Can you turn off model training on a free AI assistant?
We do not currently have an affiliate relationship with any product on this page, and no vendor paid for placement or review. Outbound product links are marked nofollow.
The Short Answer
- ChatGPT — OpenAI documents turning off "Improve the model for everyone" for signed-in and signed-out users alike, and says chat history survives the change.
- Claude — Anthropic's privacy article says chats are used for model improvement only "if … you choose to allow" it, and excludes Incognito chats even when the setting is on.
- Gemini — Google states that chats are reviewed by humans while Keep Activity is on, and that with it off they are "retained with your account for 72 hours".
| Product | Best for | From | Free plan | Free-tier training |
|---|---|---|---|---|
| ChatGPT | The switch works even signed out | — | Yes | Free opt-out |
| Claude | Framed as a choice you make | $17/mo (annual) | Yes | Free opt-out |
| Gemini | Publishes what happens after you opt out | — | Yes | Free opt-out |
What we checked, and what we did not test
The free tier of a general assistant is the most widely used free tier there is, and the thing you hand over for it is the conversation. So the question this page asks is the same one it asks of every category: what can you keep? Here, what you are trying to keep is control of your own chats.
We read the vendors’ own documentation on 6 August 2026 — OpenAI’s Data Controls FAQ, Anthropic’s article on whether your data is used for model training, and Google’s Gemini Apps privacy page — and pulled the sentences that describe the control and its consequences. Each verdict quotes and links them. We did not create accounts to inspect the default state of the toggles, which matters: where a page does not state the default, we do not supply one.
The expected split did not appear. All three publish a free opt-out, so the axis returns the same verdict three times, and the interesting differences sit just next to it: OpenAI is alone in offering the control without an account, Google is alone in publishing a retention number for opted-out chats, and Anthropic is alone in carving out a mode that ignores the setting altogether.
The switch works even signed out
1. ChatGPT — The switch works even signed out
ChatGPT is the only assistant here whose documentation describes the training control for people who never made an account.
Best for Turning training off without even having an account
From —. Free plan: Free — limited access to GPT-5.5 Instant, limited messages and uploads, limited image generation, deep research, memory and Codex.
- Go — price not shown on the page we opened · 1 seat (The page served to us priced this tier in KRW only (₩13,000 a month); we do not convert a currency into a USD claim)
- Plus — price not shown on the page we opened · 1 seat (KRW only on the page served to us (₩29,000 a month))
- Pro — price not shown on the page we opened · 1 seat (KRW only on the page served to us (from ₩159,000 a month))
OpenAI's Data Controls FAQ tells free users to "Turn off 'Improve the model for everyone'" in Settings → Data Controls, adds that "you can change this setting anytime. There are no restrictions", and uniquely offers the same switch to signed-out users. Turning it off does not delete your history: "They'll still show up in your history, but they won't be used to improve ChatGPT." Temporary Chats are "deleted from our systems after 30 days" and are not used for training.
Pros
- The training switch is available signed-out, which none of the other assistants here offer.
- The FAQ states plainly that switching it off leaves your chat history intact.
Skip it if
- The default runs toward training: the control is described as something you turn off, not something you turn on.
- The pricing page localises to your region's currency, so a USD figure is not necessarily what you will be shown.
Framed as a choice you make
2. Claude — Framed as a choice you make
Claude’s documentation frames training as conditional on your permission, and leaves the starting position of that permission unstated.
Best for Wording that puts model training behind an explicit choice
From $17/mo (annual). Free plan: Free — "Usage limits apply".
- Pro — $20/mo, or $17/mo billed annually · 1 seat ($200 billed up front for the annual option)
- Max — $100/mo · 1 seat (Published as "from $100 per month", with a choice of 5× or 20× Pro's usage)
Anthropic's privacy article says chats are used for model improvement "if: 1. You choose to allow us to use your chats and coding sessions to improve Claude", and that "your Incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your Privacy Settings". The page names Free, Pro and Max but does not state which way the setting is pre-set, so we are not claiming it is off by default.
Pros
- Incognito chats are excluded from model improvement even when the setting is on.
- Both monthly and annual prices are published in USD, which the other two assistants here did not show us.
