Free AI tiers and your data: what trains the model

The most consequential setting in any free AI account is the one almost nobody opens. On several of the largest consumer plans, what you type is used to improve the vendor's models unless you say otherwise; on others, the vendor states it does not train on your content at all; and on a few, this site could not read the policy. This note collects what the privacy pages and help centres said on October 7, 2026, product by product, and ends with a day-one routine that takes five minutes.

The default is a choice the vendor made for you

A training setting has two possible defaults. Either your content is used until you turn it off, or it is not used until you turn it on. Across the 22 reviews on this site, the first default is more common among the general assistants and the second is stated plainly by only a handful of vendors. Neither default is hidden; all of them are documented. The problem is that the document is a privacy page or help article that nobody reads on the day they sign up, and by the time they do, the material is already in. The point of this note is to move that reading forward by a week.

Plans where training is on unless you turn it off

ChatGPT: OpenAI's data controls FAQ states that by default, conversations on consumer accounts can be used to train its models. The setting is named 'Improve the model for everyone', and turning it off means new conversations are not used for training; the FAQ says the setting applies across your devices when signed in. Content from Business, Enterprise, Edu and healthcare workspaces is not used for training by default regardless of individual settings. Memory is a separate control: turning memory off does not by itself change whether conversations are used for training.

Mistral Le Chat, now called Vibe: the help centre's opt-out article states that only Enterprise customers are opted out by default. Users on Free, Pro and Education plans have inputs and outputs eligible for training unless they opt out, and the article adds a sentence worth underlining: documents attached or uploaded within Vibe are considered input data. The opt-out is documented step by step: on the web, the Privacy section of the admin panel at admin.mistral.ai has a toggle labelled 'Allow your interactions to be used to train our models'; on iOS and Android it is Settings, then Data and Account Controls, then the data-sharing checkbox. The article notes the Vibe toggle and the API toggle are separate.

Google Gemini: with the Keep Activity setting on, Google's Gemini Apps privacy hub states that your chats and what you share are saved, that human reviewers assess chats for quality and safety, and that your activity is used to improve Google's AI models. Standard activity auto-deletes after 18 months by default, adjustable to 3 or 36 months. Chats reviewed by a human are retained for up to three years, and the hub says this applies even after you delete them. The two controls, turning off Keep Activity or using temporary chats, do not require a paid plan.

Microsoft Copilot: the privacy controls page gives the path to a training toggle under the profile icon and Privacy, with a separate toggle for voice conversations. Then it qualifies the opt-out in a way that deserves quoting.

The setting will not exclude your conversations from being used for other general product or system improvements nor from use for advertising, digital safety, security and compliance purposes.

That is from Microsoft's privacy controls page for Copilot on a personal account, checked October 7, 2026. The training toggle is a training toggle, not a do-not-use-my-data toggle; conversations can still feed product improvement and advertising after you opt out. Personalisation and memory are a separate switch, and conversation history can be deleted item by item or in full.

Claude: Anthropic's privacy center article frames training on consumer chats as a choice that lives in your privacy settings. Incognito chats are not used to improve the model even when the setting is on. Raw connector content, such as files reached through a connected Google Drive, is excluded from training unless you copy it into a conversation. Feedback from the thumbs buttons is stored for up to five years and de-linked from identity before use; the article does not state a retention period for ordinary chat data used in training, and this site will not guess one. Open the setting on day one and decide it deliberately.

Grammarly: the privacy policy, which on the day of checking redirected to superhuman.com and names Superhuman Platform Inc. as the controller, states that the company uses personal data to develop and improve AI subject to your settings, and that you can decide whether it may use your content to train its models through training controls in account settings. This site cannot tell you the default state of that control on a new free account; find it before writing anything you would not want in a training set.

Otter.ai: the privacy policy states that Otter trains its proprietary AI on de-identified audio recordings and on transcriptions, and notes in the same sentence that transcriptions may contain personal information. The policy read for the review did not describe a consumer-tier opt-out; check current account settings for the tier you are on, because for a tool that records meetings this is the single most consequential setting in the product.

Vendors who state they do not train on your content

tl;dv: the privacy policy states that tldx Solutions GmbH does not use customer content to train, fine-tune or improve foundation models, will not sell collected data, and that staff do not access recordings unless a user grants support access. Retention differs by tier: free users' recordings are kept for three months, paying users' until account deletion.

