GDPR-compliant AI: ChatGPT, Gemini, Copilot, EU cloud or local?

No AI tool is GDPR-compliant by itself; it depends on how you use it. You need a legal basis, a data processing agreement under Art. 28 GDPR, a lawful basis for third-country transfers and, if high risk is likely, an impact assessment. OpenAI, Microsoft and Google say their business plans don't train on your data by default; EU hosts and local models keep it in the EU or in-house.

Is ChatGPT GDPR-compliant?

ChatGPT can be part of a GDPR-compliant setup if you use a business plan and meet your own obligations under the GDPR. Personal accounts fed with customer or employee data cannot. The difference comes down to three things: training, contract and control.

Personal plans: Free, Plus and Pro

For its services for individuals, OpenAI says it may use your content to train its models. You can switch this off in ChatGPT under Settings > Data controls ("Improve the model for everyone") or in the Privacy Portal. Temporary chats are not used for training while they remain temporary. One catch: if you give thumbs-up or thumbs-down feedback, OpenAI says the entire conversation behind it may be used for training.

OpenAI names a Data Processing Addendum (DPA) explicitly for ChatGPT Business, ChatGPT Enterprise and the API. Without a contract under Art. 28 GDPR, a key element is missing whenever a provider processes personal data on your behalf.

Business, Enterprise and the API

For ChatGPT Business, Enterprise and Edu and for the API platform, OpenAI commits to not using business data for training by default; for the API this applies to data from March 1, 2023 onwards. Deleted conversations are removed from OpenAI's systems within 30 days unless the law requires retention. API inputs and outputs may be kept for up to 30 days to detect abuse; for eligible endpoints you can request zero data retention.

OpenAI announced data residency in Europe on February 5, 2025 for ChatGPT Enterprise, ChatGPT Edu and the API. According to OpenAI, API projects created with the Europe region are processed without data retention in the region; this can only be set for newly created projects.

Does ChatGPT share your data, and who can see the history?

In a Business workspace, admins can view, export and delete users' conversations; on Enterprise they get an audit log through the Compliance API. At OpenAI itself, access is limited to authorized staff (support, abuse investigations, legal obligations) and to specialized third-party providers bound by confidentiality, according to the vendor. On the technical side, OpenAI says it encrypts data at rest with AES-256 and in transit with TLS 1.2 or higher, and Business, Enterprise and the API have completed SOC 2 Type 2 audits. Still, tell your staff openly who can read their prompts – German regulators explicitly expect this transparency when prompt histories are stored.

So can you use ChatGPT without concern? Tasks that involve no personal data at all fall outside data protection law, according to the German authorities. But personal data creeps in quickly, and not just through names and addresses.

What does the GDPR require when you use AI?

The GDPR does not mention artificial intelligence explicitly. It applies to any processing of personal data, which includes prompts, uploaded documents and generated answers. Six obligations are almost always in play:

Obligation	Where in the GDPR	What it means for AI
Legal basis	Art. 6(1); for health data also Art. 9	Needed for processing in the AI system and, per the German authorities, for the transfer to the provider.
Data processing agreement (DPA)	Art. 28(3)	Cloud AI providers usually act on your behalf; the contract sets out subject matter, duration, purpose, types of data and obligations.
Third-country transfers	Art. 44 to 46	For the US: the adequacy decision for certified companies, or standard contractual clauses.
Data protection impact assessment	Art. 35	Required where high risk is likely; German regulators expect this to apply in many AI deployments.
Data minimisation, default settings	Art. 5(1)(c), Art. 25	Turn training and history off by default, enter only what is needed, pseudonymise where possible.
Accountability and records	Art. 5(2), Art. 30	Record which AI system processes which data for which purpose.

The guidance on AI and data protection published by Germany's Data Protection Conference (DSK) on May 6, 2024 turns this into practice: define purposes before use, always involve the data protection officer and consider involving the works council, give staff company devices and accounts, and issue clear internal rules with examples of permitted and prohibited uses. The DSK prefers technically closed systems and applications that do not train on inputs and outputs.

Important: this article explains the landscape; it is not legal advice. Whether a specific processing operation is lawful is something to settle with your data protection officer or a law firm.

Which AI tools are GDPR-compliant? Vendors compared

No vendor takes your obligations off your hands, but their commitments differ considerably. The table summarizes what each vendor states in its own documentation (accessed November 4, 2026). What counts in the end is the contract you actually sign.

