Recently, Li Yizhou, who is nicknamed the "Godfather of Chinese AI", was revealed to have "transported" models from China's well-known AI models and AI creation sharing community "LiLibai" and used them for commercialization.
At present, there have been public reports that liblibai said that it has resorted to legal means.
The report also emphasized that the LiLibai team has consulted with relevant IP lawyers and believes that the issue of model property rights is even more complex than that of Wensheng Tu.
So what is the complexity of the property rights of AI models? What are the legal liabilities that may be involved in "transporting" AI models from other platforms without authorization?
oneLilibbai's rights protection has been launched to protect the AI model
Starting in 2023, with the popularity of AI, the legal community is also hotly discussing AI-related legal issues.
At that time, there was more discussion about the legal issues of the training data learned by AI models; legal characterization and infringement issues of content generated by AI models; AI-generated false information and other illegal and criminal acts, but there is little discussion about what the AI model itself is and how it can be protected.
One of the important reasons for this is that a large number of AI models in the industry are open source, and few people will pursue their legal responsibilities, or they are completely closed-source and cannot be obtained by third parties.
Therefore, the "liblibai" investigation of Li Yizhou's "handling AI model" incident will also make the industry begin to pay more attention to the legal protection of AI models.
If you want to protect an AI model, you first need to be clear about what an AI model is.
According to the inquiry to Microsoft's Copilot model, AI model refers to a mathematical model that uses methods in the fields of mathematics, statistics, computer science, and machine learning to analyze, process, and optimize data with certain regularity and appreciability. To put it simply, an AI model is a mathematical model that transforms "data" into "intelligence".
In the field of artificial intelligence, models are typically made up of two key components:
Parameters:These are similar to knobs that record the state of the model. The values of the parameters are constantly adjusted during training to accommodate different data.
Architecture:The schema defines how the parameters are connected and the hierarchy between them. Different architectures will have different effects, even if the number and value of parameters are the same.
Therefore, to put it simply, the AI model is a mathematical model, including parameters and architecture, and the output content is adjusted through parameters using the built architecture.
II. II. IIWhat legal rights do the AI models of the open source community have?
What legal rights will AI models involve in the protection of mathematical models of parameters and architectures?
Based on the analysis of AI models from the above-mentioned technical perspectives, under the law, the legal rights that the developers and publishers of AI models and the AI model community may have may involve:
Patents
As mentioned above, AI models are mainly mathematical models of parameters and architectures, and according to Article 25 of the Patent Law, the rules and methods of intellectual activity are not patentable.
However, in 2019, the Patent Examination Guidelines were revised and formulated for artificial intelligenceExamination rules for "Provisions Relating to the Examination of Invention Patent Applications Containing Features of Algorithms or Business Rules and Methods".
According to the new examination rules,Especially the parametric part in the AI modelIf it can solve new technical problems and achieve new technical effects (whether it is a basic model or an optimized model), it can meet the requirements of inventive step examination under the Patent Law and may be protected by the Patent LawAI model developers can apply for registered patents to protect their intellectual property
However, in February 2024, when SORA has been born, if the AI model is trained by another large AI model or the relevant parameters are obtained through automatic training, it needs to be considered that the AI cannot be the subject of the patent and thus cannot be granted a patent.
Software Copyright
The architecture part of the AI model is usually embodied in the form of software, which is directly executed by a computer or other device with information processing capabilities, or a symbolic instruction sequence or symbolic statement sequence that can be automatically converted into a ** instruction sequence, especially the basic model. For example, common neural networks, transformers, and convolutional neural networks will all perform operations. For the part embodied in **, it can usually be protected as a computer software copyright in law.
However, at present, AI models similar to "liblibai" are usually based on third-party basic models for further optimization (for example, "liblibai" is mostly based on various fine-tuning models of the stable diffusion underlying model, including LoRa, Control Net, etc.), and the optimization model usually does not adjust the ** of the basic model, but adjusts the parameters in the basic model. It is often difficult for AI models published on liblibai to own software copyrights for the underlying model architecture, and it is necessary to evaluate whether they can be protected by patent rights.
Trade Secrets
Although a large number of AI models, especially those released by the AI model community, are open-source models, there are still many models that are not open-source or otherwise made public. If the developer has taken technical protection measures, it can be protected through trade secrets.
Violation of open source licenses
At present, most of the models published on AI model communities (such as "Liblibai" or foreign GitHub) are open source models, and publishers can choose different open source licenses (such as Apache, BSD, GPL, etc.) when publishing. Under various open source licenses, users need to use the open source AI models according to the requirements of the open source licenses.
In particular, it is necessary to pay attention to whether the relevant open source agreement allows commercial use, if commercial use is not allowed, the use of the relevant AI model for commercial use is a violation of the open source agreement, and the AI model publisher can pursue its legal liability for breach of contract.
At the same time, many AI model communities (such as "LiBlibai") will clearly state in the user agreement that "with your authorization, we have the right to defend the rights of the content that infringes the intellectual property rights uploaded and published by you in our own name or by entrusting a professional third party when necessary, including but not limited to: monitoring infringement, sending rights protection letters, filing lawsuits or arbitrations, etc." ”
Therefore, for the infringement and breach of contract of the AI model, the AI model community also has the possibility of claiming infringement.
Data-based benefits
There is also the possibility of violating the Anti-Unfair Competition Law for the unauthorized transfer of a large amount of third-party AI model community data.
According to the Anti-Unfair Competition Law, if the unfair removal behavior substantially replaces the services provided by the moved party, affects the normal operation of the moved party, and illegally uses the business results of others, destroying the market competitive advantage of the moved party, it will also be constitutedInfringe on the data rights and interests of the AI model community being moved
ThreeConclusion
Therefore, for the unauthorized transfer of AI models, AI model creators and publishers can protect their rights through patent rights, software copyrights, open source agreements, trade secrets, etc., while AI model communities can protect their economic interests through data rights and interests under the Anti-Unfair Competition Law.
Made|Who|Introduction|Shao Shi Xiao Nan is a special expert of Internet Law Review.
Partner at Merits & Tree Law Offices.
Practice Areas: Intellectual Property, Dispute Resolution, Regulatory & Compliance.
He has been engaged in corporate compliance and intellectual property legal affairs for more than ten years, focusing on solving Internet compliance, data compliance, intellectual property compliance and other types of cases. Provide compliance solutions for a large number of telecommunications companies, Internet companies, information technology companies, and metaverse enterprises to solve difficult and complex litigation cases. He has worked in a large and well-known Internet company for many years, and is good at designing compliance solutions with business objectives as the core. Master of Laws, Hitotsubashi University, Japan.
DisclaimerThis article only represents the author's personal views and has nothing to do with this platform. The platform is neutral to the statements and opinions in the text, and does not provide any express or implied guarantee for the accuracy, completeness or reliability of the content contained therein.