DeepSeek / Safety and privacy
Is DeepSeek safe to use? Check the service and the data first
DeepSeek may be reasonable to try with public or fictional material; do not submit sensitive or confidential data without an approved route. Trying the official DeepSeek chat with a fictional question is a different decision from uploading a client file, sending company code through an API, or running a model on your own computer. Before you share information, identify who operates the service, what its current policy says, and whether you are allowed to submit that information. Do not put passwords, private client records or sensitive personal data into an unapproved service.
ChinaAI is independent of DeepSeek. This is a source-based decision guide, not a penetration test, compliance certification or guarantee about any provider’s security. Your visit to this page does not send a prompt to DeepSeek.
By China AI Chat Team · Provider documents and the official web-chat settings reviewed September 24, 2026 · Our research method
First, which DeepSeek are you using?
“DeepSeek” may be the company’s own web chat or app, its developer API, an unrelated app that offers a DeepSeek model, or downloadable model weights running elsewhere. The same model name does not give all four routes the same operator or privacy terms. Start from DeepSeek’s official website when you mean its own services; inspect a third party’s domain, app publisher and policy before using its version.
On a phone, swipe this table sideways to read the final column.
| Route | Who can receive your input? | Check before use |
|---|---|---|
| Official web chat or app | DeepSeek and the service providers described in its privacy policy. | Review that policy, your account’s data controls and the sensitivity of each prompt or upload. Use public or fictional material for a low-stakes trial. |
| Official DeepSeek API | DeepSeek processes the input you submit. Your own app or workplace can also log or process it. | Check the Open Platform terms, your application design and any applicable file-storage setting. Keep your API key on the server. See our API guide. |
| Another app or model host | The app operator, its hosting chain and possibly a model API provider, depending on its setup. | Read that operator’s own policy and find out whether it sends prompts to DeepSeek or another provider. A “DeepSeek” label alone does not answer this. |
| Model weights run locally | Potentially only systems you control, but the runtime, integrations, telemetry and logs determine the real data path. | Inspect the software and network configuration; do not assume “local” means offline. Our local setup guide shows one tested route. |
The DeepSeek Privacy Policy explicitly says it does not govern the personal-data processing rules of a developer’s downstream app. Its developer terms say the model processes that app’s users’ input and make the app operator responsible for its end-user disclosures and safeguards. You need both the app’s policy and the model provider’s role to understand the complete route.
What DeepSeek says about data in its own services
The following points come from DeepSeek’s privacy policy, updated February 10, 2026. They describe the provider’s stated practices; they are not an independent audit. Open the current policy again when your decision depends on a detail.
On a phone, swipe this table sideways to see each practical action.
| Policy statement | What to do with it |
|---|---|
| Inputs can include prompts, uploaded files, photos and chat history. Device, network and usage information can also be collected. | Treat an attachment and its surrounding conversation as information you submit, not merely the final question. Remove identifiers and unnecessary content before uploading. |
| Personal data can be used to develop and improve services and train or improve models. The policy describes a right to opt out of use for training or optimization. | Check your current account controls before a new chat. Turning off a training control should not be mistaken for a promise that no data is processed or retained to deliver the service. |
| DeepSeek says it directly collects, processes and stores personal data in the People’s Republic of China. | If your organization requires a particular data location or approved vendor, obtain approval before submitting its data. An access point in another country does not establish a different storage location. |
| Retention varies by data type, purpose and legal requirements; the policy gives account-linked input as one example kept while an account exists. | Do not assume a universal number of days or that removing a chat from view immediately removes every copy. Read the current deletion terms or request clarification for a specific need. |
| The service is not designed to process sensitive personal data. Shared conversation links can expose a published dialogue to others and crawlers. | Keep medical records, identity details and similar information out of the service. Review a conversation before creating or publishing a share link. |
The policy gives a contact route for access, deletion and training-related requests at [email protected]; rights can depend on your location and exceptions. It does not provide one short, universal retention period for every chat or upload. Its controls and terms also do not tell you how a separate third-party app handles data.
Where to check the official web-chat controls
On September 24, 2026, we opened the signed-in official DeepSeek web chat and inspected profile menu → Settings → Data without changing the account. That panel showed an “Improve the model for everyone” switch, Shared links → Manage, Export data and Delete all chats. This is a dated observation of the web interface, not a promise that every account, region or mobile app has identical labels.