Skip it if
- The article never states the default setting, so a free user cannot tell from it whether they are opted in.
- The Free plan's limits are given as the phrase "usage limits apply" with no number.
Publishes what happens after you opt out
3. Gemini — Publishes what happens after you opt out
Gemini is the only one that answers the question after the question: what is kept, and for how long, once you have said no.
Best for A published retention window for the data you opt out of
From —. Free plan: Free per Google account, with usage limits and access to 3.6 Flash plus some 3.1 Pro.
- Google AI Plus — price not shown on the page we opened · 1 seat (The page served to us priced every tier in KRW only (₩7,500 a month for this one))
- Google AI Pro — price not shown on the page we opened · 1 seat (KRW only on the page served to us)
- Google AI Ultra — price not shown on the page we opened · 1 seat (KRW only on the page served to us (from ₩119,000 a month))
Google's Gemini Apps privacy page says "a subset of chats are reviewed by human reviewers (including Google's trained service providers) to help improve Google services" while Keep Activity is on, that "you can turn this setting off or review and delete your activity anytime in Gemini Apps Activity", and that chats with the setting off "are retained with your account for 72 hours" to answer you and protect the service.
Pros
- It is the only one of the three that publishes a concrete retention window — 72 hours — for chats after you opt out.
- The opt-out lives in an activity control you can also use to delete past chats.
Skip it if
- Human review of a subset of chats is the stated behaviour while the setting is on.
- The subscription page priced everything in the local currency, so no USD figure could be recorded.
What this means if you are choosing today
If the deciding factor is the training switch, it is not a deciding factor — all three give it to free users at no cost. Pick on the adjacent facts instead. If you want to use an assistant without an account, ChatGPT is the only one documenting the control in that state. If you need to tell a colleague or a client exactly how long an opted-out conversation persists, Gemini’s 72-hour figure is the only published number of the three. If you want a per-conversation private mode rather than an account-wide setting, Claude’s Incognito chats are the only feature here described as overriding the account setting.
And on all three, do it before the first substantive conversation rather than after. Every one of these controls describes what happens to future chats; none of them is documented as retroactive.
Frequently Asked Questions
Does ChatGPT train on free-tier conversations?
Unless you turn it off. OpenAI's Data Controls FAQ instructs users to open Settings, go to Data Controls and "turn off 'Improve the model for everyone'", adds that "you can change this setting anytime. There are no restrictions", and describes the same control for signed-out users. Turning it off does not erase history: chats "still show up in your history, but they won't be used to improve ChatGPT".
Are Claude conversations used for training by default?
Anthropic's privacy article does not state a default. What it states is the condition: chats and coding sessions are used to improve models "if … you choose to allow us" to, across the Free, Pro and Max plans, and Incognito chats are excluded "even if you have enabled Model Improvement in your Privacy Settings". Because the default is not published there, we do not claim it is off.
What happens to Gemini chats after you turn Keep Activity off?
They are kept briefly and then not used for improvement. Google's page says chats with the setting off "are retained with your account for 72 hours", using the last 24 hours as context and to protect the service. While the setting is on, "a subset of chats are reviewed by human reviewers (including Google's trained service providers)".
Which one is best for privacy on a free plan?
On the switch itself, none of them wins: all three publish a free opt-out. The tiebreakers are elsewhere. ChatGPT is the only one that offers the control without an account. Gemini is the only one that publishes a retention window for opted-out chats. Claude is the only one that carves out a separate private mode that ignores the setting entirely.
Why are two of the three prices blank?
Because those pages priced in local currency for us. OpenAI's and Google's subscription pages served prices in Korean won on 6 August 2026 rather than US dollars, and converting a currency would produce a number that appears on no vendor page. Claude's pricing page published US dollars, so its figures are recorded and the other two are not.
How We Keep This Updated
Prices move and vendors rewrite their plan tables, so every value on this page carries the date we read it. When a page changes, the change lands here rather than being edited away quietly. The procedure is on the methodology page; if something above is wrong or out of date, tell us and we will re-check it.
- 2026-08-06 — First published. Training controls read in OpenAI's Data Controls FAQ, Anthropic's model-training privacy article and Google's Gemini Apps privacy page; OpenAI and Google prices recorded as unavailable in USD.