DeepSeek safety profile

What it does with your data

DeepSeek uses what staff type into it to train its models unless someone turns that off. The privacy policy says prompts, uploaded files and chat history can be used to improve the service and to train its machine learning models. The terms of use say inputs and outputs may be used to develop and improve the service after encryption and de-identification. This covers the websites, apps, SDKs and APIs. Companies in the vendor’s corporate group also process user data for purposes that include foundation model training.

The vendor pages checked do not mention whether people review chats.

Retention is open-ended. Account data, inputs and payment data are kept for as long as you have an account. Beyond that, data is kept as long as the vendor needs it for the service, legal duties, business interests and legal claims. How long depends on the type of data. After an account is deleted, some data may still be kept where the law requires it.

The pages do not split training or retention by free and paid tier.

Settings to change first

Turn off training before anyone uses DeepSeek for real work. The terms of use name the setting but do not give a menu path, so each person will need to find it in their own account settings.

1. Turn off “Improve the model for everyone” in every staff account. This stops inputs and outputs being used to develop or improve the service.
2. Decide on chat history. The privacy policy says users can copy or delete chat history through settings. Ask staff to delete chats they no longer need.
3. If you are in the EEA, Switzerland or the UK, you can also withdraw consent or object to processing based on legitimate interest. You send these requests by email to the privacy address in the privacy policy.

No admin controls are documented, so each person has to change these settings themselves.

Free and paid tiers

The vendor pages checked do not say that training use or retention differs between free and paid use. Treat every tier as training on your data by default until the vendor tells you otherwise.

What the pages do distinguish:

• Paid services on the open platform: the vendor collects order and transaction data for ordering, payment and support.
• The DeepSeek App: the cookie and payment data categories do not apply.
• Apps built by developers on the open platform: the privacy policy does not cover the data of people using those apps. The developer who built the app is responsible for that data.

On the evidence checked, paying does not get you different handling of what staff type in.

Admin controls

The vendor pages checked do not describe any admin controls. They say nothing about single sign-on, central user management, or an organization setting that turns off training for everyone. You cannot enforce settings from one place, so you depend on each staff member to get them right.

On certifications, the privacy policy shows two badges: “GDPR Certification: Art 27 representation by Prighter” and “UK-GDPR Certification: Art 27 representation by Prighter”. These relate to the vendor appointing Prighter Group as its privacy representative in the EU and UK. The policy does not describe them as security certifications or audit reports. No page checked names SOC 2, ISO 27001 or any audit report.

Using it with donor and client data

Our advice is to keep donor and client data out of DeepSeek. The vendor collects, processes and stores personal data in the People’s Republic of China, and the pages mention no choice of region. Training is on by default, and no admin controls are documented.

If you allow DeepSeek at all, limit it to public or made-up material, with training turned off. Never paste in:

• donor names with gift amounts or contact details
• client case notes or anything that identifies a person you serve
• health, payment or bank details
• staff records or internal finance reports

Before anyone uses it with donor data, get these answers in writing from the vendor:

• whether any admin controls exist for organizations
• whether a paid plan changes training or retention
• how long chats are kept after you delete them
• which security audits, if any, the service has passed

Questions people ask

Does DeepSeek train on what we type?

Yes, by default. Prompts, files and chat history can be used to train its models. Turning off “Improve the model for everyone” opts you out.

Where is our data stored?

In China. The privacy policy says the vendor collects, processes and stores personal data in the People’s Republic of China. The pages checked do not mention any way to pick a region.

Is our data deleted when we close the account?

Not all of it. Account data and inputs are kept for as long as you have an account, and the terms say some data may still be kept after deletion where laws require it. A deleted account cannot be reactivated.