AI and privacy
Where your data ends up at DeepSeek, what the supervisory authority says, and why the app differs from the open model you can run yourself.
DeepSeek's app and website are not suitable places for business data containing personal data. According to its own privacy policy, input and uploaded files are stored in China and used to improve the models. The Autoriteit Persoonsgegevens warns against it. The open model on your own infrastructure is a different situation, with its own questions.
This article describes what the sources say: DeepSeek's privacy policy and terms, the GDPR and supervisory authorities' publications, consulted on 13 October 2026. It is an explanation, not legal advice. A decision for your own organisation requires your data protection officer or a lawyer.
Someone searching “is DeepSeek safe” usually means two things at once. The first question is where data goes and what the recipient may do with it. The second is whether the model itself answers reliably. This article covers the first question. Its answer differs by method of use.
DeepSeek is a Chinese AI company developing language models and offering a chatbot under the same name. Its privacy policy and terms name the provider as Hangzhou DeepSeek Artificial Intelligence Co., Ltd., with a registered address in China. What a language model is and how it works is explained in What is an LLM?.
The name refers to three things you must distinguish:
For the R1 series of reasoning models, DeepSeek says in its release documentation that the code and model weights are under the MIT licence and commercial use is allowed. Downloading the weights lets you run the model without connecting to DeepSeek. That does not automatically apply to every model the company releases, and model names in the API change. Check the licence for each model.
According to the page itself, DeepSeek's privacy policy was last updated on 10 February 2026. It covers the apps, websites and associated services.
The policy is clear here. The company says it directly collects, processes and stores personal data in the People's Republic of China. For transfers from certain countries, it says it uses appropriate safeguards where required. It does not specify which safeguards.
The policy lists providing and improving services, security, and training and improving technology, including models. The rights section includes the right to object to using your data for training. The policy does not explain how to arrange this beyond giving a contact address and the option to manage conversation history.
No number of days or months is specified. According to the policy, account, input and payment data is retained while you have an account and thereafter as long as necessary for the stated purposes.
For users in the EEA, Switzerland and the United Kingdom, the policy has an addendum with lawful bases, rights and an EU representative. Having such an addendum does not establish that the transfer arrangements are in order. The next section covers that.
Chapter V GDPR regulates when personal data may go to a country outside the European Economic Area. Article 44 sets the starting point: only if that chapter's conditions are met.
The Autoriteit Persoonsgegevens describes three options for transfers to a country outside the EEA:
China is not on the supervisory authority's list of countries with an adequacy decision on that page. The first route is therefore unavailable. Without appropriate safeguards or an exception either, the same page says the transfer is unlawful and not permitted.
Article 46 requires more than a signed document. It requires enforceable rights and effective legal remedies for data subjects. The supervisory authority therefore points out that you assess case by case whether supplementary measures such as encryption or pseudonymisation are needed. You make that assessment yourself or have it made. This article cannot do it for you.
If an employee puts a client file into the app, your organisation sends the client's data to another country. The client does not know. Article 82 GDPR makes the controller liable for damage caused by processing that infringes the regulation.
On 3 February 2025, the supervisory authority published AP: take care when using the DeepSeek chatbot. Its main points, summarising its own words:
In the same notice, the AP announced a broader investigation into personal data transfers to China and said it was in close contact with other European supervisory authorities about DeepSeek and similar companies. No outcome specifically concerning DeepSeek could be found on the AP website on 13 October 2026. The notice remains, with 3 February 2025 as its last updated date.
On 30 January 2025, the Italian supervisory authority restricted processing of Italian users' data and opened an investigation. It found the companies' response wholly insufficient.
On 27 June 2025, Berlin's supervisory authority reported the app to Apple and Google as unlawful content. According to that press release, DeepSeek infringes Article 46(1) GDPR because it could not demonstrate equivalent protection in China for German users' data to that in the EU. What Apple and Google subsequently decided is not in the press release and was not checked for this article.
This is the distinction missing from most discussions.
The privacy policy described above applies here. Input and files go to DeepSeek, are held in China and may be used for training. There is no agreement between your organisation and the provider. The AP warning concerns this form of use.
A developer calling the API also sends input to DeepSeek. The open platform terms, effective since 29 April 2026, place obligations concerning end users' personal data on the developer and otherwise refer to the privacy policy. They are governed by the law of mainland China. A data processing agreement of the kind Article 28 GDPR requires could not be found in them in October 2026. The transfer question is the same as for the app.
