DeepSeek privacy policy: what it says now, and every version since 2023

The Morse teamPublished November 9, 202611 min read

DeepSeek's privacy policy, last updated February 10, 2026, says DeepSeek collects what you type, say, and upload to its app, stores it in the People's Republic of China, keeps it for as long as you have an account, and may use it to train its models unless you opt out. It no longer lists "keystroke patterns or rhythms," and it no longer says advertisers send DeepSeek data about you. Both lines were in the policy for 71 days, from December 5, 2024 to February 14, 2025, and a guide that describes "early 2026" still quotes them.

On November 9, 2026, we read DeepSeek's current privacy policy, its terms of use (last updated March 27, 2026), its notes on how it trains models, its Open Platform terms, its Korean-language policy, and the one old version DeepSeek still hosts, from February 14, 2025. For every other past version, we read copies of DeepSeek's own policy pages saved by the Internet Archive's Wayback Machine, each labeled "archived" below, and compared each with the one before it, sentence by sentence. We found nine versions of the privacy policy and six of the terms. A version that came and went between two saved copies could be missing.

DeepSeek deserves credit on two counts. Every version since 2023 has said plainly that your data is stored in China, and in 2025 DeepSeek replaced an email-only way to refuse training with a switch in settings.

What does DeepSeek's privacy policy say today that it didn't before?

The rules below are in the February 10, 2026 version, each with the version that first had it. What hasn't moved in years, like storage in China and keeping data for as long as you have an account, is summed up in is DeepSeek safe.

How has DeepSeek's privacy policy changed since 2023?

In nine versions, DeepSeek went from a policy that cited China's Cybersecurity Law to one with sections for Europe and Korea, added advertising language and dropped it in the next version, and widened what it says about training. Each date below is the "Last Update" printed on that version.

Three things never changed. Every version stores your data in China, none says how many days or months your chats are kept, and every version since March 2024 lets DeepSeek share data with "law enforcement agencies, public authorities" to comply with "legal process or government requests." The age line changed twice, from "not aimed at children under the age of 18" in March 2024 to a heading for "children under 14" in February 2025, and then to "not aimed at children," with no age, in April 2025.

Does DeepSeek still collect keystroke patterns?

No. The phrase "keystroke patterns or rhythms" was in DeepSeek's policy for 71 days, and the May 15, 2024 version before it never had it. It sat in a list of "Technical Information" next to device model and IP address.

The same December 2024 version said advertisers and "measurement, and other partners" shared "mobile identifiers for advertising, hashed email addresses and phone numbers, and cookie identifiers" with DeepSeek, and that DeepSeek could share your data with "advertising or analytics partners." The next version removed both. Since April 2025, the policy has said in capitals that DeepSeek does no targeted advertising and doesn't "sell" personal data.

Anonyome's DeepSeek privacy guide, which we read on November 9, 2026 and which dates its account of government bans "as of early 2026," still says DeepSeek collects "your keystroke patterns," shares data with "advertising partners," and trains on everything you type in a way that "is not a setting you can turn of[f]." All three describe the December 2024 policy, which DeepSeek replaced in February 2025. The dates of the keystroke change are also in is DeepSeek safe.

How has DeepSeek's wording on storage in China changed?

It has grown more direct, though the place has always been China, as our guide to deleting a DeepSeek account sets out. The November 2023 policy said your data "will be stored on DeepSeek servers in mainland China." From March 2024 to February 2025 it said "secure servers located in the People's Republic of China." Since April 28, 2025, it has said DeepSeek will "directly collect, process and store your Personal Data in People's Republic of China."

The English policy doesn't name the companies that handle your data. The Korean-language policy, last updated May 6, 2026, does. In our translation, it lists Beijing DeepSeek Artificial Intelligence Co., Ltd., in China, receiving "de-identified user prompts" until "model training or optimization is complete"; Beijing Volcano Engine Technology Co., Ltd., in China, receiving device details down to battery level, network details, and your usage history until you delete your account; and Intuition Machines, Inc., in the US, which runs the hCaptcha bot check. A similar table sat in the English policy from April 28 to July 4, 2025, under a South Korea heading that cited Korea's Personal Information Protection Act, before DeepSeek moved it to the Korean version. We compare where a dozen AI apps keep your data in where AI chat data is stored.

How has DeepSeek's training wording changed by version?

It went from anonymized research data in 2023 to a named use with an email-only refusal, then a switch, then training by DeepSeek's group companies. Today the opt-out is the "Improve the model for everyone" switch in section 4.3 of the terms, and what DeepSeek says it trains on is in is DeepSeek safe.

