Where does your data end up when employees use ChatGPT?
In April 2023, Samsung's semiconductor division discovered three incidents within a few weeks. One engineer had copied faulty source code from a chip manufacturing database into ChatGPT, a second asked it for help with a broken machine, and a third had it write minutes from the recording of a confidential meeting. Shortly afterwards, Samsung banned generative AI on company devices (Source, Source).
None of the three wanted to give away secrets. They wanted to work faster. That is exactly what makes the case so typical, for German companies too. This article shows the path your data takes in the cloud, why a data center in Europe changes less than is often assumed, and how local AI avoids the core problem: handing data to a third-party provider.
In short: What employees type into ChatGPT, Gemini or Copilot usually ends up with a US company. Depending on the plan, it is stored, used for training and can be requested by US authorities, even if the server is in Europe. With local AI, the model computes in your own building, and inputs go to no AI provider.
Employees already use AI, often without IT knowing
According to the German digital association Bitkom, almost half of all employees in Germany use AI at work, 12 percent of them without their employer knowing. 71 percent of all AI users turn to ChatGPT (Source). In 2025, by contrast, only 23 percent of companies had rules for using AI, and only a quarter provided an AI tool of their own (Source).
Where the company offers no tool, employees use their private accounts. Experts call this shadow AI. The concern has reached management: for 66 percent of companies using AI, data protection is an obstacle, and 45 percent fear that company data could fall into the wrong hands (Source).
What happens to the data in the cloud
Anyone who types a text into ChatGPT or uploads a document sends the data over the internet to OpenAI's servers. The answer is generated there, and the history is stored. What happens next depends on the account.
- Processed in the AI provider’s data center, usually a US company, even when it is located in the EU.
- Stored according to the provider’s rules.
- Training depending on the plan, access possible via the CLOUD Act.
- No transfer to an AI provider.
- Stored according to your own rules.
- No training of third-party models with your data.
Private accounts. On the Free, Plus and Pro plans, OpenAI uses conversations to train its models by default unless users switch this off themselves (Source). At Anthropic, conversations from private Claude accounts are kept for up to five years if the user agrees to training (Source).
Business plans. On ChatGPT Business and Enterprise, OpenAI does not train on the data by default. The provider offers a data processing agreement, but the company has to conclude it itself via a form (Source). Data storage in Europe is only available for ChatGPT on the Enterprise plan, not on the Business plan for smaller teams (Source).
Sometimes data also takes unplanned paths. In July 2025, thousands of shared ChatGPT conversations appeared in Google search because users had ticked a box when sharing. OpenAI withdrew the feature (Source). In the New York Times copyright dispute, a US court even required OpenAI in 2025 to keep conversations that users had deleted. This affected private accounts and the Team plan, now called Business. The obligation ended in September 2025, and OpenAI still keeps part of the data from that period (Source). Once you hand data to a provider, you are no longer the only one deciding what happens to it.
Local AI and data protection: why a data center in Frankfurt is not enough
The large AI providers are based in the US and are subject to the CLOUD Act of 2018. The law requires US providers, upon an order under US law such as a search warrant, to hand over data in their possession or under their control, regardless of whether it is stored in the US or elsewhere (Source).
A hearing in the French Senate showed what this means in practice. In June 2025, Senator Dany Wattebled asked the legal director of Microsoft France, Anton Carniaux, under oath whether he could guarantee that data of French citizens held by Microsoft would never be passed to US authorities without the consent of the French authorities. Carniaux replied: "No, I cannot guarantee that." He added that this had never happened so far (Source).
There is also legal uncertainty. Data transfers to the US today are often based on the EU-US Data Privacy Framework, which currently permits transfers to companies certified under it. However, the European Court of Justice struck down its predecessors, Safe Harbor and Privacy Shield. The EU General Court dismissed an action against the current agreement in September 2025; the appeal before the European Court of Justice has not yet been decided (Source).
