Is ChatGPT GDPR-compliant? What OpenAI's contracts cover

Is ChatGPT GDPR-compliant?

Depending on the tier, a company using ChatGPT may be operating with no data protection contract at all: the personal tiers Free, Plus and Pro have none. The business tiers Team, Enterprise and the API come with a data processing agreement, but it does not cover special category personal data. For holders of professional and official secrecy, OpenAI is not an option today, in the assessment of the Zurich law firm VISCHER.

In brief

• Check first whether personal tiers are running in business use at your company. For Free, Plus and Pro, no data protection contract exists to justify that.
• Switching to a business tier resolves the case for ordinary business data. For health, religious or trade-union data, you additionally need a deliberate risk decision.
• Anyone handling client-mandate, patient or official records gets no further with the direct OpenAI contract. Two routes remain: a hyperscaler contract through Azure OpenAI, or a self-hosted open model.
• Anyone required to keep data in Switzerland is ruled out here. OpenAI offers only the collective region Europe.
• If you use Codex, check its own training switch separately. The coding agent runs under the same contracts but has its own separate setting for entire working environments.

The Zurich law firm VISCHER (authors Lucian Hunger and Jonas Baeriswyl) regularly updates a public market overview of AI providers; the July 2026 edition gives OpenAI its own section. Souverana is not a law firm: what follows is VISCHER’s legal assessment, quoted and put in context, plus the architecture questions a legal review naturally does not answer: through which contract, and on which infrastructure, a company can still deploy the same model class.

This article opens our series “Provider Data-Protection Check”, in which we review the major AI providers against the same grid: contracts, data residency, professional secrecy, routes around the gaps.

What VISCHER objects to in OpenAI’s contracts

As with any provider, the basic rule applies first: only the business tiers count for business use. VISCHER states explicitly that the consumer-facing offerings are not built and not suitable for business use, at OpenAI as at any other provider.

On the business tiers, where a DPA?Data Processing Agreement (DPA)The contract that governs how a provider may process personal data on a customer's behalf. Without a signed DPA, a cloud provider is not considered bound under data protection law. exists at all, VISCHER names three concrete reservations.

The contract is built for ordinary personal data, not for special category data. VISCHER’s assessment on this is blunt: the clause is an attempt by OpenAI to step out of responsibility for a correspondingly higher data security standard. Most customers do not notice, and the firm considers it questionable whether the move even holds up under data protection law.

The clause on new sub-processors leaves OpenAI free choice of notification channel: a blog post, an in-app notice, another “appropriate” form, or email, but only for those subscribed to it. A customer cannot rely on email alone; anyone who wants to be reliably informed would have to monitor every channel continuously, blogs included. VISCHER names the risk directly: new sub-processors can be added without the customer’s knowledge, so the customer cannot actually exercise their own duty to review.

On web search, finally, the public contract texts show no exception according to VISCHER, which is why the firm assumes the same rules apply as in chat. A competitor claimed to VISCHER that OpenAI uses the same Bing search as Microsoft for web search. If that were true, OpenAI could not keep its own DPA commitment: it permits data flow only through declared sub-processors. Microsoft, however, does not offer Bing search under a data processing arrangement. VISCHER cannot verify this but considers it sensible for a supervisory authority to ask OpenAI about it.

Asked directly how ChatGPT could be used in a way that respects professional secrecy, VISCHER reports receiving only unusable answers from an AI chatbot. Their impression: OpenAI shows no interest in a solution that would satisfy customers.

Codex: the same business terms, its own switch

VISCHER’s overview does not separately cover Codex, OpenAI’s coding agent for terminal and IDE. What follows is our own research into OpenAI’s public data protection disclosures, as of July 2026, not VISCHER’s assessment.

Contractually, Codex hangs on the same hook as ChatGPT: anyone using Codex through ChatGPT Business, Enterprise or the API is subject to the same business terms and the same DPA objected to in the previous section, including the same gap for special category personal data. There is no separate Codex DPA or separate Codex terms of use; the review from the previous section applies here unchanged.

Two details are Codex-specific and easy to miss. First, OpenAI stores API inputs and outputs for up to 30 days by default for abuse detection, even with training disabled; zero data retention is available only for eligible endpoints, on request through sales, not automatically for every Enterprise account. Second, Codex has its own separate setting in the interface for whether entire working environments are released for training, a considerably larger release than individual requests. It is independent of the general organisation-level switch and belongs on the checklist for every Codex rollout.OpenAI Help CenterOpenAI Help Center · 2026Using Codex with your ChatGPT planOpen source ↗

Professional secrecy: why OpenAI falls short

For law firms, medical practices, trustees and banks, the question of professional secrecy↗Law & Regulation · Part 3Professional secrecy and AI: what may go to the cloud'That does not work for us' is the standard answer to AI in law firms and medical practices. Art. 321 of the Swiss Criminal Code says something narrower, and most of daily work falls outside it entirely.Read article is often more pressing than general GDPR or revFADP compliance. On OpenAI, VISCHER’s answer is short.

Unlike Microsoft or Google, OpenAI, as of July 2026, has no known contractual add-on and no workaround that changes this. Anyone processing client-mandate or patient data therefore needs a different route than the direct OpenAI contract; more on that below.

