Guides › DeepSeek, Grok, Mistral, Perplexity and Meta AI data policies

DeepSeek, Grok, Mistral, Perplexity and Meta AI: What They Do With Your Data

These five follow the same split as the four largest providers, with two exceptions. On enterprise plans and APIs, xAI, Mistral and Perplexity each give business customers a way to keep data out of training, and each documents a retention period or a zero data retention option. On their consumer plans, conversations are used for training unless the user opts out, and Mistral applies the same default to its Pro and Team plans. The exceptions are DeepSeek and Meta AI. DeepSeek's policy stores personal data in China and its API terms do not address training. Meta AI is a consumer product whose interactions are used to improve Meta's models, with no separate business terms on the pages reviewed. This page gives the position for each tier as of November 2026 and the checks to run before approving any of them.

The answer in four points
  • Enterprise plans and APIs. The xAI API, Mistral's Enterprise plan, and Perplexity's Enterprise plans and Sonar API state that customer data is not used for training. Mistral's API has a training opt-out in the admin panel.
  • Consumer and team plans. Grok, Perplexity and Mistral's Free, Pro and Team plans use conversations for training by default. Each has an opt-out, and the opt-out applies going forward.
  • DeepSeek. Personal data is processed and stored in the People's Republic of China, inputs are used for training unless the user opts out, and the API terms are silent on training and retention.
  • Meta AI. Interactions are used to develop and improve Meta's models. Whether you can object depends on where you live.

Provider by Provider

Each row paraphrases the provider's own policy or documentation as of November 2026. Where a document does not address a point, the table says so.

Provider and tierUsed for training?RetentionWorth knowing
DeepSeek app and web chatYes by default. The terms of use point to the Improve the model for everyone setting as the opt-out.Kept while the account exists. No fixed period is stated.The privacy policy states that personal data is processed and stored in the People's Republic of China.
DeepSeek APINot addressed in the API terms. The privacy policy lists model training among its purposes.Not stated.The privacy policy says it does not cover end users of applications built on the API.
Grok consumer appsYes by default, with an opt-out in data controls.Set by xAI's consumer terms.Grok inside X follows the X account's settings.
xAI API and enterpriseNo. xAI states it does not train on API requests or responses.Stored for 30 days for abuse auditing, then deleted. Zero data retention is available to enterprise teams.Each API response carries a header that reports whether zero data retention is active.
Mistral Vibe (formerly Le Chat), Free, Pro and StudentYes by default, with an opt-out in the account settings.Until deleted.Mistral's help center lists Pro the same as Free. The admin documentation page that listed Pro as excluded now redirects to it.
Mistral TeamYes by default. Administrators can turn training off for the whole organization.Admin-set chat retention.A business plan that trains by default. Turn the setting off before anyone uses it.
Mistral EnterpriseNo. Opted out by default, and an administrator controls any opt-in.Admin-set chat retention.Covered by Mistral's business terms.
Mistral APIOpt-out in the admin panel. Mistral does not state the default for paid API use.30 rolling days for abuse monitoring, unless zero data retention is activated.Zero data retention is granted on request. The training opt-out does not cover Labs and preview models.
Perplexity Free, Pro and MaxYes by default, through the AI data retention setting. Turning it off applies to future data.Until deleted.The default applies to anyone using a personal account for work.
Perplexity Enterprise Pro and Enterprise MaxNo.Uploaded files are kept for 7 days. Organizations with 50 or more Enterprise Pro seats, or one Enterprise Max seat, can set custom retention and force deletion.Perplexity states it has zero data retention and no-training agreements with its model providers for Enterprise data.
Perplexity Sonar APINo.Zero data retention policy.The API and the app are separate policy surfaces.
Meta AIYes. Meta states it uses interactions with its AI features to develop and improve its generative AI models.Not stated on the page reviewed.Private messages are not used unless someone shares them with the AI. An objection form exists, and its availability varies by region.

DeepSeek: Location Is the First Question

For most organizations the DeepSeek review ends at one sentence in its privacy policy: personal data is processed and stored in the People's Republic of China. A training opt-out does not change where the data sits or which legal system governs access to it. The API does not improve the picture, because the API terms leave training and retention unaddressed.

The models are a separate matter from the service. DeepSeek publishes open weights, so the same models can run in your own environment or through a cloud provider under that provider's terms, and no request goes to DeepSeek. The guides on sovereign AI deployment and cloud provider data policies cover both routes.

Consumer Plans Are the Exposure

Each enterprise tier above carries a no-training commitment, and each consumer tier above trains by default, as does Mistral's Team plan. The practical risk is an employee who signs in with a personal account because the tool is useful and nobody bought the business plan. Opt-outs help only for future data, and none of them reach content that has already been used. The guide on shadow AI covers how to find and close that path.

What to Check Before Approving Any of Them

  1. The tier and the contract.Confirm you are buying the business plan or API, and get the data processing agreement that goes with it.
  2. The training default.Read the current policy for your exact plan. Mistral's Team plan trains by default even though it is sold to businesses.
  3. Retention and zero data retention.Note the default period, who can shorten it, and whether zero data retention needs approval.
  4. Where data is processed.Ask for the processing and storage locations in writing if residency matters to your customers.
  5. The model providers behind the product.Perplexity routes requests to other providers' models. Confirm the terms that apply to them.
  6. Personal accounts.Decide the policy for consumer plans and tell people what it is.

Where Hardshell Sits

Hardshell limits what any provider receives. It connects enterprise data to the assistants and model APIs a company has approved without migrating it, enforces each user's existing permissions at query time, substitutes stand-ins for real identifiers before content reaches the model, and records every retrieval. A provider's training and retention terms then apply to stand-ins in place of customer, patient or account identifiers. See how Hardshell works.

Frequently Asked Questions

Is DeepSeek safe for company data?

Not through DeepSeek's own app or API for most organizations. Its privacy policy states that personal data is processed and stored in China and that inputs are used for model training unless the user opts out, and its API terms do not address training or retention. Running the open-weight models in your own environment is a different case.

Does Grok train on your data?

On the consumer apps, yes by default, with an opt-out. On the xAI API, no. xAI states that API requests and responses are stored for 30 days for abuse auditing, are not used for training, and are then deleted.

Does Mistral train on your data?

On the Free, Pro and Team plans, yes by default, with an opt-out. Team administrators can turn training off for the whole organization. Enterprise is opted out by default. On the API, Mistral provides an opt-out in the admin panel, and the opt-out does not cover Labs and preview models.

Does Perplexity train on your data?

On Free, Pro and Max, yes by default through the AI data retention setting. On Enterprise Pro and Enterprise Max, Perplexity states that data is never used for AI training, and the Sonar API runs under a zero data retention policy.

Does Meta AI train on your conversations?

Yes. Meta states that it uses interactions with its AI features to develop and improve its generative AI models. Meta offers an objection form through its Privacy Center, and its availability depends on the region.

Which of these providers offer zero data retention?

xAI offers it to enterprise API teams, Mistral grants it on request for its API, and Perplexity's Sonar API operates under a zero data retention policy. DeepSeek and Meta AI do not document one.