by Fabio Giacopello
January 20, 2025. While the US celebrate the new President Donald Trump inauguration, DeepSeek, a Chinese AI startup released its R1 advanced reasoning model (AI) as a direct competitor to OpenAI’s ChatGPT model.
DeepSeek has since climbed app stores becoming the most downloaded app in U.S. Not only, it created a crash into US stock market and all the business and tech community worldwide is talking about it extensively.
We then also joined this crowd of commenters trying to shade some light on the IP perspective of such phenomenon. However, before going into the IP topic, let us see what DeepSeek is and why it got so famous.
DeepSeek is an AI start-up created by Liang Wenfeng, a 40 years old Chinese tech entrepreneur, partner of a Chinese hedge fund known as “High-Flyer”active in the field of quantitative fast trading of listed stocks using AI.
DeepSeek first model was released on November 2, 2023and then the two most recent releases, V3, a general large language model (“LLM”), and R1, a “reasoning” model got the glory of the worldwide news.
The release of DeepSeek was surprising for several reasons.
Breakthrough Performance at Low Cost
DeepSeek’s new AI models deliver performance that equal rivals or even surpasses established western industry leaders. But not only this, it does it at a fraction of the cost. The company claimed that it was built at a fraction of the cost of industry-leading models like OpenAI and also it seems that it uses zero (or fewer) advanced chips. This dramatic efficiency shift caught many by surprise, as it upended the long-held assumption that high-quality AI requires massive investments.
Market Disruption
The timing of the release, coinciding with broader market events, sent shockwaves through the tech industry. Investors reacted strongly (chip-making giant Nvidia shed almost $600bn of its market value on Monday - the biggest one-day loss in US history) as the market suddenly had to re-evaluate the future demand for high-cost AI hardware in light of DeepSeek’s cost-effective solutions.
Open Source and Innovative Models
DeepSeek has open-sourced its models and made available technical reports to explain them.Also, it incorporated a number of optimizations for training. We mention only a few:
(1) improvements in Data Processing – DeepSeek V3 utilizes the technique named Multi-Head Latent Attention (MLA), a very efficient tool to process large amounts of data during inference, which result in using only half of memory than comparable techniques.
(2) Improvements to Model Architecture – DeepSeek V3 introduces a unique architecture for Mixture-of-Experts (“MoE”) models which allows that only a portion of the entire model needs to be calculated for any given token, and thus a substantial amount of computation can be avoided for each forward pass of the model.
(3) Improvements in Model Training – DeepSeek V3 uses a multi-token training objective so that that the model can “pre-plan” a few tokens in advance, making its prediction of the next token more robust.
Let's now go to check if there are IP issue in the launch of Deepseek.
In general, there are certain typical infringements in the field of AI which are the following:
Possible Copyright Infringement: (a) Training Data: if an LLM is trained on copyrighted content without authorization, it could be infringing on the copyright of those works. (b) Generated Content: if the model generates content that reflects directly the copyrighted works or is too close to them, it could also represent a copyright infringement.
Possible Patent Infringement: Model Architecture or Algorithms: these can be protected with patents and thus if they are the LLM might infringe on existing patents (for example, patented methods for machine learning or natural language processing).
Trade Secret Misappropriation: Use of leaked of Proprietary Information: for example, if the LLM is trained on data that are considered and protected as trade secrets of a specific company.
Model Usage Violations: Unlicensed Deployment: usage of a model in a way that violates the licensing terms under which it was developed (e.g., commercial application where the license prohibits that).
Given what above, it is too early to say if there is an IP issue. However, it is possible to ask the most interested party, which is ChatGPT itself. So, we asked if it believes that DeepSeek is infringing upon any IPR of OpenAI. Initially, ChatGPT was a bit reluctant to reply and pretended not to know its competitor but the later on the conversation was smoother.
This is the reply from ChatGPT.
”Ah, got it! You’re referring to DeepSeek, the Chinese AI that’s been making waves recently. From what I know, DeepSeek is another large language model similar to others like GPT-4 or even ChatGPT, but developed by Chinese companies or research institutions. While there’s no clear evidence that they are directly copying me or OpenAI’s models, there are a few things to consider:
1. Similar Technology: Most advanced language models these days, whether from OpenAI, Google, or Chinese companies, are built on similar underlying technologies—like transformer architectures. So, while the general approach may be similar, that doesn’t mean it’s necessarily “copying” in a legal sense.
2. Training Data: If DeepSeek uses a large dataset that contains copyrighted or proprietary content without proper licensing, that could potentially raise legal issues similar to what we discussed earlier about IP infringement. Each company tends to handle data licensing differently, but any overlap with OpenAI’s proprietary training data could lead to legal concerns if it’s not appropriately managed.
3. Independent Development: It’s likely that DeepSeek, while similar in many respects, has been developed independently with its own proprietary methods and datasets. It’s common for companies, even in different countries, to develop their own versions of AI technology based on public research and available data, so unless there’s evidence of a direct copy of specific code or data, it’s more likely they are building on the same general principles.
If there are concerns about DeepSeek copying specific aspects of OpenAI’s technology, it would likely fall into a gray area of IP law, depending on the exact practices in training, usage, and distribution. The specifics would come down to things like licensing agreements, patents, and whether any proprietary OpenAI data or methods were improperly used”.
It seems ChatGPT – at this stage – has no special complaints on IP infringement coming from DeepSeek.
Meanwhile, it seems that someone squatted the trademark “DeepSeek” in the US and that some concern on personal data treatment sparkled in Europe.
We will keep monitoring the situation and updated you if any interesting news arises.