Guide · updated November 4, 2026
Which AI tools train on your data, and how to opt out
What you type, upload or record can end up improving someone else’s model. Some vendors do it by default on consumer plans, some never do it, and many sit in between with a switch buried in settings.
For each tool we looked for three things in its own documents: whether user content is used for training by default, whether you can turn that off, and how long the content is kept.
A0 toolsB4 toolsC7 toolsD8 toolsE1 tool
| Tool | Grade | What the documents say |
|---|---|---|
| Aragon AI $35 one-time (headshots); app plans from $15/month | B | Face data is permanently deleted or anonymized within 30 days of the earliest of: delivery, a deletion request, account closure, an organization-set period, or one year after collection, though active subscribers may keep it until their plan ends. Aragon says it contractually bars its AI providers from training generalized models on face data and does not sell face data. It does keep a right to use anonymized or aggregate data for service improvement. Source |
| Descript $16/person/month billed yearly (Hobbyist; $24 month to month) | B | The Terms allow Descript to use inputs and outputs to train its models unless you opt out. The help centre explains how: App Settings, Profile tab, 'Share data with Descript'. It adds that current production models use no Descript user data and that opted-out users' data will not be used at any stage. Users are asked for a preference at onboarding; Enterprise drives have sharing disabled by default. Source |
| HeadshotPro $29 one-time | B | The terms say input photos are deleted automatically 30 days after your gallery is generated, or sooner if you delete them, and that input photos and outputs will not be used to train HeadshotPro's or third-party generative models. Another clause still allows training its algorithms on data derived from your content, uploaded photos excluded. The privacy policy has no biometric section and gives 14 days to erase data after a confirmed request. Source |
| Synthesia $14/month billed yearly (Starter; $19 month to month) | B | Synthesia's AI governance page states that it does not use Customer Data, including inputs or outputs, to pre-train its AI components, and fine-tuning on customer data happens only on the customer's written instruction. The Customer Terms still let it analyze usage information to improve the services. Biometric data for an avatar is kept as long as the avatar exists on the platform. Source |
| ChatGPT $8/month (Go, US price) | C | On personal Free, Plus and Pro workspaces, 'Improve the model for everyone' is on by default, so your chats can train OpenAI's models until you turn it off. The opt-out is easy and keeps your history, but a thumbs-up or thumbs-down can still send that whole conversation to training. Voice mode and Codex environments have their own separate training switches. Temporary chats are not used for training and may be kept up to 30 days for safety. Business, Enterprise and Edu workspaces are not used for training by default. Source |
| Claude $17/month (Pro, billed yearly at $200) | C | On consumer plans (Free, Pro, Max) you choose whether your chats and coding sessions are used to train models, and the pricing page lists training as something you can switch off. If you allow it, de-identified data can stay in training pipelines for up to 5 years. Deleted chats leave back-end storage within 30 days. Chats flagged for policy violations can be kept for 2 years, and anything you send through thumbs-up or thumbs-down is kept for 5 years. Turning training off does not pull your data out of models already trained. Team plans do not train on your content by default. Source |
| Cursor $8/month (Plus; $80 billed yearly) | C | The terms say Anysphere 'WILL NOT USE CONTENT TO TRAIN, OR ALLOW ANY THIRD PARTY TO TRAIN, ANY AI MODELS' unless you explicitly agree. The data-use page adds that with Privacy Mode off, Cursor 'may use and store codebase data, prompts, editor actions, code snippets' to improve features and train models. None of the pages we read states whether Privacy Mode is on by default for individuals; the docs say it is on by default only for Enterprise teams. Cursor says it holds zero-data-retention agreements with its model providers. Source |
| ElevenLabs $5/month billed yearly (Starter; $6 month to month) | C | Training is on by default. The Terms give ElevenLabs a perpetual, royalty-free license to use your content to improve the services and models, and the privacy policy confirms inputs and outputs are used for training. There is a real self-serve opt-out under the Data use menu in the account's Terms and Privacy section, but it only applies going forward. Data generated about your voice is kept up to 3 years after your last interaction. Source |
| GPTZero $12.99/month billed $155.88 yearly (Premium) | C | Text scanned through the website dashboard is stored; the help center says this data 'is only used in aggregate by GPTZero to further improve the service'. API documents are not stored by default unless you turn on scan persistence. The privacy policy (updated April 2, 2026) says personal information is kept no longer than three months after account termination, but it does not set a retention period for scanned text or offer a training opt-out for dashboard users. Source |
| Grammarly $12/month billed yearly ($144/year) | C | The support article on the Product Improvement and Training control says it is on by default for individual users (Free, Premium and single-user Pro) and for multi-user Pro bought on the website, where only admins can change it. It is automatically off for Enterprise, Pro, Business and Education accounts bought through sales. When it is off, Grammarly says the content you type is not used for training or product improvement for other users. Source |
