# Grok privacy and GDPR: training on your data and investigations in Europe

> Grok privacy and GDPR: who the controller is, how to stop training on your chats, where the data goes and which EU investigations are open (November 2026).

Source: Giallo Studio (giallostudio.it) · Page: https://giallostudio.it/en/blog/grok-privacy-gdpr · Published: 2026-11-08 · Updated: 2026-11-08

**Grok can be used in line with the GDPR, but not with its default privacy settings: on the consumer services, SpaceXSI can train its models on European users' conversations by invoking legitimate interest, and the opt-out has to be switched on manually. An Italian business needs business accounts or the API, an assessment of the data transfer to the United States and an internal policy on what can be typed in.** The data controller is a US company, transfers to vendors rely on the Data Privacy Framework or standard contractual clauses, and the Irish DPC and the European Commission have open investigations into the product. Everything below comes from the [x.ai privacy policy](https://x.ai/legal/privacy-policy), the [European addendum](https://x.ai/legal/europe-privacy-policy-addendum), the [legal FAQ](https://x.ai/legal/faq) and the developer documentation, all read on November 8, 2026.

So the question “Is Grok safe?” has two different answers. For an individual writing about travel or recipes, it is mostly a matter of personal preference. For a business pasting in a contract, a customer list or the payroll file, it becomes processing of personal data with a non-EU vendor, and has to be handled as such.

<div class="key-points">
<p class="kp-title">In summary</p>
<ul>
<li>Controller for European users: SpaceX Super Intelligence LLC (USA), with an EU representative in Dublin. Inside X, X's privacy policy and terms apply instead.</li>
<li>Training on consumer chats: on until you turn off “Improve the model”; stated legal basis, legitimate interest.</li>
<li>Private Chat: no history, no training, deletion within 30 days unless legal or security obligations apply.</li>
<li>Business and enterprise customers: SpaceXSI states it does not use their content to improve its models.</li>
<li>Open investigations: Irish DPC (2025 and 2026), European Commission under the Digital Services Act (2026), Italian Garante merged into the Irish procedure.</li>
</ul>
</div>

## Who processes the data when an Italian business uses Grok?

The party responsible for the data is not always the same: it depends on where you use Grok. On the consumer services (grok.com and the iOS and Android apps), the controller for European users is SpaceX Super Intelligence LLC, the US-based company that now develops Grok under the name SpaceXSI and that the privacy policy itself describes as separate from X Corp. The EU representative, required by [Article 27 of the GDPR](https://eur-lex.europa.eu/legal-content/IT/TXT/?uri=CELEX:32016R0679) for non-EU controllers, is Lionheart Squared (Europe) Ltd, in Dublin. Sources: [privacy policy](https://x.ai/legal/privacy-policy) and [European addendum](https://x.ai/legal/europe-privacy-policy-addendum).

If instead you query Grok inside X, the same privacy policy makes clear that this use is governed by X's privacy policy and terms, not by SpaceXSI's. The [Grok FAQ](https://docs.x.ai/grok/faq) adds that SpaceXSI provides the service inside X but has no operational oversight of the platform: for a problem on X, you go to X.

For anyone who keeps a record of processing activities, the distinction matters: the same employee who uses Grok on the website and then on X is dealing with two different controllers, with two different privacy notices. The overall picture of roles and responsibilities is in our guide [AI and GDPR](https://giallostudio.it/en/blog/ai-and-gdpr).

## Does Grok use conversations to train its models?

Yes, on the consumer services, unless you opt out. The [legal FAQ](https://x.ai/legal/faq) explains that SpaceXSI may use users' prompts and responses to train its models; for European users, the [addendum](https://x.ai/legal/europe-privacy-policy-addendum) names legitimate interest as the legal basis, not consent. This means users don't have to say “yes” for training to start: they have to say “no” for it to stop.

Three details are often missed:

- **The opt-out applies going forward.** Once the option is off, the FAQ specifies that it is new conversations that stop being used: so it's best to do it before you start working with Grok, not after.
- **Feedback is an exception.** Even with the opt-out on, feedback you give voluntarily (for example, rating a response) can be used for training.
- **Private Chat is a different thing.** In this mode, conversations stay out of your history, are not used for training and are deleted within 30 days, unless legal or security obligations apply.

The issue of work data inside public assistants goes beyond Grok: we cover it in [company data and AI: the privacy risks](https://giallostudio.it/en/blog/ai-privacy-for-business).

## How do you turn off training on Grok?

The procedure given in SpaceXSI's legal FAQ is short. Do it on every device and every company account, and keep a record of it.

<ol class="steps">
<li><strong>iOS or Android app.</strong> Settings → Data Controls → turn off “Improve the model”.</li>
<li><strong>grok.com website.</strong> Settings → Data → same option.</li>
<li><strong>For sensitive requests.</strong> Use Private Chat, knowing that conversations are still retained for up to 30 days.</li>
<li><strong>Feedback.</strong> Ask staff not to rate responses that contain customer or colleague data.</li>
<li><strong>Shared links.</strong> Don't publish links to conversations: anyone with the link can read them and, if the link is published, they can be indexed by search engines.</li>
</ol>

The last point comes from the same [legal FAQ](https://x.ai/legal/faq) and is the most underrated: a link pasted in a forum or a public document makes the conversation visible to anyone who finds it.

