its Gemini Apps Activity documentation) is the most aggressive published retention figure of the bunch, and Google’s own “don’t enter confidential information” line tells you how seriously to take it. Anthropic’s shift to an opt-out-by-default consumer model was widely reported in 2025; the enterprise/API side did not change and remains the privacy-respecting tier. Microsoft’s enterprise commitment is real and documented, but it is specifically the paid business product — free consumer Copilot is governed by Microsoft’s consumer privacy terms, which are not the same promise.
The consistent pattern: consumer tiers train by default; paid enterprise/API tiers usually don’t. If you’re on the free version of anything, assume the most data-hungry default is in effect until you’ve personally verified otherwise.
Do deleted conversations still exist? Backups, safety holds, and legal holds
This is where “delete” stops meaning what you think it means. Across providers, hitting delete typically does three predictable things and then hits a wall:
- The conversation disappears from your UI immediately. Good.
- A copy persists on backend systems for a retention window — commonly cited as up to 30 days for OpenAI — for abuse and safety review before scheduled permanent deletion.
- Anything already used to train a model stays baked into that model. You cannot un-bake a cake. Deletion removes the record, not the statistical imprint your text may have left in a trained model.
Then there’s the wall: legal holds. When a company is under litigation, a court can order it to preserve data it would otherwise delete — overriding its own retention policy. OpenAI publicly acknowledged exactly this dynamic in the New York Times litigation, where it was compelled to preserve output data it would normally have deleted. The lesson is structural, not specific to OpenAI: any centrally stored data can be frozen in place by a third party’s lawsuit, no matter what the provider’s privacy policy promises. Your delete button is subordinate to a judge’s order.
Why “opt out” is not “never stored”
Put the pieces together and the gap is obvious. “I opted out of training” means:
- ✅ My future chats shouldn’t feed the next model.
- ❌ My chats are still stored, still tied to my account, still inside the retention window.
- ❌ Flagged content can still be human-reviewed.
- ❌ A legal hold can still freeze everything.
- ❌ Anything already trained on is already gone.
Opt-out is a real and worthwhile setting — turn it on everywhere. But it is a policy promise about one data flow, enforced by the same company that profits from your data, revocable by a terms-of-service update, and overridable by a court. It is not a technical guarantee. If you wouldn’t email a sentence to a stranger, an opt-out toggle is not the thing that makes it safe to type into a cloud chatbot. This is the same reason your employer can sometimes see your usage — covered in can your employer see your ChatGPT history — and it’s related to why cloud models behave the way they do, which we unpack in why cloud AI censors you.
The only structural guarantee: local
There is exactly one configuration where “does it train on my chats?” becomes unanswerable-because-impossible: when the model runs on your own machine and your text never leaves it.
Run a model locally with Ollama and the conversation happens entirely on your hardware. The API listens on loopback only — 127.0.0.1:11434 — so there is no server, no account, no retention window, and no opt-out toggle, because there is nothing on the other end to store or train on. You can verify it the blunt way: pull the network cable, and a local model keeps answering. A cloud one goes dark.
# install Ollama (Linux/macOS)
curl -fsSL https://ollama.com/install.sh | sh
# run a model fully on your machine — no cloud, no account
ollama run llama3.1The cost is real: you need a capable GPU (VRAM drives how big a model you can run), and a local model won’t match the absolute frontier of the biggest hosted systems. But for privacy, local isn’t a policy — it’s physics. There is no toggle to misconfigure, no terms update to ambush you, no subpoena that can reach a machine that never sent your words anywhere. For the full case, see private ChatGPT alternative with no data collection.
Zero-retention hosted as a middle ground
Local is the strongest guarantee, but it asks you to own and run hardware. The honest middle ground is a hosted service that doesn’t retain or train on your conversations — you give up the physics-level guarantee of local, but you get a much narrower promise than a default consumer chatbot: no training on your chats, minimal or zero message retention, no scraping your conversation history into a model.
That’s the lane Freya occupies for people who want a private AI companion without buying a GPU or configuring anything — a hosted experience built so your conversations aren’t mined for training the way default consumer chatbots are. It’s not the same airtight guarantee as running the model yourself; it’s a deliberately smaller, clearer promise than “we train on you unless you find the toggle.”
Your action plan
If you take nothing else from this page, do these, in order of how much they actually protect you:
- Turn off training everywhere now. ChatGPT: Data Controls → “Improve the model for everyone” off. Gemini: turn off Gemini Apps Activity. Grok: Data → “Improve the model” off. Claude: flip the training toggle in Privacy Settings off. It’s the weakest guarantee, but it’s free and takes two minutes.
- Assume stored ≠ deleted. Treat anything you type into a cloud chatbot as retained for a window and potentially freezable by a legal hold. Don’t paste anything you couldn’t survive being read.
- Heed Google’s own warning — “don’t enter confidential information” — and apply it to every cloud assistant, not just Gemini.
- For genuinely sensitive work, move the model, not just the toggle. Use a zero-retention hosted service or run it local so the question can’t be asked.
The toggle is the floor. Local is the ceiling. If you want the version of this where “does it train on my chats?” is structurally impossible — because the model lives on your machine and the conversation never leaves it — start with how to run AI locally, or skip the hardware entirely with a hosted companion built not to mine your chats.