| Magica $12.50/month billed yearly ($14.99 month to month) | C | The privacy policy (updated October 6, 2026) states that Magica does not train or fine-tune AI models, so no opt-out is needed on its side. Your content is passed to providers including OpenAI, Anthropic, Google, OpenRouter, fal, Replicate, xAI and others, and the policy says their retention and training terms depend on the service and feature. No fixed retention periods are given, and Magica does not promise that providers delete content immediately. Source |
| Gamma €8/seat/month billed annually (Plus; the pricing page lists euros, no USD list published) | D | On individual plans (Free, Plus, Pro, Ultra) the default is 'Data is allowed to be used to train AI models'. You can opt out any time under Settings, Data Controls. Business and Enterprise workspaces are excluded automatically and the setting is locked. The terms grant Gamma a perpetual, irrevocable license to use your content for training whenever those privacy controls are not switched on. Source |
| Google Gemini $4.99/month (Google AI Plus, US price) | D | While Keep Activity is on, Google says it uses your activity to provide and improve its services, including training generative AI models, with help from human reviewers. Activity auto-deletes after 18 months by default, and you can change that. Chats seen by reviewers are disconnected from your account and kept for up to three years, even if you delete your activity. With Keep Activity off, or in temporary chats, conversations are held for 72 hours and are not used for training unless you send feedback. Source |
| HeyGen $24/month billed yearly (Creator; $29 month to month) | D | For paid plans, the Terms grant HeyGen a royalty-free, transferable, sublicensable, worldwide and irrevocable license to your content, including to train or improve its AI and machine-learning models. The privacy policy lists text, voice, scripts, images and videos as inputs used to train its models, on a legitimate-interest basis. There is no settings toggle: you opt out by contacting HeyGen's privacy team. Deletion requests are handled within 72 hours, with backups kept 60 days. Source |
| InVideo AI $4/seat/month billed yearly (Basic; $9 month to month) | D | The Terms and privacy policy both say InVideo may use your User Data to train, fine-tune or otherwise adjust its AI and machine-learning models. There is no settings opt-out: you stop it only by deleting the data or closing your account, and that does not undo earlier training. Avatar facial data and voice profiles are kept until you request deletion or delete the account; other content is deleted within 90 days of a deletion request. Source |
| Jasper $59/month per seat, billed yearly | D | The pricing page states that Jasper does not use your data to train AI models, and the trust page says your data is never used to train third-party LLMs. The terms say the opposite for Jasper's own models: Jasper uses inputs and outputs to develop and enhance its AI models, and you grant a perpetual, irrevocable license that covers improving AI models. Only EU and UK users are offered a way to object, by emailing legal@jasper.ai. Source |
| Otter.ai $8.33/user/month (Pro, billed annually) | D | The privacy policy (effective June 16, 2026) lists 'training our proprietary AI technology on de-identified audio recordings and on transcriptions (which may contain Personal Information)'. Otter's security page says de-identification is automatic, recordings are not manually reviewed without explicit customer consent, and customer data does not train its AI providers' models. No training opt-out is described. Retention is 'as long as necessary', with no fixed period. Source |
| Perplexity $17/month (Pro, billed yearly) | D | For Free, Pro and Max users, the AI Data Retention setting is on by default, so some of your data may be used to train AI models until you switch it off under Preferences. The opt-out only covers data collected after you opt out, and Perplexity says data already collected for training cannot be deleted. Enterprise data is never used for training, and Perplexity says its contracts bar third-party model providers from training on its data. Source |
| Undetectable AI $5/month (10K words, billed $60 yearly) | D | The privacy policy was last updated 03/18/2023. It covers account and usage data and says information is processed to 'provide, improve, and administer' the service, but it never says whether pasted text is stored, for how long, or whether it trains models. No training opt-out is described. The Ethics page separately says the company used 'the data we've collected over the years' to build deepfake-detection tools. Deletion is by email request. Source |
| Midjourney $8/month (Basic, billed yearly at $96) | E | Midjourney's public training summary answers 'Yes' to whether data from user interactions, such as prompts, was used to train its models, and we found no opt-out in the documents we read. The Terms of Service grant Midjourney a perpetual, worldwide, irrevocable license to your inputs and generated assets, and that license survives termination. Your creations are publicly viewable and remixable by default. Only Pro and Mega include Stealth mode, which hides creations on the website on a 'best efforts' basis. Source |
Before you pay
- Business and team plans often have stricter defaults than personal plans from the same vendor.
- An opt-out usually applies going forward, not to content already used.
- Uploading other people’s faces, voices or documents can require their consent, whatever the plan says.
Each finding comes from the vendor's own page, read on the date shown in the full review. Terms change: re-read them before you buy. How we grade.