Offering	Training on your inputs	DPA	Data location and notes
ChatGPT Free, Plus, Pro	Possible; opt out in data controls	Not named for these plans	Temporary chats not used for training
ChatGPT Business, Enterprise, API	No, by default	Yes	EU data residency for Enterprise, Edu and API (new projects)
Microsoft Copilot (Microsoft 365, business customers)	No, not for foundation models	Yes, Product Terms and Data Protection Addendum	EU Data Boundary; Anthropic models currently excluded
Gemini in Google Workspace	Not without your permission	Yes, Cloud Data Processing Addendum	Data region settings do not apply to Gemini Notebook
Gemini app with a personal account	Yes when "Keep Activity" is on; human review	Not named	Chats reviewed by humans kept for up to three years
Gemini API	No in the EEA, Switzerland and the UK	Yes, Data Processing Addendum	Time-limited logs for abuse detection; may be cached transiently in any country with Google facilities
Claude for Work, Anthropic API	No, by default	Yes, with standard contractual clauses	–
Mistral Vibe	Yes, with opt-out; Enterprise no by default	DPA available	EU by default; some features transfer data outside the EU
IONOS AI Model Hub	No; prompts not logged	Yes, part of the contract	Data centers in Germany, no sub-processors
STACKIT AI Model Serving	No; requests not stored	Check your contract	STACKIT data centers in Germany and Austria
Local with Ollama	No vendor access when run locally	Not needed if you run it yourself	Switch off cloud features

Is Gemini GDPR-compliant?

It depends on the product. In Google Workspace, prompts count as customer data under the Cloud Data Processing Addendum, and Google does not train models on them without your prior permission. In the Gemini app with a personal account, by contrast, Google also uses your activity for training when "Keep Activity" is on, a subset of chats is reviewed by humans, and Google itself advises against entering confidential information.

If you build the Gemini API into your own application: under the Gemini API terms, applications offered to users in the EEA, Switzerland or the UK may only use paid services, and there the data rules for paid services also apply to free quota.

Copilot and the GDPR

For Microsoft 365 business customers, Microsoft commits that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models; for EU customers, Copilot is an EU Data Boundary service. Check two things: Anthropic models, which Microsoft uses as a sub-processor, are currently excluded from the EU Data Boundary. And Copilot only surfaces content a user can at least view, so Microsoft stresses that permissions, for example in SharePoint, must be right – clean them up before rollout.

What about Mistral?

According to its help center, Mistral hosts data in the EU by default but, depending on the feature, may transfer it temporarily to sub-processors outside the EU, covered by standard contractual clauses. In its Vibe assistant (formerly Le Chat), users are not opted out of training by default; Enterprise customers are, and on Team plans admins can switch training off. An EU data location does not automatically mean "no training".

Can data still go to the US? The Data Privacy Framework

Yes, for now. On July 10, 2023, the European Commission adopted its adequacy decision for the EU-US Data Privacy Framework (DPF), so transfers to participating US companies rest on Art. 45 GDPR. For recipients without certification you need appropriate safeguards under Art. 46, such as standard contractual clauses.

That foundation is not stable. The EU General Court dismissed the Latombe action for annulment on September 3, 2025 (T-553/23); the appeal C-703/25 P has been pending before the Court of Justice since October 31, 2025. On June 29, 2026, noyb pointed to the US Supreme Court ruling on the independence of the FTC (Trump v. Slaughter), said that in its view the framework has collapsed, and announced a lawsuit. Formally, the decision stays in force until the Commission repeals it or the Court of Justice annuls it.

In practice: check whether your vendor participates in the DPF or relies on standard contractual clauses, use EU data residency where it is offered, and plan an exit route. If you build applications against an OpenAI-compatible interface, switching later is easier: IONOS and STACKIT offer one, and Ollama supports a subset of the OpenAI API.

Local AI or EU hosting: when is it worth it?

When data should not leave your premises or you want to avoid transfers to the US. Legal basis, transparency and, where needed, the impact assessment remain your job here too; both routes do reduce the number of recipients and transfers.

EU hosting: IONOS and STACKIT

With the IONOS AI Model Hub, inference runs exclusively in data centers in Germany, according to the documentation. Prompts and outputs are not logged, not stored and not used for training, no sub-processors are involved, and operational access happens only from within the EU. The DPA is part of the contract. IONOS also points out that it does not pseudonymise content: if you want to replace names, do it in your own application before the data leaves.

