A new foreign exchange student—DeepSeek v2--appeared in the AI classroom late last year. And its newer, smarter, faster, younger sister—v3--just entered the scene, taking a lot of people by surprise and shaking up more than just the AI industry. DeepSeek is an AI model created by a Chinese artificial intelligence company of the same name. DeepSeek has recently gained rapid and significant attention for its advanced AI models, particularly in natural language processing and reasoning tasks. Founded in 2023 and backed by the hedge fund High-Flyer, DeepSeek has developed several notable models:

DeepSeek-V3—Comparable to ChatGPT 4o: Is an open-source model with 671 billion parameters, trained on 14.8 trillion tokens. It excels in various natural language and coding tasks, achieving top-tier performance among open-source models. Parameters are generally considered to indicate the level of complexity of a model, and experts have questioned whether V3 really has that many parameters. While 671 billion parameters sounds impressive, it likely lags behind powerhouses like ChatGPT 4o. Although neither OpenAI nor Anthropic have officially disclosed the number of parameters in their models, educated estimates suggest that GPT4o may have around 1.8 trillion parameters.

DeepSeek-R1—Comparable to ChatGPT o1: Is designed for complex problem-solving. This open-source model has been praised for its reasoning capabilities, rivaling OpenAI's models like ChatGPT’s o1 in benchmarks such as the American Invitational Mathematics Examination (AIME) and MATH, according to tech magazine WIRED.

BackGround of DeepSeek AI Models

Open-Source LLMs

Open-source large language models (LLMs) like DeepSeek are AI systems trained on massive datasets to generate human-like text. Unlike proprietary models like Claude, ChatGPT, and every legal-specific LLMs, open-source LLMs make their models freely available for anyone to use, modify, or build upon. Open-source LLMs also provide a greater level of transparency, enabling users to understand how they work and tailor them to specific needs, such as legal research or drafting. This flexibility promotes innovation and reduces reliance on costly commercial tools. However, open-source models also pose risks, such as potential misuse or unvetted data influencing their outputs, requiring a great deal of programming necessary before they are used in legal practice.

Work Around for Less Powerful Chips

What’s interesting about DeepSeek is that in creating R1 and V3 it accomplished essentially the same results as OpenAI and Anthropic, but in a much shorter time, and with far less expense, and using less capable chips. Ok, taking into account the fact that DeepSeek almost certainly used a distillation method to quickly train its model (more on that in the next section), perhaps the real surprise is that DeepSeek was able to use slower, less-capable chips to run its model, while remaining competitive against models like OpenAI and Anthropic which both use the more capable chips. While the big U.S. companies are using powerful H100 chips, Chinese companies can’t legally purchase those chips thanks to Biden’s 2022 and 2023 bans. DeepSeek’s workaround for that is to use less powerful, slower H800 chips in their computers, but it uses them in a more specialized way.

This more specialized architectural structure is called a Mixture of Experts (MoE) architecture. This MoE style essentially passes the question you ask (your input) to the expert or range of experts best suited to answer your question effectively. Imagine you are writing a motion for summary judgment regarding a real property dispute, and you want help from your computerized assistant. You could ask an assistant who is great at everything and wait while the assistant shuffles through car manufacturing information, child-rearing expertise, and the like, until it returns information on your legal and real property questions. Alternatively, you could ask a group of experts the same question. Only the experts on law and real property would search through their information to give you the same or similar information as the one expert. Thus, you avoided needing the faster, more powerful expert assistant because it had more material to wade through than the two slower, less knowledgeable assistants. This is essentially the way the MoE architecture works. Because it takes up less computing power to spin up a couple of H800 chips than it does to spin up a couple of H100 chips, DeepSeek models use less computing power, and therefore less energy.

The Distillation Effect

DeepSeek-V3 was developed at a cost of approximately $5.5 million, significantly less than many Western counterparts. Remember how I said DeepSeek accomplished essentially the same model capability but faster and cheaper than OpenAI and Anthropic? That is almost certainly due to the benefit of distilling knowledge from already existing models. There are several ways that companies—not just DeepSeek—can benefit from the expense and time that went into creating powerful, usable models.

First, DeepSeek might have started with an open-source (freely available) model like Llama, which is free for anyone to download from a site like HuggingFace. The challenger model can then use API’s or chat clients, to shortcut the process of further training their own model. An API is an application programming interface. API’s are like the international translators for computer programs—they allow one computer program to talk to another. Using these API translators, or the chat clients, companies could ask questions of the parent model—for example ChatGPT 4o—to train and program their challenger models. Starting from scratch is almost always more difficult and time-consuming, and therefore more expensive, than getting a shortcut by copying someone else. Here is an excellent article that explains this concept in more detail. Not surprisingly, OpenAI has claimed that DeepSeek wrongfully used their model's data to shortcut training DeepSeek. We can appreciate the irony of this allegation since many authors have sued OpenAI for alleged copyright infringement.

Should We Care?

Another AI tool has arrived on the scene; should we care? The innovation of DeepSeek has several real-world implications. First, it has shown us that China can compete in the AI realm. DeepSeek's V3 and R1 models have been recognized for their efficiency, achieving high performance with relatively lower computational costs. But is there reason for concern about a Chinese company like DeepSeek developing a highly proficient AI model? Here are some key areas of concern and their implications:

1. Privacy and Data Security Issues:

Pass or Fail? ❌ Fail

There may be reason for apprehension about how these AI systems manage and store user data, especially if the tools are deployed internationally. There may be reason to worry about the potential for surveillance or misuse of data by the Chinese government, given its history of closely overseeing domestic companies. This is a valid concern considering the Typhoon Salt Hack that the U.S. government only recently uncovered. For more information on how Chinese government-backed hackers gained access to U.S. government and healthcare information, check out this article.

Is the information you put into DeepSeek safe? When I asked DeepSeek if it used the information I put into a chat, this was the response I got:

On the Terms of Service page it clearly indicated that yes, it does use my personal information and distributes it to third parties.

Further, Chinese AI companies may face government-imposed restrictions or directives that could influence how their models are trained, used, or modified. This raises concerns about bias, censorship, or the promotion of state narratives, as discussed in the next section.

Note that this privacy concern does have a mitigating factor. Because DeepSeek is an open-source LLM, you can download it from HuggingFace or another site, and customize or personalize it. Essentially, you are using it as a base to build an AI tool on top of. If you are self-hosting or deploying DeepSeek (running it locally from your own computers) it is far more secure.

Bottom line: The information you put into DeepSeek is not secure. If you have downloaded a copy and are self-hosting it, it has far greater privacy.

2. Copyright and Trademark Issues

Pass or Fail? ✅ Pass-ish

One of OpenAI’s biggest stressors since coming into the spotlight has been copyright and trademark infringement concerns. Here is an article that explains a number of key copyright issues more deeply. But putting that aside, there was a time where you could ask an AI tool for a copyrighted article—say from the New York Times—and it would allegedly spit out the article word-for-word. AI tools have not completely cleared their copyright and trademark problems, but they have come a long way. When asked to write in the style of Maggie Haberman, a well-known New York Times writer, ChatGPT was very careful to answer the question concisely and blandly, and then walk away. The ChatGPT output is here.

Claude was similarly unwilling to play the game that used to work in getting some copyrighted material. Instead, its cautious reply was to refuse to replicate Haberman’s writing and instead give a list of characteristics of her writing style. Claude’s output is here.

DeepSeek, not yet having been burned by this particular hot-button issue, was less cautious, because it did provide actual text of an article, but it looked like a first year college student who was copying a paper written by his best friend, Maggie Haberman. Despite spitting out an article, DeepSeek likely still did not violate any copyrights. DeepSeek’s verbose response is here.

Bottom line: More questionable, but there are guardrails for copyright protections in place.

3. Bias and Censorship

Pass or Fail? ❌ Fail

Chinese-developed models often include censorship mechanisms to comply with the Chinese government’s strict regulations on sensitive topics. Some of DeepSeek's models incorporate censorship mechanisms, particularly concerning topics sensitive to the Chinese government. For example, when I asked about Tiananmen Square protests, this was the response I got:

This lackluster attempt to pivot the conversation in another direction was in stark contrast to ChatGPT’s page-long essay on the events leading up to the Tiananmen Square protests and the aftermath.

Whether DeepSeek programmed out specific topics, or accomplished the censorship through training and weights and biases, DeepSeek steers well clear of topics controversial in China. The output of DeepSeek’s models might very well be under the control of the Chinese government. The bottom line is that we don’t even know or understand all the limitations of the DeepSeek model yet. But this model is clearly more limited in its information. This results in a less accurate and less robust model. Such limitations make DeepSeek’s models less attractive. Responsible users should keep this in mind.

Bottom line: Judiciously review all DeepSeek output, as you should with any AI Tool, and be aware of potential bias.

4. Innovation and Collaboration

Pass or Fail? ✅ Pass

On a more positive note, companies like DeepSeek advancing the field of AI can spur global innovation. Open-sourcing their models could provide valuable tools for researchers and developers worldwide, fostering collaboration. However, openness might also enable adversarial uses of the technology, such as misinformation campaigns, advanced surveillance systems, or cyberattacks.

DeepSeek's rapid advancements have positioned it as a significant player in the global AI landscape, challenging existing models and prompting discussions about the future of AI development and international competition.

Bottom line: In addition to causing the share price of chip maker NVIDIA to drop by around 20%, the emergence of DeepSeek will likely keep GenAI tools costs to consumers down.

5. Regulatory and Ethical Gaps

Pass or Fail? ⚠️ Use with Extreme Caution

Internationally, AI regulations and ethical guidelines are inconsistent. A company operating under a regulatory framework that prioritizes state control and rapid deployment might not align with more stringent privacy and ethical standards in places like the EU or U.S. That being said, however, in my initial queries to DeepSeek, it was mindful of staying in compliance with regulatory frameworks like the GDPR and the EU’s AI Act.

Bottom line: Verify compliance with all applicable regulatory frameworks. Be especially mindful of GDPR Article 45.

6. Knowledge of U.S., European, and Non-Chinese Legal Systems:

Pass or Fail? 🤔 Meh

Does DeepSeek have the knowledge of U.S. and European legal systems that lawyers need? It seems to have some knowledge of U.S. laws and legal structure; most likely it was trained on a great deal of English text and western legal system documentation. But it lags behind legal-specific AI tools, and even non-legal specific AI tools. For example, here is a very basic legal question given to DeepSeek, and ChatGPT:

PROMPT: I was told that I should challenge jurisdiction in my misdemeanor case. I refused to provide ID to a police officer in a traffic stop and was arrested. Walk me through the approach to this and if it is a good defense.

DeepSeek’s Response ChatGPT’s Response

In this comparison, I think most lawyers would agree that ChatGPT had the better response. It discussed both personal and subject matter jurisdiction both for definitions and in the context of the hypothetical, talked about a Motion to Dismiss (albeit in the context of having a lawyer), and advised of potential consequences. But DeepSeek made a respectable showing.

Bottom line: It lags behind private non-legal-specific AI tools, like ChatGPT and Claude, but it is not too far behind.

7. Other Notable Limitations:

DeepSeek is subject to many of the same limitations as ChatGPT and other LLMs. Remember the article I wrote about OpenAI’s o1 model? Model o1 could count the number of R’s in strawberry, whereas ChatGPT 4o at the time could not. OpenAI fixed the defect in 4o shortly after I wrote that article. Can DeepSeek pass the Strawberry test? Yes, but likely only because it was “hard-coded” into the knowledge base. I say that because DeepSeek could not accurately complete a very similar task—counting the number of I’s in “jurisdiction.”

No, DeepSeek, the answer is not 2. ChatGPT 4o got the answer to the question of how many I’s correct the first time. The reason this matters is that it shows a lack of reasoning ability, and potentially a limitation on the MoE approach. DeepSeek probably doesn’t have a “letter counting expert” in its arsenal. It’s also a good reminder that humans must always check all AI model output that they intend to use or rely on.

If you are planning on using DeepSeek for anything other than some non-personal entertainment or a very basic starting draft, follow the Best Practices, below.

Conclusion: Use Extreme Caution when Using DeepSeek in Legal Practice

DeepSeek represents both an innovation in AI technology and a cautionary tale for legal professionals. While its technical achievements are impressive—particularly its efficient architecture and reasoning capabilities—the decision to use DeepSeek in legal practice requires careful consideration of several key factors.

For Legal Research and Non-Confidential Tasks:

🔹 DeepSeek can serve as a free supplementary research tool for general legal concepts

🔹 Its open-source nature allows for transparency in how it processes information

🔹 The model's efficiency may provide cost advantages over other AI tools--use it for more basic use cases

📌 Perhaps the most valuable use of DeepSeek in your law firm would be to download a copy, fine-tune it, and self-host it as a secure AI tool. Watch for a follow-up article as I test this out.

However, significant risks persist. The future of AI in legal practice will likely include multiple tools, each serving specific purposes. DeepSeek may find its place in this ecosystem, particularly for preliminary research or general legal education. However, given its current limitations and risks do not rely upon it as a primary tool for sensitive legal work. As with any technology, the key is understanding its capabilities and constraints, and using it judiciously within appropriate ethical and professional boundaries.

Regardless of how you use DeepSeek, remember that AI tools like DeepSeek are supplements to, not substitutes for, professional judgment. The responsibility for protecting client interests and maintaining professional standards remains squarely with the attorney, requiring careful evaluation of any technology before incorporation into legal practice.

© 2025 Amy Swaner. All Rights Reserved. May use with attribution and link to article.