DeepSeek went from an unknown Chinese lab to the #1 free app on the US App Store in about a week, and it's still free. That's exactly what makes people nervous. So I read its full privacy policy, both US government evaluations of its models, and every ban I could verify for this DeepSeek review, to answer one question: is DeepSeek safe to use, and when is it worth it?
How I researched this review
I read the current privacy policy in full, both NIST CAISI evaluation reports, the official API pricing page and the release changelog. I checked every government ban against its official announcement. DeepSeek has no affiliate program and pays me nothing.
Some links are affiliate links · they support this site at no extra cost to you · Full disclosure
1. What Is DeepSeek? The Quick Version
It's an AI lab based in Hangzhou, China, funded by the hedge fund High-Flyer. It builds large language models, the same technology behind ChatGPT and Claude, and releases most of them as open weights under the MIT license. Anyone can download, run, and modify them.
The chatbot went viral when the R1 reasoning model launched. Within a week it passed ChatGPT as the most-downloaded free app on the US App Store, and Nvidia's stock fell 18% in one day on fears that cheap Chinese AI would undercut US chipmakers.
The lineup has moved fast since then. The current workhorse on the DeepSeek API is V4.1-Flash, a 552-billion-parameter mixture-of-experts model with a 1-million-token context window. DeepSeek is retiring V4-Pro and routing those requests to Flash, because its own tests put the smaller model ahead.
Why is the app free? There's no subscription and no ads in the chat app. The company earns money from developers who pay for the API, and its privacy policy lets it use what app users type to improve its models. So the price you pay is your data, which is why the policy matters more than the price tag.
There are two DeepSeeks: a free app that sends your chats to China, and an open model you can run on hardware you control. Most of the safety confusion comes from mixing them up.
2. Is DeepSeek Safe? 5 Proven Risks I Found
I kept to what's documented: DeepSeek's own policy, NIST's model evaluations, and security reports that the company later patched. Here's what holds up.
Risk 1Your chats are stored in China
The DeepSeek privacy policy says it collects, processes and stores personal data in the People's Republic of China. That covers your text and voice prompts, uploaded files, chat history, device model, IP address, and approximate location.
China's National Intelligence Law can require companies to cooperate with state intelligence work. That clause, more than model quality, is behind almost every government ban.
Risk 2Your prompts can train future models
The policy lists "training and improving our technology" as a use of your inputs. You have the right to opt out, but there's no toggle: you request it by email at [email protected]. Your data is kept for as long as your account exists.
Risk 3Its guardrails are weak
NIST's Center for AI Standards and Innovation (CAISI) found that DeepSeek agents were 12 times more likely than US frontier models to follow malicious instructions hidden in the content they were working on. In jailbreak tests, DeepSeek models complied with over 90% of harmful requests, compared with 6–10% for OpenAI's open model.
For casual chat, that just means weaker filters. If you build an AI agent that reads your email or browses the web, it's a real attack surface.
Risk 4Some answers are censored
CAISI found that DeepSeek models echoed Chinese Communist Party narratives more often than US models. Ask the app about Tiananmen Square or Taiwan and you'll get a refusal or the official line. It won't affect a spreadsheet formula, but don't use DeepSeek for research on politics or history.
Risk 5Its security track record is patchy
Security firm Wiz found an exposed DeepSeek database holding over a million log lines, including plain-text chat history and secret keys. It was locked down about 30 minutes after Wiz reported it. Researchers at NowSecure also flagged weak encryption in the iOS app. Both issues were fixed, but they show how mature the company's security is.
🇺🇸 Is DeepSeek banned in the US? Not for personal use. It is banned on government devices: NASA, the US Navy, the Pentagon and Congress restricted it, and so did several states, starting with Texas, then New York and Virginia. Many employers block it too, so check your company's AI policy before you use it for work.
Myth vs reality
🚦 What's safe to type into the DeepSeek app?
The problem with DeepSeek isn't the model. It's the app's address. Your chats go to a jurisdiction where the government can ask for them.
3. DeepSeek vs ChatGPT vs Claude: Side by Side
All three have a free tier, so price alone doesn't decide it. The real differences are where your chats go, how much control you get over them, and what each assistant does best.
If you're choosing a daily assistant for work, ChatGPT and Claude win on privacy controls. If you want ideas for everyday use, see how I use ChatGPT every day and my Claude AI daily workflows.
DeepSeek's advantage is on the developer side: open weights and rock-bottom API prices. That's a different buyer from someone looking for a chat app.
