Which AI plans train on your work — and how to opt out
Five assistants, five different defaults. One of them is opt-out, one is opt-in, one hides the switch in a Google-wide activity page, and none of them behave the way "I pay for it, so it's private" would suggest. Here is what each provider's own documentation says, and what the off switch really covers.
The short answer
Two things decide whether your input can be used to train a model, and neither of them is the price of your plan.
The first is which way the provider ships the default. On consumer ChatGPT tiers, OpenAI's data controls FAQ describes content as something that may be used to improve models unless you opt out — that covers Free, Plus, Go and Pro alike (OpenAI help centre, checked September 2026). Perplexity's help centre describes its Free, Pro and Max tiers the same way: retention for model improvement is on unless you turn it off (Perplexity help centre, checked September 2026). Anthropic runs the opposite default on consumer Claude, where the privacy centre states that chats and coding sessions are used to improve models only if you have explicitly chosen to allow it (Claude privacy centre, checked September 2026).
The second is the kind of account you are signed in with. This is the part worth remembering, because it is the only lever that changes the default without you touching a switch. OpenAI states that ChatGPT Business, Enterprise and Edu, and the API, do not use your content to train models by default (OpenAI, checked September 2026). Microsoft says data from users signed in with an organisational Entra ID account is not used to train foundation models. Perplexity says Enterprise data is never used for training or fine-tuning, and that its agreements with third-party model providers forbid them training on it too (Perplexity Enterprise data page, checked September 2026).
So "I pay for Plus, therefore my work is private" is the wrong sentence. A consumer Pro subscription and a free account sit under the same policy at the two providers that default to on. If you want to know what the extra money does buy, we went through that separately in what $20 a month actually gets you.
ChatGPT: on by default, one switch
The path is Profile → Settings → Data controls, and the setting is labelled "Improve the model for everyone". Turn it off and, per OpenAI's own description, new conversations stop being used for model training. The label here is one of the few we are quoting without hedging, because OpenAI's help article names it in those exact words (OpenAI help centre, checked September 2026).
The switch has a narrow reach, and that is where people get caught out. That article exists precisely to explain that you can keep chat history on while training is off; the sidebar does not empty itself. Deletion is a third, separate thing again — OpenAI's retention page says deleted conversations leave your view immediately and are permanently deleted within 30 days, and Temporary Chats may be kept up to 30 days for safety purposes (OpenAI, checked September 2026).
Claude: off unless you said yes
Consumer Claude — Free, Pro and Max — is the opt-in case. The control sits in Privacy Settings, under the Model Improvement item (label as of September 2026; the live settings page wins).
What deserves more attention than the switch is the price of saying yes. Anthropic's retention page and its consumer terms update both state that allowing your data to be used for model improvement extends retention from 30 days to five years, on de-identified data (retention page; consumer terms update, both checked September 2026). That is a much bigger swing than the toggle itself implies, and it is the reason this one is worth a deliberate decision rather than a shrug.
Anthropic's retention page fills in the rest. Deleted conversations leave your history immediately, are removed from the back end within 30 days, and are not used for future training. Incognito chats are excluded from training even when model improvement is switched on. And whatever you choose, sessions that automated trust and safety systems flag as violating policy can be retained for up to two years, with classification scores kept for up to seven.
Existing users had to accept the updated consumer terms and make a training choice by 8 October 2025 (Anthropic, updates to our consumer terms, checked September 2026) — so if you have used Claude since before then, you already answered this once, possibly by clicking through a dialogue at speed. Worth going back to look.
Gemini: an activity switch with a three-year asterisk
For Gemini the control lives one level up, in Google's account-wide activity system. It is called Gemini Apps Activity, and it sits at myactivity.google.com/product/gemini. While it is on, Google's support page describes an automatic deletion window defaulting to 18 months, adjustable to 3 months, 36 months, or no auto-delete at all (Google support, page last updated 10 August 2026, checked September 2026).
Switch it off and future conversations stop being saved to your activity and stop being used to improve Google's AI models. They are still stored for up to 72 hours so the service can return responses, handle feedback and protect Google — a short window, but not zero.
Then there is the asterisk, and it is the single most surprising line we read while checking this. Some conversations are reviewed by human reviewers, including trained service providers. Those are disconnected from your account before review — and, per Google's page, a conversation that has been selected for human review is kept separately for up to three years and is not deleted when you delete your Gemini Apps Activity. If you assume "I cleared my history, so it's gone", that assumption has a three-year hole in it.
Copilot: two switches instead of one
Microsoft's consumer path is the most explicit of the five. On the web, go to copilot.com, click your avatar, then your user name, then Privacy. You are looking for "Training on conversation activity" and "Training on voice conversations". In the Windows and macOS apps it is avatar → Settings → Privacy, and the same pair. Microsoft's page notes you can switch training off while keeping personalisation on, which is a distinction the other providers do not offer as cleanly (Microsoft support, checked September 2026). These two labels, like OpenAI's, we are quoting as documented rather than hedging.
On the commercial side Microsoft is unusually direct: data from users signed in with an organisational Entra ID account is not used for training; Microsoft 365 Copilot prompts, responses and the file content it accesses are not used to train foundation models; and Copilot inside apps covered by a Microsoft 365 Personal or Family subscription is likewise not used for training (Copilot privacy FAQ and Microsoft Learn, checked September 2026). That Personal/Family line is the one place on this page where consumer money does change the answer.
Perplexity: default on, label unstable
Free, Pro and Max all ship with AI data retention for model improvement switched on (Perplexity help centre, checked September 2026). The control is in Settings, under AI Data Usage / AI Data Retention (label as of September 2026; the live settings page wins). We flag that harder here than elsewhere because the two Perplexity help articles we read did not phrase the same control identically, which usually means the interface has moved and the documentation is catching up. Enterprise accounts are documented separately as never used for training or fine-tuning.
What opting out does not do
In rough order of how often we see them assumed:
- It does not delete your history. This is the OpenAI case above, and it generalises. Training controls and history controls are separate everywhere on this list.
- It does not mean zero retention. Google's 72-hour window is the explicit example, and safety, abuse-monitoring and legal retention exist at every provider here. Opting out narrows the purposes your input may serve; a short working copy still gets made.
- It does not reach backwards. These switches are forward-looking by design. If you spent six months with the default on, that period was governed by the default.
- It does not change your account type. If the real problem is that client work is going through a personal account, a toggle is the wrong fix.
Turning training off on a consumer account is worth doing, and we would recommend it.
FAQ
Does paying for a Plus or Pro plan stop my chats being used for training?
No. Price is not the variable. OpenAI's help pages put Free, Plus, Go and Pro in the same consumer bucket, where content may be used to improve the model unless you opt out, and Perplexity's help centre describes Free, Pro and Max the same way. What changes the default is the kind of account — business, enterprise, education, organisational or API — not how much the consumer plan costs.
If I turn training off, does my chat history disappear?
No, and OpenAI documents this explicitly: you can disable model training and keep your history, in which case new conversations still appear in the sidebar but are not used to improve the model. Turning off training and deleting history are two separate actions with two separate switches.
I deleted my conversations. Is that the same as opting out?
Not quite, and Gemini is the clearest counter-example. Google's support page says conversations selected for human review are kept separately for up to three years and are not deleted when you delete your activity. Deletion is about what sits in your account; opting out is about what happens to material before it ever gets there.
Which account types are documented as not training on your input?
As the providers' own pages read in September 2026: OpenAI says ChatGPT Business, Enterprise and Edu, and the API, do not use your content to train models by default. Perplexity says Enterprise data is never used for training or fine-tuning. Microsoft says data from users signed in with an organisational Entra ID account is not used to train foundation models. For Anthropic's commercial tiers we did not obtain a clean primary source, so we make no claim here — check the official commercial privacy page.
Does switching training off mean nothing is retained at all?
No. Every provider here keeps some material after you opt out — for abuse monitoring, safety review, support and legal obligations. Google, for example, says conversations are still stored for up to 72 hours after Gemini Apps Activity is off, so the service can respond and process feedback. Opting out changes what your input is used for, not whether it exists.
If the answer to all this is that you want fewer accounts rather than better settings, start with how to cancel the ones you don't use — or, if the charge has already landed, how to ask for that money back. Then look at which free tiers are good enough to fall back to. Our ChatGPT review covers what the paid tier actually buys, and how we test explains what we do and don't verify ourselves. Settings pages move; if a path here no longer matches what you see, the provider's current page wins — and tell us so we can correct it.
Filed under：Managing Your Subscription