Is ChatGPT GDPR Compliant? A 2026 Guide for European Teams
Team Sonomos · Sonomos
The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.
Short answer: ChatGPT is not "GDPR compliant" by default for European teams. Compliance is a posture you assemble, not a checkbox a vendor ticks. To use ChatGPT lawfully under the GDPR in 2026, an EU controller has to (1) sign OpenAI Ireland Limited's Data Processing Addendum, (2) operate on a tier that excludes prompts from training and supports EU data residency where applicable, (3) identify a lawful basis under Article 6 (and Article 9 for special-category data), (4) run a DPIA when the processing is likely to result in high risk, (5) implement Article 32 technical measures including pseudonymization at the source, and (6) document the lot in the Article 30 record. None of that is automatic on ChatGPT Free, Plus or Pro. This guide walks through, in plain English, what each ChatGPT tier means for GDPR in 2026 — and the controls that hold up under a supervisory-authority audit.
What the GDPR actually requires of an AI tool
The GDPR doesn't have an "AI annex." It applies to ChatGPT the same way it applies to any other processor: through six obligations that all attach the moment a prompt containing personal data leaves the user's device.
- Lawful basis (Article 6). Each processing activity needs a basis: contract, legal obligation, vital interest, public task, legitimate interest, or consent. For "draft an email mentioning a customer," legitimate interest is usually the operative basis — but only if the balancing test is documented and the data subject's expectations support it.
- Special-category data (Article 9). Health, racial or ethnic origin, political opinion, religious belief, biometric and genetic data, sex life or orientation, and trade-union membership. Sending any of these to ChatGPT requires explicit consent or one of the narrow Article 9(2) exemptions. There is no "legitimate interest" path for special-category data.
- Transparency (Articles 13–14). Data subjects must be told that their personal data is being processed by an AI tool, the categories involved, the recipients (OpenAI Ireland Limited, plus any sub-processors), the retention period, and their rights.
- Data subject rights (Articles 15–22). Access, rectification, erasure, restriction, portability, objection, and the right not to be subject to solely-automated decisions with legal or similarly significant effects. ChatGPT outputs that make such decisions trigger Article 22.
- International transfers (Chapter V). OpenAI's EU customers contract with OpenAI Ireland Limited (OpenAI's single EU establishment since 15 February 2024, per the EDPB ChatGPT Taskforce report), but processing still occurs in part on US infrastructure, so a transfer mechanism — Standard Contractual Clauses with supplementary measures, or the Data Privacy Framework where the importer participates — is needed. The Schrems II analysis applies.
- Security of processing (Article 32). Pseudonymization, encryption, ongoing confidentiality and integrity, and the ability to restore availability after an incident. Article 32(1)(a) calls out pseudonymization by name — which is what masking identifiers before a prompt leaves the device implements.
A vendor cannot make a controller "GDPR compliant." A vendor can only avoid being the thing that breaks compliance. Read every "GDPR-compliant ChatGPT" claim in that light.
ChatGPT tier-by-tier GDPR posture (April 2026)
OpenAI's tier documentation could not be re-read from primary pages at the November 2026 review; confirm each row against OpenAI's current help pages before relying on it.
| Tier | DPA available? | Default training | EU data residency option | Sub-processors disclosed | Contractual data-processing terms for special-category data |
|---|---|---|---|---|---|
| ChatGPT Free | No | Yes (opt-out in Data Controls) | No | Limited | No |
| ChatGPT Plus | No | Yes (opt-out) | No | Limited | No |
| ChatGPT Pro (consumer) | No | Yes (opt-out) | No | Limited | No |
| ChatGPT Business (formerly Team) | Yes — business plan, DPA available | No | Limited; varies by SKU | Yes | Limited |
| ChatGPT Enterprise / Edu | Yes — DPA + EU data residency for eligible customers | No | Yes (for stored content; inference compute may transfer) | Yes | Yes (with appropriate contractual terms) |
| OpenAI API (default) | Yes — DPA | No | EU residency available on eligible plans | Yes | Yes |
| OpenAI API + Zero Data Retention | Yes | No | Yes | Yes | Yes (strongest posture) |
A few things to notice:
- Free, Plus and Pro are not covered by a DPA. That alone makes them unsuitable for processing personal data of EU data subjects on behalf of a controller. They are a B2C contract between the individual user and OpenAI; the user is the controller of their own personal data, not the employer.
