Is Claude GDPR compliant? What Anthropic does with your data

In short: Claude can be used in line with UK GDPR, but not on every plan and not with every kind of data. Free, Pro and Max are consumer plans without a data processing agreement, and Anthropic uses chats there for training while the model improvement setting is on. Team, Enterprise and the API run under the commercial terms, with a data processing addendum and no training. In every case your data goes to a US-based company.

The most common mistake: a paid Pro or Max subscription is not a business plan. According to Anthropic, the consumer terms cover Free, Pro and Max alike, including Claude Code used with those accounts. If staff paste client data into Claude on a personal Pro account, your business has no processing agreement for it.

According to its privacy policy, the controller for users in the UK, the EEA and Switzerland is Anthropic Ireland, Limited in Dublin. For transfers to affiliates and service providers in countries without an adequacy decision, Anthropic relies on standard contractual clauses. Its data processing addendum for business customers includes the UK International Data Transfer Addendum.

The rules differ sharply between consumer and business plans. This overview summarises Anthropic’s Privacy Center:

On every plan, if safety systems flag a chat as breaching the usage policy, Anthropic keeps inputs and outputs for up to two years and the classification scores for up to seven. Thumbs up or down feedback keeps the conversation for five years. According to the privacy policy, flagged chats and reported material are used for model improvement even if you have switched training off. Anthropic names its legitimate interests as the legal basis for training. Incognito chats are never used for training, Anthropic says.

Where the servers are: via the API you can only choose whether inference runs solely in the US or in any region (“global”), according to Anthropic’s documentation. The only storage location currently available is “us”. Anthropic does not offer UK-only or EU-only processing directly. If you use Claude through Amazon Bedrock or Google Cloud, you choose the region there, and the cloud provider is the processor.

New since 9 June 2026: for its most capable models, which Anthropic designates “Covered Models”, prompts and outputs are kept for 30 days on every platform, including for organisations that otherwise use zero data retention. Anthropic’s reason: spotting patterns of misuse. Consumer plans are not affected.

We found no fine or completed investigation by the ICO or a European data protection authority against Anthropic (as of September 2026). These points are on record:

1. 28 Aug 2025
   Training on consumer chats, five-year retention. Anthropic announced it would use chats and coding sessions from Free, Pro and Max for training unless users opted out, and keep that data for five years instead of 30 days. Existing users had until 8 October 2025 to choose. According to TechCrunch, the pop-up showed a large “Accept” button with a much smaller training toggle below it, set to “On”. Sources: Anthropic, TechCrunch.
2. Ongoing
   Risk when creating files. In its help article on creating and editing files, Anthropic itself warns that with network access switched on, Claude could be tricked through prompt injection, hidden instructions in files or web pages, into sending data to external servers. Source: Claude Help Center.
3. 31 Mar 2026
   Claude Code source code exposed. A packaging error in a public npm release made the source code of the Claude Code tool readable. According to Anthropic, no customer data or credentials were exposed. Source: heise online.
4. 8 Jul 2026
   ID checks for some users. Since the privacy policy of 8 July 2026, Anthropic may ask for photo ID in certain cases, handled by the provider Persona. According to heise, Anthropic now stores this identity data rather than deleting it straight away; mainly accounts suspected of fraud are affected. Sources: Anthropic, heise online.
5. 30 Jul 2026
   Models broke into third-party systems during tests. Anthropic reported that, because of a misconfiguration, Claude models had internet access during security evaluations and gained access to systems of three organisations, including credentials and several hundred rows of production data. Those affected were notified on 27 July 2026. Source: Anthropic.

The pattern applies to every cloud service, not just Anthropic: terms for the same app change, retention periods get longer, and security depends on features you do not control. For OpenAI, see our guide Is ChatGPT GDPR compliant?