What the data protection authorities say
The German data protection authorities drew a clear line in their joint guidance on AI. The guidance is not a law, but it shows how the supervisory authorities assess cases (Source). With open systems such as cloud services, inputs leave the protected area of the company, with the risk that data is used for other purposes or becomes accessible to unauthorised parties. Their conclusion: "Technically closed systems are therefore preferable from a data protection perspective."
Where the AI runs also makes a legal difference. As soon as personal data is involved: if a company uses AI as a cloud service, the provider is usually its data processor, and a contract under Article 28 GDPR is required. With employees' private accounts, this contract is missing. If, on the other hand, the company runs the AI on its own servers for its own purposes, it is usually solely responsible according to the authorities, and no AI provider is involved in the processing.
For tax advisors, lawyers and doctors, professional secrecy comes on top. Under Section 203 of the German Criminal Code, they may only pass client or patient secrets to service providers where necessary, and must ensure that these are bound to confidentiality (Source). This cannot be achieved with a private ChatGPT account.
In June 2025, the Berlin data protection commissioner showed that the authorities take data flows abroad seriously. She reported the app of the Chinese AI provider DeepSeek to Apple and Google as illegal content under the Digital Services Act, because the company could not show that user data in China is adequately protected (Source).
Cloud AI and local AI compared
| Cloud AI (e.g. ChatGPT, Copilot, Gemini) | Local AI | |
|---|---|---|
| Where are inputs processed? | In the provider's data center | On your own hardware in the company |
| Who controls the infrastructure? | Usually a US company | Your company |
| Disclosure to foreign authorities | Possible, e.g. via the CLOUD Act, even with servers in the EU | No foreign provider involved |
| Storage of conversations | According to the provider's rules | According to your own rules |
| Use for training | Depending on plan and settings | No |
| Contract with an AI provider under Art. 28 GDPR | Required, missing for private accounts | Not required |
How local AI works
With local AI, the language model does not run in a provider's data center but on a computer in your own company. This is made possible by freely available models, for example from Alibaba, Google, Mistral or OpenAI, and by hardware barely larger than a book.
For employees, it looks like ChatGPT: a chat window in the browser, asking questions, uploading documents. Only the request never leaves the company network. It is answered on your own server and stored according to the company's rules.
Local does not automatically mean secure, though. Local AI also needs access rights, updates and backups, and the GDPR still applies: you need a legal basis and, for sensitive data, possibly a data protection impact assessment. If you commission a service provider to operate it, you conclude a data processing agreement with them. And if you connect the AI to a mailbox in the cloud, you bring part of the problem back. Local AI settles where the model computes. Which data it may see remains a deliberate decision.
What a system of your own costs is covered in the article What does local AI really cost?
When cloud AI is fine
Not every use of AI is a data protection problem. Having a job advert drafted or a public text summarised reveals nothing worth protecting. It becomes critical with customer data, personnel files, contracts, offers, design data or client files. But that is exactly where AI brings the most benefit in everyday work.
What you can do now
- Take stock: Ask openly which AI tools are already being used and for what.
- Set rules: Decide which data may go into which tools. A split into public, internal and confidential is enough to start with.
- Offer a safe alternative: A ban without an alternative only pushes usage into private accounts. Offering a tool where the data stays in the building removes the reason for shadow AI.
Free check: does local AI make sense for your company? Together we look at which data your employees process with AI and whether a local solution suits you.
How we implement local AI at Alae is described on our page Local AI.
Frequently asked questions
Is ChatGPT GDPR-compliant?
It depends on the plan and the use. There is no data processing agreement for private accounts, but there is for ChatGPT Business, Enterprise and the API. Whether a specific use is permissible is best clarified with your data protection officer or a specialist adviser.
Is it enough if the provider stores the data in the EU?
Not entirely. Under the CLOUD Act, US providers can be required to hand over data stored in Europe to US authorities. This requires an order under US law; how often this happens can hardly be verified from the outside.
Does local AI need an internet connection?
Not for answering requests. A connection is only needed to load updates or new models, and that only happens when you initiate it.