• November 2023: de-identified data could be used "for academic research or statistical analysis to improve algorithm models" (archived).
• May 2024 terms: "DeepSeek collects and analyzes users' input and output information to enhance recognition and response capabilities" (archived). The August 30, 2024 terms dropped the sentence.
• December 2024: the policy named "training and improving our technology" as a use, with no way to refuse training in particular.
• February 2025: the way to refuse was an email to privacy@deepseek.com.
• April 2025: the email line left the policy, and the terms added the switch.
• December 2025: the policy listed opting out of training among your rights.
• February 2026: the policy added "foundation model training and optimization" to what DeepSeek's group companies do with your data.

What do DeepSeek's terms of use say, and what changed?

They put any dispute under "the laws of the People's Republic of China in the mainland," heard by a court where Hangzhou DeepSeek is registered, as every version since March 2024 has. The other changes, by version:

The March 2026 terms also reworded section 3.3. The April 2025 version gave DeepSeek the right to review "the behavior and information of users," "including but not limited to reviewing inputs and outputs." The current one says "the behavior of users," with "risk filtering mechanisms" as its example. DeepSeek hasn't said whether the reworded clause changes what it reviews, and the policy still says it uses your data to detect "abuse, fraud, and illegal activity." Section 11.1 says that if you keep using DeepSeek after new terms are posted, you've accepted them.

What does DeepSeek's privacy policy say for users in Europe?

It has had a section of its own since February 14, 2025, for the EEA, Switzerland, and the UK, with a table of legal bases. It rests training on "legitimate interests" to improve the platform and "conduct research," or on "your consent when we ask for it," and it rests teenagers' use, for those "14 or older but under the age of 18," on legitimate interests where a contract can't apply. It lists the GDPR rights to access, erase, object, and complain to a regulator. Since July 4, 2025, it has named Prighter as DeepSeek's representative in the EU and UK.

On January 30, 2025, two weeks before that section appeared, Italy's data protection authority ordered an urgent limit on DeepSeek's processing of Italian users' data (Garante). We cover Italy and the other government actions in is DeepSeek safe. How a European objection to training works in general is in how to object to AI training under the GDPR.

Where can you read DeepSeek's old privacy policies?

On DeepSeek's own site, only the February 14, 2025 version. It carries no notice that it has been replaced, and it came second for "deepseek privacy policy" in the search results we read on November 9, 2026, so it's easy to quote by mistake. The April 2025 policy linked to it with "You may find the previous versions of the Privacy Policy here." That link was gone by July 2025, and the current policy links to no old version. Everything else on this page came from the Wayback Machine.

The policy tells you to review it "each time you access our Services," and says a change is announced by a new "Last Updated" date, a posted policy, or "any other notice required by applicable law." There's no change log. We wrote about how 13 apps announce changes in AI privacy policy changes.

How to keep your chats out of DeepSeek's next policy version

DeepSeek changes its policy by posting a new "Last Update" date, and its terms say that using the service after a change means you accept it. Each of the nine versions above set the rules for whatever DeepSeek already held about you, on servers in China, under your account. The fix is a chat that never lands in that account. On Morse, DeepSeek is a model you pick from the list, with no DeepSeek account to sign up for, so there's no account of yours for a new DeepSeek version to apply to.

Your Morse chat history is kept in your browser, encrypted, and locked when you sign out. Our servers don't keep your messages or log what you write. The account holds sign-in details, your plan, your balance, and a usage ledger with no content (privacy policy). There's no training clause to watch for rewording, because Morse doesn't train AI on your content, and no advertising clause to come and go, because Morse doesn't sell your data or show ads. Because we don't store chats, we usually have very little to give authorities.

Each chat runs in the privacy mode you pick:

• Private, to use DeepSeek with no DeepSeek account behind it. New chats start here, on open-weight models like DeepSeek, Kimi, and GLM. Morse has zero data retention agreements with its inference, search, and other providers, so they keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't; otherwise Morse relies on the provider's commitment.
• E2EE, for questions you'd keep from every server, ours included. DeepSeek runs here in secure enclaves operated by third-party providers, not by Morse. Before anything leaves your device, your browser checks the enclave's hardware attestation and encrypts the message for that enclave alone, so our servers only pass along ciphertext they can't read. The protection depends on the enclave hardware and attestation working as designed. E2EE has no web search or image generation.
• Anonymous, to ask GPT, Claude, or Gemini the same thing. Our proxy strips your account, email, and IP address and sends the message under Morse's own account. The provider sees the message without knowing it's you and may keep it for a limited time to check for abuse, under its own policy, which can change on its own schedule, so leave out names.

Opening your history takes two keys, one of which never leaves your device. Clearing your browser deletes your chats for good, and Morse can't recover them, so copy out anything you'll need.

Send a first message from the Morse home page without an account, or start free with no card. DeepSeek V4.1 Flash and E2EE chats come with Pro, at $19 a month.