An open model can also be offered by another party on its own infrastructure. Your input then goes to that provider rather than DeepSeek. What happens to your data depends on that provider's terms: where it processes, whether it trains on your input, how long it retains data and whether it signs a data processing agreement. These are the same questions as for any other AI service.
Running the model on your own server means input does not leave your organisation. There is no transfer because there is no recipient. Your own duties remain: security, access, logs, a retention period and whether the model answers well enough for the purpose. The AP mentions concerns about the underlying model without explaining them. Include this in your assessment and have the model tested before relying on it.
| Method of use | Where the data goes | What to arrange |
|---|---|---|
| App or website with your own account | To DeepSeek, storage in China according to the privacy policy | Do not use for personal data or confidential business information; include in the AI policy |
| DeepSeek API in your own software | To DeepSeek | A transfer basis, an agreement and a risk assessment; no personal data without all three |
| Open model with another provider | To that provider, on its infrastructure | Check data processing agreement, data location, training and retention period in its terms |
| Open model on your own hardware | Nowhere | Security, access, logs, retention period and a quality test |
| A colleague uses the app on a personal phone | To DeepSeek, outside your view | Policy, explanation and a usable alternative; if personal data is involved, assess whether it is a data breach |
As of October 2026, based on DeepSeek's privacy policy and terms. Check them yourself before deciding.
Which types of data do and do not belong in a chatbot, regardless of provider, is in May you put client data into an AI chatbot?.
The question with DeepSeek is the same as with every chatbot, only sharper: what goes with it, and where? PrivaSearch addresses the first part. Personal data in documents is detected within PrivaSearch and replaced with consistent pseudonyms, and a colleague checks the result before a document becomes searchable. The steps are on How PrivaSearch works.
You can search entirely within PrivaSearch, with passages and source references, per project and only for those with access. Semantic search uses a European AI provider and only pseudonymised, reviewed text. Processing and storage take place in data centres in the EU. The parties involved at each stage are in the sub-processor overview, and the principles on Hosting in the EU.
Anyone wanting an AI assistant to help search can connect Claude, ChatGPT Business or Mistral Vibe (formerly Le Chat). The assistant then receives only released, pseudonymised passages and no original files. What goes to the assistant's provider is described on the page about AI chatbots and the GDPR.
Pseudonymisation is not anonymisation: for your own organisation, it remains personal data. The lawful basis, transfer assessment and DPIA remain your responsibility and your data protection officer's.
Want to see how that works in practice? Book a demo.
That depends on how you use it. With the app and website, input and files go to DeepSeek, which says in its own privacy policy that it stores them in China and uses them to improve models. The Autoriteit Persoonsgegevens therefore warns users to exercise great caution and restraint. The open model on your own hardware sends nothing to DeepSeek but requires your own assessment.
The privacy policy says the company directly collects, processes and stores personal data in the People's Republic of China. This covers account data, everything you enter and uploaded files. The policy specifies no fixed retention period: data is kept while you have an account. This reflects October 2026. Read the policy yourself, as it is regularly updated.
For work involving personal data or confidential information, the app is not suitable. There is no adequacy decision for China, so transfers require appropriate safeguards or a GDPR exception and a risk assessment. An employee with a personal account has arranged none of that. Record what is and is not allowed in your AI policy and offer an alternative.
On 3 February 2025, the supervisory authority warned users to exercise great caution and restraint with the app. It expressed serious concerns about the privacy policy and noted that users' data is stored in China. Anyone uploading others' data may be liable. The AP also announced a broader investigation into personal data transfers to China.
No. The app is a DeepSeek service where your data goes to the company. An open model is a file containing model weights that you can run yourself or with another provider. Where the model runs then determines where your data goes. The Autoriteit Persoonsgegevens does say similar concerns also exist about the underlying model.
The privacy policy mentions the right to object to using your data to train models. It does not describe how this works in the app and provides a contact address for requests. Such a setting changes nothing about where your data is stored. For an organisation, it therefore does not resolve the transfer question.
Where an uploaded file ends up, how long it stays, who can access it and what deletion means. With four approaches for files containing personal data.
How schools use AI, why pupil data receives extra protection, which applications are prohibited or high risk, and what a school needs to arrange.
What ChatGPT stores, who can access it, how to turn off training, and what temporary chats, memory and shared links do. Includes a practical checklist.
We show it with fictitious client files from your sector: upload, detection, review, search and the connection with your AI assistant. Afterwards you'll know exactly what your team can use safely.
Book a demo →