- The training default is the loudest GDPR signal in the table. Article 5(1)(b) ("purpose limitation") and Article 5(1)(c) ("data minimisation") are difficult to reconcile with prompts being used to train future models. OpenAI's enterprise tiers exclude training by default, which removes that conflict.
- EU data residency does not eliminate transfers. Even on Enterprise with EU data residency, OpenAI is a US-headquartered company, parts of the model lifecycle run in the US, and US government access requests under FISA 702 remain a Schrems II concern. The mitigation is a combination of SCCs, supplementary measures, and — most directly — minimising the personal data that ever leaves the device.
The Italian Garante decision and what happened to it
The Garante per la protezione dei dati personali was the first EU supervisory authority to act against ChatGPT. On 31 March 2023 it imposed a temporary limitation on OpenAI's processing of Italian users' data after the data breach of 20 March 2023. On 20 December 2024 it announced the close of its investigation: decision No. 755 of 2 November 2024 fined OpenAI €15 million and ordered a six-month information campaign. The Garante's findings, simplified:
- OpenAI had not identified a lawful basis under Article 6 for the personal data it processed to train ChatGPT.
- Transparency to data subjects under Articles 5(1)(a), 12 and 13 was insufficient.
- Age-verification controls were inadequate.
- The 20 March 2023 breach had not been notified under Article 33.
On 18 March 2026 the Tribunale di Roma (R.G. 4785/2025; the Garante cites the judgment as No. 4153/2026) annulled decision No. 755 in its entirety on competence grounds: OpenAI Ireland Limited became OpenAI's EU main establishment on 15 February 2024, so under the GDPR's one-stop-shop mechanism the Irish Data Protection Commission was the lead supervisory authority and the Garante could not continue the proceeding alone. The court did not rule on the substance of the findings. The Garante's press-release page now notes that the decision was removed after the judgment.
What survives the annulment is the premise, not the penalty. The EDPB's ChatGPT Taskforce report (23 May 2024) and Opinion 28/2024 on AI models (17 December 2024) both treat model training as processing that needs its own lawful basis, and both reinforce that controllers using ChatGPT must identify their own lawful basis for their own processing, separately from whatever basis OpenAI relies on for training.
For a 2026 European deployment, the practical implication is that controllers must document their own basis for every category of prompt — not rely on OpenAI's broader posture, and not read the annulment as a ruling on the merits.
What goes wrong in real European deployments
Across the deployments we've seen and the supervisory-authority decisions that have leaked into the public record, four patterns recur:
-
Personal data in prompts on Free, Plus or Pro. The single most common breach pattern. An employee pastes a customer email into ChatGPT to translate or rewrite it. The customer's personal data is now processed by OpenAI under consumer terms with no DPA, with training enabled by default, with no documented lawful basis. This is an Article 6, Article 28, and (if special-category) Article 9 breach in a single prompt.
-
Special-category data via consumer Plus. A clinician summarising a patient note. A HR manager analysing an absence pattern that reveals a disability. A nonprofit logging refugee case files. Article 9 explicit consent has not been obtained — and even if it had, Plus does not have the contractual measures to support special-category processing.
-
Automated decisioning without an Article 22 carve-out. Using ChatGPT to triage CVs, score loan applications, or generate disciplinary recommendations and acting on the output without meaningful human review. The data subject has rights under Article 22 that the workflow has not honoured. The EU AI Act overlay compounds this: following Regulation (EU) 2026/1744, Annex III high-risk obligations for employment and creditworthiness systems apply from 2 December 2027.
-
No DPIA where one was required. Article 35 requires a Data Protection Impact Assessment whenever processing is likely to result in a high risk to data subjects' rights and freedoms. The ICO says AI processing of personal data triggers a DPIA in the vast majority of cases, and the CNIL's AI how-to sheets include a sheet on carrying out one.
The control layers that hold up under audit
A defensible 2026 deployment is a stack, not a single product:
Contract layer
- OpenAI Ireland Limited DPA signed for every EU controller relationship — Business (formerly Team), Enterprise, Edu, or API.
- Standard Contractual Clauses for the international transfer to US infrastructure, with the supplementary measures specified in EDPB Recommendations 01/2020.
- Sub-processor list reviewed against the controller's own register and updated when OpenAI publishes changes.
