When DeepSeek exploded into the headlines in early 2025, two reactions arrived almost at once: “this is an amazing free AI” and “this is a Chinese app stealing your data.” Both are too simple. So is DeepSeek safe to use? The honest answer is it depends entirely on which DeepSeek you mean — and once you see why, you’ll be able to make the call for yourself instead of trusting a scary headline or a breathless rave.
The single most important thing to understand is this: DeepSeek is two different products wearing one name. Get that distinction right and almost every safety question answers itself.
The Two DeepSeeks
Product one: the open-weight model. DeepSeek released its models (like R1 and V3) as open weights under a permissive MIT license. That means anyone can download the actual model and run it on their own computer or server. When you run it locally, nothing you type ever leaves your machine — there’s no company on the other end receiving your prompts, because there’s no “other end” at all.
Product two: the hosted app, website, and API. This is the free DeepSeek chatbot most people actually used — the app that topped the App Store. It’s operated by a Chinese company, and like every free consumer AI service (including many American ones), it logs your prompts and stores them on its servers.
Almost all of the privacy and legal risk you’ve heard about belongs to product two, the hosted service — not the weights. Keep that split in mind as we walk through the three genuinely distinct concerns: privacy, security, and censorship. Each has a different level of severity, and — crucially — a different fix.
Risk 1: Privacy and Jurisdiction (the Hosted App)
This isn’t speculation; it’s DeepSeek’s own privacy policy. The policy states plainly that the company collects and stores your personal data — including your prompts, chat history, uploaded files, IP address, and device identifiers — on servers in the People’s Republic of China.
Why does that matter more than, say, a US app storing your data in America? Because data physically stored in the PRC falls under Chinese law, which includes national-security and intelligence statutes that can compel companies to hand over data to the state. To be precise and fair: this is a standing jurisdictional concern — a risk based on what the law permits — not a documented case of your specific chats being misused. But for anyone typing sensitive personal, business, or legal information, it’s a real consideration.
One widely-shared 2025 claim deserves a correction, because it’s a lesson in reading tech news carefully. Early versions of DeepSeek’s policy mentioned collecting “keyboard input patterns” — a genuine fingerprinting concern that made a lot of headlines in late January 2025. But DeepSeek removed that clause around February 14, 2025, and it’s not in the current policy. The PRC data-storage fact didn’t change; the keystroke clause did. Always date-check a viral privacy claim before you act on it.
Risk 2: The Security Findings (Also the Hosted Product)
In the weeks after launch, security researchers put DeepSeek’s hosted infrastructure under a microscope and found real problems. Each is worth attributing precisely:
- The exposed database (Wiz Research, Jan 29, 2025). Wiz found a publicly reachable DeepSeek database with no authentication required — over a million log entries including plaintext chat history and API keys, fully open to anyone who found the address. Important context: Wiz disclosed it privately and DeepSeek secured it quickly. This was an infrastructure-security failure, not proof of deliberate data theft, and it was fixed — but it shouldn’t have been possible in the first place.
- iOS app flaws (NowSecure, Feb 6, 2025). Researchers found the iPhone app transmitted some data unencrypted, disabled a key Apple transport-security protection, and used outdated encryption with hardcoded keys. NowSecure recommended enterprises remove the app from work devices.
- Weak safety guardrails (Cisco, Jan 31, 2025). Cisco tested DeepSeek-R1 against 50 standard “harmful prompt” attacks and reported a 100% jailbreak success rate — every guardrail-bypass attempt worked. For comparison on the same test, OpenAI’s o1-preview blocked far more. Read this as a snapshot of the model’s out-of-the-box alignment at that moment (hosted guardrails do get patched over time), but it signaled the safety tuning was thin.
The pattern here is engineering immaturity in a product that scaled to tens of millions of users almost overnight — not necessarily malice, but not reassuring either.
Risk 3: Censorship (Both Products, Strongest in the App)
DeepSeek’s hosted chatbot demonstrably avoids topics sensitive to the Chinese government. Reporters at outlets including the New York Times and Wired have repeatedly documented it deflecting questions about Tiananmen Square, giving notably thin answers about Xi Jinping, and sidestepping Taiwan’s political status.
Two things make this more nuanced than “it’s censored.” First, the filtering appears to work at more than one layer — some is baked into the model’s training, and some is applied by the app as you use it — which is why the raw downloaded weights behave differently from the app. Second, because the weights are open, outside researchers have been able to identify and reduce the censorship: Perplexity’s R1-1776 is a publicly released version fine-tuned specifically to answer previously-refused questions while keeping DeepSeek’s reasoning intact. Self-hosting a de-censored version reduces the behavior, though it does nothing to address the separate privacy question of the hosted service.
To be clear, this is a documented fact about a product’s alignment — nothing here is a comment on Chinese people or culture.
What the Government Bans Actually Restrict
You’ve probably seen that Italy, Australia, South Korea, Taiwan, several US agencies, and others have “banned DeepSeek.” It’s worth knowing what those actions actually cover, because the headlines oversell them. Nearly all of these restrictions apply to the consumer app on government-owned or work devices and networks — Italy’s data-protection authority went furthest, forcing an app-store removal in January 2025. They are institutional data-handling decisions about the hosted service. They generally do not ban ordinary citizens from using DeepSeek, and they say nothing about the open weights, which any developer can still download and run. (A note on freshness: most of these actions date from 2025; be skeptical of “2026 ban” headlines that are really just re-dated 2025 stories.)
So — Is DeepSeek Safe? A Decision Framework
There’s no single yes or no, but there is a clear way to decide for your situation:
- Casual, non-sensitive use of the app (brainstorming, general questions): the main cost is that your prompts are logged and stored in China. If that’s acceptable to you for a free tool, the practical risk is modest — just never paste anything you’d mind a company (and potentially a government) reading.
- Anything sensitive — personal, financial, medical, legal, or proprietary work data: don’t put it in the hosted app. This is the clearest call in the whole article.
- You want the model without the privacy risk: run the open weights locally or through a trusted provider. Nothing you type leaves your control. (This takes real hardware and some technical comfort — it’s not a one-tap option for a beginner, but it’s genuinely private.)
- You’re bothered by the censorship: a self-hosted or de-censored version (like R1-1776) largely addresses it; the app never will.
The fairest summary: DeepSeek the hosted app carries meaningful, well-documented privacy and security baggage and deserves caution with anything sensitive. DeepSeek the open model is one of the more transparent frontier AIs precisely because you can download it, inspect it, and run it disconnected from any company. Same name, very different safety stories — and now you can tell them apart.
❓ Frequently Asked Questions
Is the DeepSeek app safe to use? The hosted DeepSeek app logs your prompts and stores them on servers in China, per its own privacy policy, and researchers found real security flaws in early 2025 (an exposed database, iOS app weaknesses, thin safety guardrails). It’s fine for casual, non-sensitive questions if you accept that logging — but you should not enter personal, financial, medical, legal, or proprietary information into it.
Does DeepSeek steal your data? There’s no documented case of DeepSeek misusing individual users’ chats. The concern is jurisdictional: data stored in China is subject to Chinese law, which can compel disclosure to the state. That’s a risk based on what the law permits, not proof of theft — but it’s why sensitive data shouldn’t go into the hosted app.
Is it safer to run DeepSeek locally? Yes, for privacy. Because DeepSeek’s models are open-weight, you can download and run them on your own hardware, and nothing you type is sent to any company. It requires a capable computer and some technical setup, so it’s not effortless for a beginner — but it removes the hosted-service privacy risk entirely.
Why won’t DeepSeek answer some questions? DeepSeek’s hosted chatbot is documented to avoid topics sensitive to the Chinese government, such as Tiananmen Square and Taiwan’s status. The filtering happens partly in training and partly in the app. Because the weights are open, researchers have released de-censored versions (like Perplexity’s R1-1776) that answer those questions.
Did governments ban DeepSeek? Several governments restricted the DeepSeek app on official devices and networks (Italy went as far as an app-store removal), citing data-handling concerns. These are mostly institutional restrictions on the hosted service, not bans on ordinary citizens or on the downloadable open-weight model. Most date from 2025.
📚 Keep Learning
- Learn AI and ML Concepts Through the Lens of DeepSeek — how the model itself actually works.
- What Is Mixture of Experts? — why the full model needs serious hardware to self-host.
- Did DeepSeek Really Cost $6 Million? — separating the real story from the viral headline, just like we did here.
📚 Recommended Resources
Disclosure: some links below are Amazon affiliate links. If you buy through them, I earn a small commission — at no extra cost to you — which helps keep this site running.
Artificial Intelligence: A Guide for Thinking Humans
Melanie Mitchell · Book · ~$28 on Amazon
The best single book for doing exactly what this article asks of you: judging AI tools for yourself instead of trusting the hype or the fear. Mitchell explains, in plain language, what these systems really can and can’t do — the grounding you need to weigh any “is this AI safe?” question with a clear head.
Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow
Aurélien Géron · Book · ~$70 on Amazon
If the “run the open model yourself” path appeals to you, this is the reference that demystifies how these models actually work and how to work with them hands-on. The more you understand the machinery, the better you can decide what to trust with your data — and what to keep on your own hardware.