## Does Grok data leave the European Union?

Largely yes, and it is stated openly. The [European addendum](https://x.ai/legal/europe-privacy-policy-addendum) says transfers to vendors take place through the EU-US Data Privacy Framework or standard contractual clauses, the two most common mechanisms for transfers to the United States: the adequacy decision on the Data Privacy Framework (Article 45 of the [GDPR](https://eur-lex.europa.eu/legal-content/IT/TXT/?uri=CELEX:32016R0679)) and standard contractual clauses (Article 46).

On the API side, the documentation is even more explicit. The [regional endpoints page](https://docs.x.ai/developers/advanced-api-usage/regions) explains that `https://api.x.ai` does not guarantee the region where the request is processed; there is a dedicated US endpoint, priced 10% higher, but the documentation lists no EU endpoint. The release notes report that Grok 4.5 was made available in the API console for EU users in July 2026: model availability, not processing in Europe.

For a business, this means one concrete thing: the transfer assessment and the privacy notice to your own customers must state that data may be processed in the United States. We explain how to approach this in [GDPR and AI for businesses](https://giallostudio.it/en/blog/ai-and-gdpr).

## What changes between consumer chat, business accounts, the API and Grok Bot?

Data rules change with the product. This table sums up what SpaceXSI states on the pages we read on November 8, 2026.

| Product | Training on content | Retention | Source |
| --- | --- | --- | --- |
| Grok consumer chat | Yes, unless you opt out in Data Controls | Private Chat deleted within 30 days | [legal FAQ](https://x.ai/legal/faq) |
| Grok for business and enterprise customers | Stated “no” | Enterprise: custom retention policies | [legal FAQ](https://x.ai/legal/faq), [user guide](https://docs.x.ai/grok/user-guide) |
| API | No training without explicit permission | 30 days, encrypted, for abuse monitoring; then deleted | [Security FAQ](https://docs.x.ai/developers/faq/security) |
| API with Zero Data Retention | No training | No retention; you lose features such as Files, Batch API and stateful Responses API | [Security FAQ](https://docs.x.ai/developers/faq/security) |
| Grok Bot | Follows the privacy settings of the Cursor account | Cloud storage required; Legacy Privacy Mode not supported | [Grok Bot FAQ](https://docs.x.ai/grok-bot/faq) |

Two notes for developers. Zero Data Retention is enabled at team level, but the documentation itself advises most customers against it precisely because it turns off several features. The `safety_identifier` field is used to attribute any violations to the individual end user rather than to the whole API key: the documentation asks for an opaque, hashed identifier, never an email or a phone number. It's a good example of data minimization worth replicating with other vendors too.

On compliance, the vendor's own sources don't match. The [business page](https://x.ai/grok/business) lists SOC 2 Type II, HIPAA, GDPR and CCPA; the [security FAQ](https://docs.x.ai/developers/faq/security), asked “Is SpaceXSI GDPR and SOC II compliant?”, answers by citing only SOC 2 Type 2 compliance and points to a Trust Center accessible under NDA. We read them for what they are: vendor statements, to be checked against the contract documents.

## Which European investigations involve Grok?

Grok is one of the AI tools most closely watched by European authorities. These are the documented milestones.

| Date | Authority | Subject | Source |
| --- | --- | --- | --- |
| September 4, 2024 | Irish DPC | Proceedings closed after X committed not to use EU users' personal data to train Grok (press report) | [The Record](https://therecord.media/europe-data-privacy-watchdog-case-closed) |
| April 11, 2025 | Irish DPC | Inquiry into X Internet Unlimited Company over the processing of EU/EEA public posts used to train Grok: lawfulness and transparency | [DPC](https://dataprotection.ie/en/news-media/latest-news/data-protection-commission-announces-commencement-inquiry-x-internet-unlimited-company-xiuc) |
| January 2, 2026 | Garante privacy | Investigation into Grok, merged into the Irish procedure because X is based in Ireland | [Garante](https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10306949) |
| January 8, 2026 | Garante privacy | Warning to anyone using services like Grok to generate content from real third parties' images or voices without consent | [Garante](https://www.garanteprivacy.it/home/docweb/-/docweb-display/docweb/10207147) |
| January 26, 2026 | European Commission | Formal proceedings under the Digital Services Act against X over risks linked to Grok and its recommender system | [Commission](https://digital-strategy.ec.europa.eu/en/news/commission-investigates-grok-and-xs-recommender-systems-under-digital-services-act) |
| February 17, 2026 | Irish DPC | Inquiry into non-consensual intimate or sexualized images generated with Grok on X, including images of minors; GDPR Articles 5, 6, 25 and 35 | [DPC](https://dataprotection.ie/en/news-media/press-releases/data-protection-commission-opens-investigation-x-xiuc) |

According to the news item published by the Garante on October 8, 2026, the Irish authority had already confirmed on January 8, 2026 that X cooperated until the “nudification” feature was disabled. As of November 2026 we have found no final decisions on the DPC inquiries or on the Commission proceedings: anyone writing that Grok “has been fined” or “has been cleared” is anticipating an outcome that has not been published.