- EU data residency elected on Enterprise / API where the workload supports it.
Tier-and-settings layer
- No personal data on Free, Plus or Pro. Period. This is a policy, not a guideline.
- Training disabled on every tier, by configuration where required.
- Retention configured as low as the workload allows — Zero Data Retention on the API for sensitive workloads.
- Memory features off unless their use is documented in the Article 30 register and surfaced in transparency notices.
Data-minimisation layer (the one that prevents the breach)
This is the layer where Locke sits. Articles 5(1)(c) and 32(1)(a) push controllers toward processing the minimum personal data necessary and toward pseudonymisation. Locke runs on the user's device as a desktop app; its companion browser extension screens AI websites, and a local proxy screens command-line, script and SDK traffic. Locke 1.0 detects personal data and replaces it with one-way placeholders before the request leaves the device; reversible tokenization is planned shortly after 1.0. Coverage depends on the app, platform and settings.
- The prompt OpenAI receives carries placeholders in place of the detected identifiers. Masking removes identifiers before transfer, which is data minimisation (Art. 5(1)(c)) and a security measure (Art. 32(1)(a)); it does not by itself take the processing outside the GDPR. Recital 26 treats pseudonymised data as personal data, so Chapter V still applies to whatever personal data remains in the prompt, with materially reduced risk.
- The user still gets a useful response, because the surrounding context of the prompt is preserved.
- Detection happens on the device. The content is processed locally and Sonomos does not receive it; whether a processor relationship arises is for the controller's assessment.
EDPB Recommendations 01/2020, Use Case 2, treat pseudonymisation before transfer as an effective supplementary measure where the data can no longer be attributed to a specific person without additional information, the re-identification key stays in the EEA under the exporter's sole control, and the data cannot be re-identified even when cross-referenced with information the recipient state's authorities may hold. The Recommendations treat pseudonymised data as personal data; the measure reduces, rather than removes, the transfer.
Governance layer
- Article 30 record updated to include each ChatGPT processing activity, lawful basis, recipient, retention, and transfer mechanism.
- DPIA completed for any high-risk processing. Refresh it whenever the tier, model, or data scope changes.
- Transparency notices updated to disclose AI processing where required by Articles 13–14 and the EU AI Act's Article 50.
- Data subject request workflow that can locate, export, rectify, and erase personal data within ChatGPT — including the memory feature.
- Training and acceptable-use policies that name ChatGPT explicitly. Generic "be careful with AI" guidance is not a control.
EU AI Act overlay
The EU AI Act doesn't replace the GDPR; it stacks on top of it. For ChatGPT specifically:
- General-purpose AI obligations (Articles 51–55) attach to OpenAI as a GPAI provider, including transparency about training data and copyright, model evaluation, and serious-incident reporting. Most of these are OpenAI's obligations, but a deployer needs evidence the provider is meeting them.
- Annex III high-risk classifications apply when ChatGPT is used as a component of a system in education, employment, access to essential services, law enforcement, migration, justice, or critical infrastructure. Following Regulation (EU) 2026/1744, these obligations apply from 2 December 2027 (Article 113 as amended). High-risk obligations include risk-management systems, data governance, technical documentation, logging, transparency, human oversight, and post-market monitoring.
- Article 50 transparency obligations — synthetic content must be marked, people must be informed when interacting with an AI system — applied from 2 August 2026 (with a transition to 2 December 2026 for Article 50(2) marking by generative systems already on the market) and reach deployers, not just providers.
- Prohibited practices (Article 5) include emotion recognition in workplaces and schools, untargeted scraping of facial images, and certain biometric categorisations. ChatGPT processes input that touches these areas if you let it; the prohibition runs to the deployer, not OpenAI.
The EU AI Act's risk classifications cascade into GDPR through Article 35 — high-risk use under the AI Act will almost always meet the Article 35 threshold under the GDPR.
Frequently asked questions
Is ChatGPT Enterprise GDPR compliant?
ChatGPT Enterprise is the closest commercially available off-the-shelf ChatGPT tier to a GDPR-defensible deployment, but "compliant" is a posture, not a label. Enterprise gives you the DPA, training-excluded-by-default, EU data residency for stored content on eligible plans, audit logs, and SAML SSO. The deployer is still responsible for lawful basis, transparency, data subject rights, DPIA where applicable, and Article 32 measures — including data minimisation. Enterprise plus device-level masking plus a documented DPIA plus an updated Article 30 record is a materially stronger posture. Enterprise alone is not.
