Which AI assistants train on your data?
Of the 25 AI assistants on file, 24 are graded: 21 train on your data by default and 3 don't. One more has a policy that doesn't say either way. The best graded is Cohere (B) and the worst is Runway (F).
Do they train on your prompts, your files and your recordings?
What these apps hold is the conversation itself: the prompt, the file you attached, the recording you made, the answer you rated. They’re graded here on whether that trains a model, and how hard it is to stop. Read together, they divide by the way out their documents offer. None, one that’s sold, a switch, or permission asked first. Or a privacy policy that never settles it.
The ones with no way out, or a way out that is sold
Otter lists “training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)” among its uses. De-identification is claimed for the audio and not for the transcripts, in the same sentence, and the bracket concedes those may carry personal information. The opt-out language the policy does contain covers marketing email, analytics and state-privacy disclosures. None of it touches training. The right to object that would reach it is gated by where you live.
Runway’s privacy policy never reaches the subject. Its terms of use do: “You acknowledge that Inputs and Outputs may be used by the Company to train and improve its AI models, algorithms and related technology”. The licence behind that the terms call “non-exclusive, irrevocable, perpetual, worldwide, royalty-free, fully paid, transferable, sublicensable”. “Opt out” appears nine times in the terms and every one is the arbitration clause. In the privacy policy it’s marketing email. No document names a way to decline, which is an F.
DeepL’s policy splits by product. The free version: “We process the content you upload and their translations or improvements for a limited period of time to train and improve our neural networks and algorithms.” Pro: “your texts will not be used to improve the quality of our services”. No setting sits between them. The way to stop training on a free account is to pay, which is what an E is.
The ones with a switch, and what the document says about it
Mistral’s clause is a denial wrapped around an admission. “We do not use Your Data to train our artificial intelligence models except (a) when you have not opted out of training or (b) when you provide Feedback to us.” Training runs until you turn it off. The terms state that default, and the privacy policy says the control is in your account, which is what a C means here. Limb (b) is the hole. A rated answer goes to training whatever the switch says.
Claude trains on Free and Pro “unless you opt out through your account settings”. The next sentence carries the exceptions: “Even if you opt-out, we will use Inputs and Outputs for model improvement when: (i) your conversations are flagged for safety review”. The location of the control is named. The setting isn’t, and nothing says which way it ships. The business terms run the other way: “Anthropic may not train models on Customer Content from Services.”
Google’s hub for the Gemini app says it uses your activity “to provide, develop, and improve its services (including training generative AI models)”, with the help of human reviewers. The switch is Keep Activity. The page states a default for audio and video, never for Keep Activity itself, and that’s the line between C and D here. Chats reviewers have read outlive your deleting your activity.
The ones that ask first
Cohere’s privacy policy points to a separate Model Training Privacy Notice. The notice answers that in most cases “Cohere has no access to inputs submitted to its Models and other products”. And that “Where a user has given Cohere permission to use inputs and outputs for model training, we take steps to de-identify” them first. Permission before use is what a B is.
Descript’s privacy policy reads as if training were on: you opt out, it says, by disabling a setting. The help article it links to says the opposite: “for any in-house AI training we maintain a clear data sharing opt-in, and otherwise data is not used for AI training.” Descript disputed an earlier reading of the policy and was right. The reply is printed on the verdict.
You.com’s privacy policy now says it doesn’t use “the substance of your prompts, files, or generated outputs” to train its own models, “except where you have separately opted in”. Its terms of use, dated 27 August 2024, still let it use your Content to “train and improve the Services”, and neither document says which wins. B on the newer and narrower promise, with the conflict noted on the entry.
The ones whose privacy policies do not settle it
Manus’s policy, revised on 28 September 2026, now says de-identified data made from your personal information may be used “to develop and improve our models and related technologies”. Then: “Depending on your jurisdiction, you may opt out by contacting us at the email address provided below.” It also mentions content used “when you allow us to use your content to improve our models”, and names no setting. The entry has no grade until Manus has had the chance to answer.
Side by side
6 pairs of AI assistants people choose between, with both verdicts on one page.