STACKIT AI Model Serving, the inference service of Schwarz Digits' cloud brand, says in its FAQ that it stores no customer data from requests and trains no models on it; STACKIT's seven data centers are in Germany and Austria. Both providers run open models, and the line-up changes: IONOS retired Llama 3.1 405B, Llama 3.1 8B and Mistral Nemo in September and October 2026. Plan against an interface, not a model name. For a wider look at EU clouds, see our article on sovereign cloud in 2026.

Self-hosting AI: Ollama, hardware, costs

Local AI means an open model runs on your own hardware, for example with Ollama. Ollama says it does not see your prompts when you run locally, and by default it only listens on 127.0.0.1, port 11434. But Ollama also offers cloud models and web search. If you want to stay strictly local, switch the cloud features off (OLLAMA_NO_CLOUD=1 or disable_ollama_cloud).

On hardware: Ollama supports NVIDIA GPUs with compute capability 5.0 or higher and AMD GPUs via ROCm, with further support via Vulkan. ollama ps shows whether a model is loaded fully into GPU memory, into system memory or split between the two. Costs cover purchase, power, operations and updates; compare them with the usage-based prices of the cloud instead of relying on a rule of thumb.

In data protection terms, if you run the model yourself for your own purposes you are usually the sole controller. Still check where the model comes from: according to the European Data Protection Board, a model developed with unlawfully processed personal data could affect the lawfulness of its deployment unless it has been duly anonymised. How to make your own documents usable for a model without handing them over is covered in RAG systems explained.

Special cases: health data and professional secrecy

Health data is a special category under Art. 9 GDPR; processing it is prohibited unless one of the exceptions in Art. 9(2) applies. If you operate in Germany, two national rules tighten the frame:

• Section 393 SGB V (Social Code, Book V): healthcare providers, statutory health and long-term care insurers and their processors may process social and health data in the cloud only in Germany, the EU, equivalent states or third countries with an adequacy decision. The processing entity needs an establishment in Germany and a current C5 attestation; since July 1, 2025 that generally means a Type 2 attestation.
• Section 203 StGB (Criminal Code): doctors, lawyers, tax advisors and other professionals bound by secrecy may disclose secrets to service providers only as far as needed for their work. They must ensure these providers are bound to confidentiality; if a provider that was not bound discloses a secret without authorization, the professional is liable to prosecution too.

Checklist: GDPR-compliant AI in ten steps

1. Define the purpose: what will the AI be used for, which data will it see, and are there tasks with no personal data at all?
2. Involve your data protection officer and works council before choosing tools.
3. Pick a plan that excludes training and switch training and history off by default.
4. Document the legal basis (Art. 6, plus Art. 9 for health data).
5. Sign the DPA and review the list of sub-processors.
6. Settle third-country transfers: DPF participation or standard contractual clauses, EU data residency where offered.
7. Assess the risk and carry out an impact assessment where high risk is likely.
8. Provide company accounts and an AI policy with examples of permitted and prohibited uses.
9. Minimise inputs: replace names and identifiers before sending; simply deleting names is often not enough.
10. Document, train, plan your exit: keep your records of processing up to date, brief your staff, and keep a switching path to an EU or local model ready.

From choice to integration

Whether a business plan, an EU host or a local model fits best is often decided at the interface: where does the data sit, who calls the model, what gets logged? That is where our AI integration work starts: we connect models via APIs to your CRM, ERP, ticketing system or document storage. The legal assessment stays with your data protection officer or a law firm.

Sources

1. Regulation (EU) 2016/679 (General Data Protection Regulation) – EUR-Lex, Official Journal of the EU L 119, May 4, 2016
2. Orientierungshilfe Künstliche Intelligenz und Datenschutz, Version 1.0 (in German) – Datenschutzkonferenz (DSK), May 6, 2024
3. Enterprise privacy at OpenAI – OpenAI, updated January 8, 2026
4. Data, Privacy, and Security for Microsoft Copilot – Microsoft Learn, accessed November 4, 2026
5. Generative AI in Google Workspace Privacy Hub – Google, accessed November 4, 2026
6. AI Model Hub: Data Handling – IONOS Cloud documentation, accessed November 4, 2026
7. Press release No 106/25, Case T-553/23, Latombe v Commission – Court of Justice of the European Union, September 3, 2025
8. US Supreme Court just blew up EU-US Data Transfers – noyb, June 29, 2026