- Confirm you are on an official DeepSeek destination, then open your profile menu and Settings → Data.
- Read the explanation next to the training/improvement switch and choose the setting you want before adding new information. If the control is absent, check the current policy or ask DeepSeek rather than guessing.
- Review shared links and existing chat history. Export or delete through the available controls if that serves your purpose, and check the provider’s policy for how deletion requests are handled.
We did not switch the setting, delete a chat, test a data export or verify what happens to earlier training use or backend copies. Changing a toggle is not a substitute for deciding whether the content should be sent at all.
Before you paste, upload or connect a tool
Use this quick test for the specific data in front of you:
- Classify it. Public text, a fictional example and your own non-sensitive draft may be suitable for an ordinary trial. A customer contract, patient record, private codebase, access token or unreleased business plan needs a different decision.
- Check authority. For work or another person’s data, follow your organization’s approved-tool rules and obtain the necessary permission. If the service or data location is not approved, do not submit it.
- Reduce the input. Replace real names, account numbers, addresses and secrets with fictional placeholders where the task still works. A redacted excerpt can still reveal a confidential idea or enough context to identify someone; inspect it again.
- Trace the route. Verify the domain or app publisher, the operator’s current policy, training controls and any upload-retention setting. For an integrated app, check whether it forwards the prompt to another provider.
- Verify the result. Read cited documents and check calculations, code and claims independently before acting. Privacy controls do not make a model’s answer accurate.
Example: To ask for a summary of a private client PDF, first seek approval for that service. If there is none, make a fictional document with the same structure or use an approved route. Our DeepSeek PDF guide demonstrates the upload workflow with fictional source files; it is not permission to submit a real client’s file.
API file nuance: DeepSeek’s Files API guide says an uploaded image file can be assigned an expiry from one hour to 30 days; omitting the parameter keeps that file permanently. This is about that API file object, not a published retention period for consumer-chat attachments, all requests or logs. Developers should set the needed expiry deliberately and examine the rest of their application’s data path.
Security history is one part of the decision
On January 29, 2025, security company Wiz reported finding an unauthenticated DeepSeek database containing logs, including chat-history material and API secrets. Wiz says it disclosed the issue and DeepSeek promptly secured the database. That dated finding matters when weighing a provider’s history. It does not establish that the database remains exposed, that every DeepSeek chat was in it, or that we have audited the current system.
Account security and answer quality are separate questions. Use an authentic site or app, protect your login and API key, and do not accept a generated answer as verified fact. DeepSeek’s own policy warns that outputs can be inaccurate. Our DeepSeek overview links to small, dated task tests and the relevant guides; those examples are not a general safety or reliability benchmark.
Common questions
Does DeepSeek keep my chats?
Its policy lists chat history and inputs among collected data and says retention varies. It gives no single lifetime for every chat. The observed web settings offered history controls, but a screen-level deletion action should not be interpreted as immediate deletion from every system.
Does turning off model improvement make a chat private?
No setting alone establishes that. The policy describes a training opt-out right, while the service still needs to process input to answer and may retain data for other stated purposes. Check the current explanation and avoid submitting information that you cannot share.
Is the API more private than the official chat?
It is a different workflow, not a blanket privacy guarantee. DeepSeek processes input submitted to its Open Platform; an application built on it may add its own storage, logs and users. Check the exact app and provider terms. Protect keys server-side.
Is running DeepSeek locally fully private?
Running DeepSeek locally does not by itself guarantee complete privacy. Downloadable weights can run in an environment you control, but the app may still connect to cloud services, save prompts or send usage data. Check those connections, where chat history and logs are stored, and who can access them. See our tested local setup and inspect the actual configuration you choose.
Can I use DeepSeek for work?
Use your employer’s approved-tool and data rules. If approval is unclear, ask before entering company information. You can often explore a public question or fictional example without using internal material, subject to your organization’s policy.
Sources and scope
We reviewed DeepSeek’s Privacy Policy (last updated February 10, 2026), Open Platform Terms (effective April 29, 2026), Files API guide, its official service links, and Wiz’s original 2025 disclosure on September 24, 2026. The settings path above came from a read-only inspection of the official signed-in web chat on that date. We did not inspect backend systems, test account deletion or run a security audit. Policies, UI labels and model hosts can change. For a correction, contact ChinaAI; our methodology explains how we separate provider statements, observations and tests.