As soon as staff enter client, patient or employee data, your business is the controller under UK GDPR. That means:

• A processing agreement (Article 28 UK GDPR): only on Team, Enterprise or the API. Anthropic’s DPA becomes part of the commercial terms automatically and commits Anthropic to report security breaches without undue delay, and in any event within 48 hours.
• International transfers: data goes to the US. Document the transfer mechanism and, where you rely on the UK Addendum to the standard contractual clauses, a transfer risk assessment. The ICO’s guidance on international transfers explains the steps.
• A data protection impact assessment: likely needed once you process special category data such as health data, or large volumes of client data. The ICO’s guidance on AI and data protection covers what regulators expect.
• A staff AI policy: which plan, which accounts, which data must never be entered, which settings are mandatory. Ban personal Free, Pro and Max accounts for business data explicitly.
• Professional confidentiality: solicitors, accountants and medical practices owe duties of confidentiality that go beyond data protection law. A consumer account does not meet them; check business plans with your regulator’s guidance and your data protection officer.
• EU customers: if you offer AI systems or their output in the EU, the EU AI Act may apply, including the AI literacy duty. See our guide to the EU AI Act.

Anthropic publishes its list of sub-processors at anthropic.com/subprocessors.

To switch off training on a Free, Pro or Max account, according to Anthropic’s help:

1. Open claude.ai and click your name in the bottom left.
2. Select Settings.
3. Open Privacy, or go straight to claude.ai/settings/data-privacy-controls.
4. Switch off the model improvement toggle. Anthropic has called it “Help improve Claude”; the current help calls it “Help improve our AI models”.
5. In the mobile app: your name, Settings, Privacy, same toggle.

Note: switching it off only affects future chats. According to Anthropic, your data stays in training runs that have already started and in models already trained. For one-off confidential questions, use an incognito chat, which is not used for training. Delete individual chats in the sidebar; Anthropic says they disappear from your history at once and from its systems within 30 days.

On Team and Enterprise, your organisation sets the rules. Enterprise admins can configure custom retention periods.

Every agreement, toggle and retention period on this page has one cause: your input leaves your computer and ends up with Anthropic in the US. A local assistant turns that around. Speech recognition and the language model are files on your PC and run there.

• No recipient: dictations, emails, documents and recordings go to no provider, so there is no processing agreement and no international transfer for that content.
• No training, no reviewers: there is no toggle anyone can forget and no retention period at a third party.
• No terms that change overnight: what the software does does not depend on a new version of someone’s privacy policy.
• Confidential matters stay in-house: client, patient and customer data need no redacting, because they are not transferred anywhere.

SpeechToWork is such an assistant for Windows: dictate, write emails, ask questions about documents, take meeting minutes, transcribe audio and read receipts, all on your own computer. It has limits: the local language model is smaller than the big cloud models, knows less about the world and does not search the web. It runs on Windows 10 and 11 only and needs a reasonably recent PC, with 16 GB of memory recommended. For office work where all the information is already in the text, it is enough.

More on our pages about GDPR-compliant AI and local AI.

• Anthropic: Updates to Consumer Terms and Privacy Policy, 28 August 2025
• TechCrunch: Anthropic users face a new choice, 28 August 2025
• Anthropic: Privacy Policy
• Anthropic Privacy Center: Updates to our Privacy Policy
• Anthropic Privacy Center: How long do you store my data? (updated 1 July 2026)
• Anthropic Privacy Center: How long do you store my organization’s data?
• Anthropic Privacy Center: Is my data used for model training?
• Anthropic Privacy Center: How do I change my model improvement privacy settings?
• Anthropic Privacy Center: Data Processing Addendum (DPA)
• Anthropic: Data Processing Addendum, effective 24 February 2025
• Claude docs: Data residency
• Claude Help Center: Data retention practices for Covered Models (from 9 June 2026)
• Claude Help Center: Create and edit files with Claude
• heise online: Anthropic may require an ID for using Claude in the future, 25 June 2026
• heise online: Claude Code unintentionally open source, 31 March 2026
• Anthropic: Investigating incidents in cybersecurity evaluations, 30 July 2026
• ICO: Guidance on AI and data protection
• ICO: International transfers

Last reviewed: September 2026. We checked every statement against the sources listed in September 2026. Providers change plans, settings and terms often; the provider’s current version always prevails. This guide is not legal advice.