Deepseek or Chat GPT? The challenge (also) affects Intellectual Property

Recently, there have been rumours that the much-talked-about AI software, DeepSeek, used Open AI training models to more easily achieve the same results as Chat GPT. Can it be true? Let us analyse the situation with a look at intellectual property.

DeepSeek recently shook up the world of artificial intelligence-based chatbots as it overtook its rival ChatGPT, from US-based OpenAI, in the list of free downloads in digital stores, both in China and in the US.

In particular, it is surprising that DeepSeek, which can be traced back to the Chinese start-up High-Flyer, has apparently achieved similar computation performance as its American rival, but using chips with lower computation capacities and a significantly lower investment. Of course, these data still need to be verified.

Deepseek and Chat GPT: the comparison in numbers

Let’s look at a specific comparison of the main data:

Feature	Deep Seek	Chat GPT
Chips used	2000 Nvidia H800 chips	30000 Nvidia H100 chips
Expenditure	About 5 m USD	Over 100 m USD
Training time	Approximately 55 days	Approximately 35 days

In this comparison, it should be noted that the Nvidia H800 chips have lower performance than the Nvidia H100. Among other things, Deepseek had to overcome US measures pertaining to export of IT to the East, precisely to prevent China from having the same computing power as America.

This notwithstanding, with this new AI chatbot  Deepseek was claimed to have caught up with (and perhaps overtake?)  Chat GPT.

All this is happening in the political context in which the newly installed US President Donald Trump has announced 500 billion USD in investments in companies developing AI tools. Therefore, the topic appears to be very important.

Copyright and software protection: the Intellectual Property battle

Now let’s look at some intellectual property issues.

It seems that DeepSeek has used a technique called “distillation” with respect to OpenAI models; a process in which more powerful models are used to improve the efficiency of “smaller” software programs. Although this method is common in the industry, OpenAI believes that Deep Seek violated the terms of its platform by creating its own competitive AI.

However, to  allege “intellectual property theft” by DeepSeek against OpenAI may be misleading as it would be necessary to understand whether  Deep Seek also copied or reproduced part of the code.

As a matter of fact, from a copyright point of view, it should be noted that software protection is mainly confined to the expressive form of the software (typically the source code) or in any case to the total or partial reproduction of the program by any means.

If, on the contrary, DeepSeek successfully proved to have made new code (albeit inspired by the other), the question of mere copying of the software could be dropped.

What is more relevant, if proven, would be the violation of the terms of use of OpenAI. In fact, at the “contractual” level, it is clearly stated in the OpenAI Terms and Conditions of Use themselves that it is forbidden to use the output to develop competing models, to automatically or programmatically extract data or output, and the so-called practice of reverse engineering of its systems.

Therefore, if OpenAI were to prove that DeepSeek used its training models (which remains very difficult to prove), this would constitute a breach of contract situation. It all remains to be seen.

In any case, it is to be expected that OpenAI will certainly not sit idly by and legal battles are expected in the near future.

DeepSeek, a trademark already registered in the USA

In addition, there also seems to be trouble in the US on the DeepSeek trademark registration front.

A few weeks ago, DeepSeek filed an application with the United States Patent and Trademark Office (USPTO) to register a trademark for its AI applications, products and tools. But the application came a little too late.

In fact, it seems that shortly before DeepSeek’s application, another US company had already applied for the ‘DeepSeek’ trade mark: a Delaware-based company called “Delson Group Inc.”.

Delson Group claims to have been selling AI products under the DeepSeek brand since early 2020. The CEO of Delson Group, by the way, is of Chinese origin and his name is Willie Lu. Lu seems to have spent most of his career in the telecommunications industry, so it may be true that Delson has been using the brand since before the inception of DeepSeek chat.

Conclusions

In any case, DeepSeek’s options at this point are rather limited. Under US law, in fact, the first user of a trademark is generally considered the rightful owner, unless it can be proven that the trademark was registered in bad faith.

Of course, DeepSeek could always try to obtain a coexistence agreement if it could prove that it operates in AI fields other than those in which Delson Group operates.

Meanwhile, in Italy, the DeepSeek application has been blocked. In fact, the Italian Data Protection Authority announced that it had “ordered, as a matter of urgency and with immediate effect, the restriction of the processing of Italian users’ data with regard to Hangzhou DeepSeek Artificial Intelligence and Beijing DeepSeek Artificial Intelligence, the Chinese companies that provide the chatbot service”. Put another way, it means that these companies could violate the  EU General Data Protection Regulation n. 2016/679  (GDPR) and were therefore prevented from doing so.

In short, DeepSeek will certainly not have an easy life outside China’s borders. And this is only the beginning…