For a chat app, all three cost the same: nothing. You're choosing who keeps your conversations.
4. Where DeepSeek Is Worth It
None of the risks above make the technology bad. NIST called V4 Pro the most capable Chinese model it has evaluated, with its best results in math, software engineering and natural sciences. It was also more cost-efficient than GPT-5.4 mini on 5 of 7 benchmarks.
Developers on a budget
V4.1-Flash output costs $1.20 per million tokens at peak and half that off-peak. Premium US models can charge $50 for the same volume.
Math & step-by-step logic
DeepThink mode shows its reasoning step by step, which helps when you're checking a proof or a tricky calculation.
Huge documents via API
A 1M-token window fits an entire codebase or a stack of reports in a single request.
Offline tinkerers
Open weights let you run smaller versions on your own machine. Nothing leaves your computer.
Switching is easy for developers too. The API accepts both OpenAI-style and Anthropic-style requests, so most existing apps and coding agents can point at it by changing a base URL and a key. That's why it shows up in so many budget coding setups.
DeepSeek's real edge isn't the free app. It's cheap, open models that developers can run anywhere.
5. 3 Safer Ways to Use DeepSeek
If you still want to try it, these three options go from least effort to most privacy.
1Use the app, but lock it down
- Sign up with a separate email, not your work account.
- Don't upload files or photos.
- Delete your chat history regularly in Settings.
- Email [email protected] to opt out of model training.
- Remove names, numbers and codes before you paste anything.
2Run it on your own computer
Free tools like Ollama and LM Studio run DeepSeek models fully offline. Install Ollama, open a terminal, and type:
ollama run deepseek-r1:8b
One caveat: the full V4.1-Flash is far too big for a laptop. What you run at home is a smaller distilled version. It's slower and less capable than the app, and 16GB of RAM is a comfortable minimum. For more options, see my guides to open-source AI models and offline AI tools.
3Use a host outside China
Because the weights are open, other cloud providers can serve DeepSeek models from their own servers in the US or EU. Before you paste anything, check the provider's data-location and retention pages. The model is the same, but the privacy rules are the host's. If a provider won't tell you where it stores prompts, treat it like the official app.
Same model, different address. Running DeepSeek locally removes the risk that makes governments ban the app.
6. Who Should Use DeepSeek (and Who Should Skip It)
👍 Good fit
- Developers who need cheap API calls
- Students working on math and logic
- Hobbyists who want to run AI offline
- Anyone asking casual, non-personal questions
👎 Skip it
- Government workers and contractors
- Anyone handling client or patient data
- Researchers covering politics or history
- Teams building agents that touch email or the web
7. Best DeepSeek Alternatives (No Data Sent to China)
If the data question is a dealbreaker, these assistants are run by US companies and have free plans to start with.
✳️ Claude
Free · Pro $20/mo
Best for writing and long documents. Training on your chats can be switched off in settings.
💬 ChatGPT
Free · Go $8 · Plus $20/mo
The all-rounder: chat, images, voice and search in one app.
🔎 Perplexity
Free · Pro plan available
Research with sources you can check. More in my Perplexity review.
🧩 Monica
Free daily credits · paid plans
GPT, Claude and Gemini in one browser sidebar, so you don't need three subscriptions.
Worried about free tools in general, not just this one? Read what free AI tools do with your data.
▶️ Video credit: "Deepseek did it again..." by Matthew Berman. Worth watching for a look at DeepSeek in action.
8. My Verdict: Is DeepSeek Worth Using?
As technology, yes. DeepSeek builds strong, cheap, open models, and even the US government's own evaluators rank them as the best to come out of China. As a free chat app for everyday life, I'd only use it for questions I wouldn't mind seeing on a public forum.
For personal or work use, an assistant from a US company with a training opt-out in settings is the lower-risk default. For developers and tinkerers, running DeepSeek locally gives you the best of both.
What I'd do in your shoes
- Just curious? Try the app with a throwaway email and zero personal details.
- Using AI for work? Pick Claude or ChatGPT, and switch off training in settings on day one.
- A developer watching costs? Test the API on public data, or self-host the open weights for anything sensitive.
Want AI that works for you, not the other way around?
Get my weekly picks of AI tools I've checked, with the privacy fine print summarized.
Join the free newsletter →Frequently Asked Questions About DeepSeek
🛡️ Lock Down the Rest of Your AI Setup
Now you know where DeepSeek sends your chats. These three reads cover the rest of your AI toolkit: