DeepSeek AI: risks and legal insights

There has been a huge amount of chatter over the last couple of weeks about DeepSeek, a new AI product from China. We have waded through some of that, so you don’t have to.  Here is what you need to know.

DeepSeek is an AI research lab in China. On January 20, 2025, it released an open-source AI model called DeepSeek-R1.  They claim it gives better results than currently available AI.

While competing AI models tend to leapfrog one another in their abilities, the big deal with DeepSeek is its training efficiency. It was trained for a fraction of the cost, using a fraction of the computing resources than current AIs. It uses a different, more efficient approach than typical brute force models. Traditional AI models use huge amounts of energy — so much that it will require a significant increase in electrical production. A nuclear reactor at Three Mile Island is being restarted to power Microsoft’s AI.

Some question whether the DeepSeek approach means the U.S. AI industry has totally missed the mark on how to develop AI.

DeepSeek apparently hallucinates as badly as other AI, so no advantage there.  But it does suffer from heavy Chinese government censorship. Testers report that many prompts that would deliver results that might differ from Chinese government views are met with canned refusal responses. Sometimes it starts to provide an answer and stops midway to refuse to respond.

Risk of overlooking malicious prompts

Most AI language models include safety guardrails to prevent them from providing harmful content, such as hate speech and bomb-making instructions. Researchers tested DeepSeek with 50 malicious prompts, and it didn’t detect or block any of them.

A common issue with AI is some products retain prompts and information that is input for training purposes which means that it could be spit back out based on prompts from anyone. And it could possibly be used for other purposes such as collecting info on people and industrial espionage. This is a huge problem if anyone inputs any information that is sensitive, confidential, proprietary, or personal. DeepSeek’s privacy policy clearly states that it keeps everything, and it keeps it in China. It also automatically collects data from users, who must sign on with an account. So from that perspective, it is not safe to use.

Third parties have hosted DeepDeekR1 on servers in the United States, stating that data stays on those servers. While that helps, questions remain about what is done with the input, and whether the code contains anything that might phone home.

Controlling chip exports

The United States imposed export controls on AI chips to China a while back to try to slow Chinese advances in AI. The question is whether those export controls had the intended effect, or whether they just forced the Chinese to innovate around it.

OpenAI says DeepSeek may have “inappropriately” used OpenAI data for training through a process called distillation. That’s rather ironic given the number of claims, including ongoing lawsuits, that OpenAI and others have trained their models using material scraped off the web without permission.

Expect businesses and governments to ban the use of DeepSeek. It is no longer available on app stores in Italy after Italian authorities questioned how it uses people’s info.

DeepSeekR1 is available on app stores and on the DeepSeek website. Anyone can try it if they sign up — but be very careful what you do with it.

David Canton is a business lawyer and trademark agent at Harrison Pensa with a practice focusing on technology, privacy law, technology companies and intellectual property. Connect with David on LinkedIn, Bluesky, and Twitter.

Image credit: ©Mojahid Mottakin – stock.adobe.com