An open investigation does not ban the use of a tool. It does weigh on a business's risk assessment, though, especially if Grok handles customer data or images of people.

## Is Grok safe for a business that processes personal data?

It depends on which Grok and which data. A consumer account with default settings is not suitable for processing customer, employee or patient data: training is on, the controller is outside the EU and, according to the privacy policy, SpaceXSI acts as controller on the consumer services, so there is no data processing agreement with the business. A business account or the API changes the picture, because SpaceXSI states that it does not train on the content and offers retention controls; contract, transfers and security measures still need to be checked.

Then there is the AI Act. Since November 7, 2026, responses from Grok's text models via the API include an invisible text watermark, on by default, which the [documentation](https://docs.x.ai/developers/advanced-api-usage/watermarking) ties to AI Act compliance; admins can turn it off per model. Anyone publishing generated text should know this before turning it off. The transparency obligations for businesses using AI are summarized in our [guide to the AI Act for businesses](https://giallostudio.it/en/blog/ai-act-business-guide).

Finally, age: Grok is not intended for children under 13, and between 13 and 17 a parent's permission is required, according to the [legal FAQ](https://x.ai/legal/faq). For schools, gyms or youth associations, it's a limit worth knowing.

## What should an internal Grok policy cover?

Before banning or allowing it, write a few clear rules. The ones we suggest to clients when a team already uses Grok are these:

<ul class="checklist">
<li>Use allowed only with company accounts, with “Improve the model” turned off and verified.</li>
<li>No personal data of customers, employees or suppliers in consumer chat; sensitive data goes only into tools covered by a data processing agreement.</li>
<li>No uploading images or voices of real people without documented consent.</li>
<li>No sharing links published outside the company.</li>
<li>Grok inside X treated as an X service, with its own privacy notice.</li>
<li>Record of processing activities updated with controller, purpose and transfer to the United States.</li>
</ul>

A complete template, adaptable to other assistants too, is in [corporate policy for the use of AI](https://giallostudio.it/en/blog/company-ai-use-policy). For the broader picture on confidentiality risks, see [AI and corporate privacy](https://giallostudio.it/en/blog/ai-privacy-for-business).

## How we handle AI agents and the GDPR

When a process touches personal data, our starting point is not the model but the processing: what data is really needed, where it is processed, how long it is kept, who can see it. Only then do we choose the model, among Claude, GPT and Gemini, looking first at the contract, the processing region and retention; and as Refreesh we sign the DPA with the client.

If someone on your team already uses Grok or another public assistant and you want to understand what can stay and what should move to a controlled agent, we start with an analysis through our [AI consulting](https://giallostudio.it/en/ai-consulting). If the process is already clear, you can see how we build [custom AI agents](https://giallostudio.it/en/ai-agents) or [write to us directly](https://giallostudio.it/en/contact).

## Frequently asked questions

### Who is the data controller when I use Grok from Italy?

For Grok's consumer services used by people in Europe, the controller is SpaceX Super Intelligence LLC, a company based in the United States, which has appointed Lionheart Squared (Europe) Ltd, in Dublin, as its EU representative. If you use Grok inside X, X's privacy policy and terms apply instead. Source: privacy policy and European addendum on x.ai, read on November 8, 2026.

### Does Grok use my conversations to train its models?

On the consumer services, yes, unless you turn it off: SpaceXSI may use prompts and responses for training and, for European users, cites legitimate interest as the legal basis. You switch it off under Data Controls → “Improve the model” in the app or Settings → Data on grok.com; from then on, new conversations are no longer used.

### What is Grok's Private Chat?

It's a mode in which conversations stay out of your history, are not used to train models and are deleted within 30 days, unless legal or security obligations apply. It's useful for occasional questions, but it doesn't replace a company rule on what data may be entered.

### Is Grok GDPR compliant?

Grok's business page lists GDPR and CCPA among its compliance claims, while the developer security FAQ, asked about GDPR and SOC 2, cites only SOC 2 Type 2 compliance and points to a Trust Center under NDA. These are vendor statements: the assessment remains the responsibility of whoever uses the tool on their own data.

### Are there European investigations into Grok?

Yes. The Irish DPC opened an inquiry on April 11, 2025 into the use of European users' public posts to train Grok, and another on February 17, 2026 into non-consensual sexualized images; the European Commission opened DSA proceedings on January 26, 2026, and the Italian Garante opened an investigation on January 2, 2026, later merged into the Irish one. As of November 2026 we have found no final decisions.

### Does Grok Bot have different privacy rules from Grok chat?

Yes. Grok Bot uses the authentication and data settings of your Cursor account: the training opt-out follows Cursor's privacy settings, and Legacy Privacy Mode is not supported because the service needs cloud storage. Source: Grok Bot FAQ, read on November 8, 2026.