Fireflies.ai: the privacy policy states that the vendor does not use personal information for AI model training and contractually prohibits its vendors from doing so, and describes a zero-data-retention arrangement with the third parties that process meeting content. Account data is deleted within 30 days of closure.

Adobe Firefly: the product page states that Adobe does not train Firefly models or partner models on any Creative Cloud subscribers' personal content, and that Firefly models are trained on licensed Adobe Stock and public-domain content. The statement names subscribers; whether an identical commitment covers a free account with no subscription should be confirmed in the terms for the free tier rather than assumed.

Canva: the AI page states that Canva does not use your content to improve AI-powered features unless it is consistent with your privacy controls, and that granular privacy settings exist on Free and Pro. That sentence is conditional on the controls, so the controls, not the sentence, decide what happens; find the default on a new free account.

Where this site could not read the answer

QuillBot's privacy policy returned an access error to every fetch attempted on October 7, 2026, so that review cannot say whether pasted text trains models or how long it is kept; it says so plainly rather than paraphrasing a third party. Runway's documentation reviewed was about credits, not data, and the review lists the questions to ask instead. ElevenLabs, Suno, Descript and Cursor are in the same position: the pages checked were pricing and usage pages, and the training default is a question for the current policy. Gamma offers one useful inference: its Business plan lists a locked training opt-out so that workspace members cannot change the data-usage setting, which implies an unlocked version at lower tiers that you should find and read. NotebookLM and Perplexity reviews likewise send you to the policy for your exact tier, because the terms differ between consumer and organisational accounts.

Temporary modes and what they actually do

Several vendors offer a chat mode outside history and training, and the details differ in ways that matter. ChatGPT's Temporary Chat is not used for training and does not appear in history, but the FAQ says it may be retained for up to 30 days for safety purposes, and saving it turns it into a regular chat. Gemini's temporary chats, with Keep Activity off, are still kept for 72 hours so the product can respond, process feedback and protect Google. Claude's incognito chats are excluded from training even when the general setting is on. Mistral's developer documentation says temporary conversations are not saved to history or used for training, with only usage logging except where legally required. A temporary mode is the right tool for a one-off sensitive question; it is not a substitute for setting the default.

The switch is not the whole answer

Four things survive any toggle. Human review: Gemini's hub states that reviewers may read chats with activity on, and reviewed chats outlive deletion by up to three years. Other uses: Copilot's page says the training opt-out does not cover product improvement or advertising. Uploads: Mistral treats attached documents as input data, and the reviews for Descript, Runway and ElevenLabs point out that recordings of a voice or face are a more sensitive category than text. And the tier line: ChatGPT, Gemini, Mistral, Copilot and NotebookLM all draw a documented distinction between consumer accounts and organisational workspaces, and that distinction is the main reason a business should not let staff paste client material into personal free accounts. The same prompt in a Business workspace sits on different terms.

A day-one routine

• Before the first real prompt, open privacy or data settings and find the training control. Decide it; do not assume it.
• Find the temporary or incognito mode and learn its retention window, because it is not zero.
• If the tool records people, find the consent mechanism and say out loud that you are recording anyway; the policy puts that obligation on you.
• If you upload files, read whether uploads count as input data and how long they are retained after deletion.
• Keep client, employee, patient and contractual material off personal free accounts entirely, whatever the toggle says. That is a tier question, and the decision belongs to whoever owns the risk.
• Re-read the policy when the product renames itself; Mistral's and Grammarly's both moved in 2026.

Bottom line

As of October 2026, the largest consumer assistants train on free-tier conversations unless you turn it off, and the switch is documented on every one of them: ChatGPT's 'Improve the model for everyone', Mistral's admin-panel toggle, Gemini's Keep Activity, Copilot's Privacy toggle with its advertising caveat, Claude's privacy settings. tl;dv, Fireflies and Adobe state they do not train on customer content; Canva and Grammarly make it conditional on settings you have to find; QuillBot's policy could not be read. None of this is a reason to avoid the free tiers. It is a reason to spend five minutes on day one, and to treat a prompt as a disclosure regardless of what the toggle says. How this site grades a free tier explains why documented controls count toward the support score, and the method page explains why every statement here carries a date.