Can EU teams use ChatGPT Free or Plus for personal data?
No, and this is not a close call. Free, Plus and Pro are consumer subscriptions without a controller-processor DPA. Using them to process personal data of EU data subjects on behalf of an employer is an Article 28 breach (no DPA), almost always an Article 6 or 9 breach (no documented lawful basis), and usually an Article 32 breach (no documented technical measures). The exposure is structural, not theoretical: the DPA, the lawful-basis record and the technical measures are all missing at once.
Is ChatGPT covered by the EU-US Data Privacy Framework?
Not through OpenAI itself, which is not certified. The EU-US Data Privacy Framework rests on the Commission's adequacy decision of 10 July 2023 and remains in force. OpenAI does not appear on the Department of Commerce's Data Privacy Framework List, checked on 18 November 2026. Its DPA (version v.010126) has OpenAI Ireland transfer EEA and Swiss data onward under SCCs, so that is the mechanism to document. Where a provider participates, the DPF covers data transferred from the EEA to its US operations subject to the framework's principles. It is one mechanism among several. The DPF does not, on its own, cure deficiencies in lawful basis, transparency, retention, or data minimisation; those are still the controller's obligations.
How does the new memory feature affect GDPR analysis?
Memory turns ChatGPT into a system that retains personal data across sessions and can use it to personalise future responses. Under the GDPR, that retention is itself processing — and it has to be on a lawful basis, surfaced in transparency notices, and subject to data subject rights. In practice, memory features should be off for any work account unless the controller has explicitly evaluated them, included them in the DPIA, and built a workflow to honour erasure and access requests against the memory content. Check OpenAI's current documentation for the memory controls each tier exposes to users and to administrators.
What is the role of pseudonymisation in a 2026 GDPR posture?
Article 32(1)(a) names pseudonymisation as a security measure. Recital 28 frames it as a way to reduce risks to data subjects and help controllers meet their obligations. EDPB Recommendations 01/2020 treat pseudonymisation in the EEA before transfer as a supplementary measure that can support international transfers, under the Use Case 2 conditions described above. In an AI workflow, pseudonymisation at the source is one of the most direct Article 32 measures available: it removes identifiers before any cross-border transfer happens. It does not by itself take the processing outside the GDPR — Recital 26 treats pseudonymised data as personal data — but it reduces what reaches the processor and supports a defensible minimisation posture. A tool that masks on the device, across the AI websites, command-line tools and SDKs staff already use, is one practical way to deliver pseudonymisation across an organisation; the legal determination remains the controller's or counsel's.
Do we need a DPIA before letting employees use ChatGPT?
If the use is "likely to result in high risk to the rights and freedoms of natural persons," yes — Article 35 makes a DPIA mandatory. The ICO says AI processing of personal data triggers a DPIA in the vast majority of cases, and the CNIL's AI how-to sheets include a sheet on carrying out one; Article 35(3) separately mandates a DPIA for systematic evaluation of personal aspects and for large-scale special-category processing. In our reading, organisation-wide deployment of ChatGPT on personal data will usually meet that threshold. The scope can be tier-specific (separate DPIAs for Free/Plus blanket-prohibited use vs. Enterprise sanctioned use) and should be refreshed whenever the model, the tier, or the data scope materially changes.
The bottom line
ChatGPT is GDPR-compliant in exactly the same sense that AWS is GDPR-compliant: it is the building block of a defensible posture, not a finished one. The deployments that hold up under EU supervisory-authority scrutiny are the ones that combine the contract layer, the tier-and-settings layer, the data-minimisation layer, and the governance layer — and that put pseudonymisation as close to the user's keyboard as possible.
If you only do one thing this quarter, it is not "buy ChatGPT Enterprise." It is "stop personal data from reaching ChatGPT in the first place." Everything else in the posture compounds from there.
Sources and primary references
Reviewed November 2, 2026. Regulations are cited to their official text. Vendor statements are cited to the vendor's own documentation as it read on that date; vendors revise these pages, so confirm before relying on them.
- GDPR Article 5 — Principles relating to processing — purpose limitation, data minimisation and accuracy.