EU data residency: closer than the US, but not Switzerland

One feature deserves its own attention, because at first glance it promises more than it delivers: EU data residency. OpenAI processes data for the API and for ChatGPT Enterprise, Edu and Healthcare in a chosen region on request, with zero data retention on eligible API endpoints.OpenAI Help CenterOpenAI Help Center · Januar 2026Data residency and inference residency for ChatGPTOpen source ↗

Key figure

9 countries

Source: OpenAI Help Center, Data residency and inference residency, January 2026

Two limitations remain. The choice applies only to newly created projects, not retroactively, and for the API only per project. And “Europe” is a collective region, not a Swiss answer: anyone needing that precision is more likely to find it with a provider that actually names a Swiss location, or by self-operating an open model.

Where ChatGPT stands on the sovereignty scale

How sovereign ChatGPT is for your company does not hinge on a single question but on six, each answerable independently of the others. Here is where OpenAI’s offering stands today on each one:

less sovereign more sovereign →

Jurisdiction

US (standard contract)

EU region, optional for new projects

Switzerland: no dedicated option

Model

GPT-5 class, closed

gpt-oss, open weights, Apache 2.0

no model with open training data

Software

ChatGPT app, closed

API, de facto standard, openly replicated

gpt-oss via your own inference software

Operation

OpenAI's own cloud

Azure OpenAI Service, hyperscaler contract

gpt-oss in your own or a Swiss data centre

Data

Free/Plus/Pro: no DPA

Business tiers: DPA excluding special category data

technically impossible only when self-hosting gpt-oss

Integration

proprietary ChatGPT interface

OpenAI API, widely adopted format

open in-house inference

The picture shows why “ChatGPT, yes or no” is the wrong question. OpenAI sits somewhere different on every axis, and only together do they show how sovereign a given setup actually is. The highlighted point marks where OpenAI’s standard hosted offering (the business tiers) sits on every axis. Individual dials can be moved: the two escape routes further below, self-hosted gpt-oss and Azure OpenAI Service, pull certain axes to the right. (We explain the full framework behind the six axes in our foundational article, What is sovereign AI?.) On operation, the difference lies between plain ChatGPT consumption and the API route through Azure OpenAI Service: the same GPT models also run under Microsoft’s own, hyperscaler-typical contract terms. That is a different legal starting point from OpenAI’s own terms of service; we place it in this series’ article on Microsoft Copilot. The model axis is the most surprising: OpenAI is not uniformly closed. With gpt-oss-120b and gpt-oss-20b, the company released two models under the open Apache 2.0 licence in August 2025, downloadable, self-hostable, fine-tunable.OpenAIOpenAI · August 2025Introducing gpt-ossOpen source ↗ As a hosted business tier, though, OpenAI does not offer them, more on that in the next section.

The escape route: gpt-oss

For tasks the smaller gpt-oss models can handle, self-hosting is a genuine sovereignty gain↗Local Inference · Part 1Self-hosting an LLM: GPT-5 performance for CHF 8’500Two RTX 4090s, CHF 8’450, three weeks of measurements. Which tasks a machine of your own can carry, which it cannot, and where the real limit sits.Read article: your contract then sits with a host such as Infomaniak, OVHcloud or PHOENIQS, not with OpenAI. It covers exactly what your company negotiates with that host, independent of OpenAI’s own business terms.

Anyone needing the full GPT-5 class instead finds the second route not at OpenAI directly but at the hyperscaler: Azure OpenAI Service offers the same models under Microsoft’s contract terms, with its own conditions for professional secrecy and abuse monitoring. For a company with mixed data classes, that is often the more pragmatic combination than a single provider for everything.

Where to start

Before your next ChatGPT contract or the next renewal, clarify four points. First, whether a personal tier (Free, Plus, Pro) is running in business use at your company despite no DPA existing to justify it. Second, whether ChatGPT processes special category personal data or even professional-secrecy data, for which the business contract, according to VISCHER, is not sufficient. Third, whether a specific region matters for your data, and whether the collective region Europe is enough or you need a more precise answer. Fourth, whether a hyperscaler contract through Azure OpenAI or a self-hosted gpt-oss would be the better fit for the application in question than OpenAI’s own business terms.

Against the same grid, the series also reviews Anthropic, Google, Microsoft, and Proton, a provider that does it differently.

Does ChatGPT run in business use at your company? Then the tier is worth checking, not the model.

That is exactly what decides whether a data protection contract exists at all. Free, Plus and Pro have none; the personal tiers are not built for business use, and OpenAI trains on Codex tasks by default through Plus and Pro.

Team, Enterprise and the API come with a data processing agreement, but it does not cover special category personal data. For professional and official secrecy, OpenAI is not an option at all today, according to the Zurich law firm VISCHER, and there is no Swiss data residency, only the collective region Europe.

Our conclusion: the model class is not the problem, the route to it is. The same GPT models run through Azure OpenAI under Microsoft's contract terms, and for suitable tasks, gpt-oss is available under Apache 2.0 for self-hosting at a Swiss host. Then the contract is with your host.

Analysis, not legal advice. The article compares the four routes to the same models, the Codex details, and six questions for the next contract. Link in the comments.

#AI #DataProtection #OpenAI #Switzerland