Learning how to use DeepSeek safely starts with one line in its privacy policy: the company says it stores the information it collects on secure servers located in the People's Republic of China. That covers what you type, the files you upload, your feedback and your chat history, plus account details and device information such as your IP address and device model. (DeepSeek Privacy Policy (opens in a new tab))

The upside is real. The DeepSeek chatbot is free on the web and in its iOS and Android apps, and the company released its V4 models on April 24, 2026, with a context window of one million tokens that it says is now the default across its official services. So the goal here is not to scare you off. It is to set up DeepSeek so you share as little as possible, and to know what never to type into it. (DeepSeek API Docs: DeepSeek V4 Preview Release (opens in a new tab); Wikipedia: DeepSeek (chatbot) (opens in a new tab))

What DeepSeek does with your data

The privacy policy lists training and improving DeepSeek's technology, including its machine learning models, among the uses of the information it collects. Account data, including your inputs, is kept while your account is active, and the policy says some information can be kept longer to meet legal obligations or legitimate business interests. (DeepSeek Privacy Policy (opens in a new tab))

Kaspersky adds two points worth knowing. As a company based in China, DeepSeek may be required to hand data to Chinese authorities. And because chats can end up in training data, Kaspersky warns that a model can sometimes repeat training data to other users, which is why it advises keeping sensitive information out entirely. (Kaspersky (opens in a new tab))

Governments have reacted. On October 20, 2026, the Federal Register published a Department of Homeland Security order that tells federal civilian agencies to remove DeepSeek products from their systems and stop buying them, including third party products developed using DeepSeek. According to The Hacker News, the order does not restrict companies or the public. Earlier, Italy's data protection authority ordered DeepSeek blocked in January 2025, and Australia, Canada and South Korea barred it from government devices. (The Hacker News (opens in a new tab); Wikipedia: DeepSeek (chatbot) (opens in a new tab))

How to use DeepSeek safely: the settings to change

1. Sign up with email, not a social login

Kaspersky recommends creating your account with an email address and a strong password you use nowhere else, rather than signing in with Google. At the time of its review, DeepSeek offered no two factor authentication or passkeys, so the password is your only lock.

2. Turn off training on your chats

On the web, sign in at chat.deepseek.com, open the sidebar, click the three dots next to your name and choose Settings. Open Data and switch off the option to improve the model for everyone. Kaspersky found the same switch in the mobile app under Settings, then Data controls. This only covers future chats: Kaspersky says data already used for training needs an email to privacy@deepseek.com.

3. Delete shared links you do not need

Anyone who has a shared chat link can read that chat. In Settings, open Data and click Manage next to Shared links. Click the bin icon next to each link you want gone and confirm with Delete.

4. Clear old chats

To remove one chat, hover over it in the sidebar, click its three dots and choose Delete. To remove everything, go to Settings, then Data, and click Delete all next to Delete all chats. Kaspersky notes that this removes chats from your account, not necessarily from DeepSeek's servers.

5. Export, then delete the account if you are done

If you want a copy first, use Settings, then Data, then Export. DeleteMe says the download link expires after seven days, and Kaspersky says JSON exports are available only on the web. Then open Settings, then Profile, click Delete next to Delete account, type DELETE MY ACCOUNT and confirm. DeepSeek's policy says a deleted account cannot be reactivated or recovered.

What never to type into DeepSeek

Tip: Treat every message as something stored on a server you do not control. Keep out passwords and login codes, ID and account numbers, health details, client or employer documents, private photos, and anything about other people that they did not agree to share.

Tip: If your job involves a government agency or a federal contract, check your employer's rules before you open DeepSeek on a work device. The DHS order covers federal civilian agencies and, through a contract clause, the contractors working for them.

DeepSeek works best for tasks where nothing is private: explaining a concept, drafting a generic email, checking code that contains no secrets, or summarizing a public article. When you need help with something personal, remove the identifying details first. The prompt below lets you keep the structure of a text while the real names and numbers stay with you. (Kaspersky (opens in a new tab))

Prompt: rewrite a text with the private details removed

The text below uses placeholders like [NAME], [COMPANY], [CITY] and [AMOUNT] instead of real details.
Keep every placeholder exactly as it is, in the same place.
Rewrite the text so it is clear, polite and under 150 words.

Text:
[paste your text with placeholders here]

Tip: Once you get the answer back, swap the real names and numbers in yourself. It takes a minute, and the private part never leaves your computer.

The most private option: run DeepSeek on your own computer

DeepSeek publishes the weights of its models under the MIT License, while the web and mobile apps are proprietary. That means you can run a version of the model locally, where your chats never reach the company's servers. Kaspersky suggests LM Studio, which works offline, and says smaller models in the 7B to 13B range suit a laptop with 16 to 32 GB of RAM. The trade off, according to Kaspersky: a local model cannot look up current information online. (Kaspersky (opens in a new tab); Wikipedia: DeepSeek (chatbot) (opens in a new tab))

Running it locally does not remove every issue. Reports collected on Wikipedia's DeepSeek page say locally hosted versions still sometimes give answers consistent with Chinese Communist Party narratives, and the hosted chatbot limits answers on topics such as the Tiananmen Square massacre and Taiwan's political status. (Wikipedia: DeepSeek (chatbot) (opens in a new tab))

DeepSeek is a strong free model, and for public, low stakes questions it does the job. The problem is not the quality of the answers but where the conversation lives afterward. Switch off training, clear your shared links, keep personal and work details out, and use a local model or a different assistant when the subject is private.