- GDPR Article 9 — Special categories of personal data — the Article 9(2) conditions contain no legitimate-interests ground.
- GDPR Article 13 — Information to be provided — recipients, transfers and retention period in transparency notices.
- GDPR Article 32 — Security of processing — pseudonymisation and encryption named in Article 32(1)(a).
- GDPR Article 33 — Notification of a personal data breach — the 72-hour notification duty at issue in the Garante case.
- GDPR Article 35 — Data protection impact assessment — the high-risk trigger and Article 35(3) cases.
- GDPR Recital 26 and Recital 28 — pseudonymised data remains personal data; pseudonymisation reduces risk.
- Garante — ChatGPT: il Garante privacy chiude l'istruttoria (20 December 2024) — the €15 million fine, the six-month campaign, the findings, and the notice of the March 2026 annulment.
- Garante — ChatGPT press release of 31 March 2023 — the temporary limitation on OpenAI's processing after the 20 March 2023 breach.
- Tribunale di Roma, judgment of 18 March 2026, R.G. 4785/2025 (PDF, copy hosted by Sapienza University of Rome) — annulment of Garante decision No. 755/2024 on one-stop-shop competence grounds.
- EDPB — Report of the work undertaken by the ChatGPT Taskforce (PDF) — OpenAI's EU establishment from 15 February 2024, lawful basis for training stages, and Article 13 duties for prompts.
- EDPB Opinion 28/2024 on AI models (PDF) — legitimate interest for development and deployment; pseudonymisation and masking as mitigating measures.
- EDPB Recommendations 01/2020 on supplementary measures, v2.0 (PDF) — Use Case 2 conditions for pseudonymisation before transfer.
- European Commission — EU-US data transfers — the 10 July 2023 Data Privacy Framework adequacy decision.
- EU AI Act, Article 5 — Prohibited AI practices — workplace emotion recognition and untargeted facial scraping.
- EU AI Act, Article 50 — Transparency obligations — provider and deployer disclosure duties.
- EU AI Act, Chapter V — General-purpose AI models — Articles 51–56 obligations on GPAI providers.
- EU AI Act, Article 113 (as amended) — Annex III high-risk obligations from 2 December 2027.
- OpenAI API — Your data — API content not used for training, 30-day abuse-monitoring retention, Zero Data Retention approval, EU region requirements.
- ICO — Accountability and governance implications of AI — DPIA triggered in the vast majority of AI cases.
- CNIL — What is the scope of the AI how-to sheets — the how-to sheets include one on carrying out a DPIA when necessary.
Related GDPR guides
- GDPR and AI: A 2026 Compliance Guide for European Teams
- Is Claude GDPR Compliant?
- Is Gemini GDPR Compliant?
- Is Microsoft Copilot GDPR Compliant?
- Is Perplexity AI GDPR Compliant?
- Is Grok GDPR Compliant?
- GDPR Data Subject Rights and AI
- Claude EU Data Residency in 2026
- Which AI Tools Train On Your Data by Default? The 2026 Audit
Send the AI less personal data
Locke runs on your own computer and replaces the names, email addresses and other personal data it detects with placeholders before a prompt reaches the AI tools it screens, so less personal data is transferred to the AI provider. It does not make an AI tool GDPR compliant or replace a DPA, a transfer mechanism or a DPIA, and anything it does not detect is sent as typed.
Sonomos, which publishes this blog, makes Locke. Free for one person on x64 Windows and Linux; macOS is coming soon.
This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.
Reference pages for this post
Related Articles
Best Privacy-Focused AI in 2026: Chatbots That Keep Your Data Private
The most private AI runs locally, via Ollama or LM Studio. Lumo, Brave Leo and Venice say they keep no chat logs; Duck.ai hides your IP. Compared, Nov 2026.
Can My Employer See My AI Chats? ChatGPT, Claude, Gemini and Copilot
Yes, if your employer controls the account, the computer or the network. Work accounts on ChatGPT, Claude, Gemini and Copilot let it read or keep chats. A side-by-side table of what each vendor documents, as of November 2026.
Can My Employer See My Copilot Chats? Work vs Personal Accounts
Copilot chats on a work account sit in your mailbox, where your employer can search, hold and export them with Purview. A personal account sits outside the tenant, but a work laptop can still capture it. What Microsoft documents, as of November 2026.