Can I patent my AI system?
Most AI can be considered software. In many jurisdictions, software cannot be protected by patents. To understand the nuance here, we need to break down the key terms in this area.
Software is a computer program that instructs hardware (i.e., the computer’s physical equipment) to perform specific tasks. A computer program is an algorithm, which is a series of instructions to perform a computation.
Algorithms are usually written in a programming language, like Java or C++ or Python, which is similar to natural language. This is called the source code. The source code is then usually translated to machine language (object code), which consists of binary values (0s and 1s) that provide operations a computer can execute directly.
The European Patent Office (EPO) does not consider a computer program to be patent eligible subject matter (and considers the term “software” to be ambiguous). Specifically, in its 2018 annual update to its “Guidelines for Examination,” the EPO notes that AI and machine learning are based on algorithms that are of an “abstract” (and so, unpatentable) nature and, irrespective of training, cannot be patented. However, the EPO may allow patents on computer-implemented inventions (CII), which functionally allows patents on some software, including AI applications.
A CII “involves the use of a computer, computer network or other programmable apparatus, where one or more features are realized wholly or partly by means of a computer program.” Eligible patent claims must have a technical character, which may involve claiming a computer-implemented method rather than an algorithm. Claims must also have a technical contribution to an inventive step.
For example, a method for having a computer generate random numbers by itself would not be patentable, but the EPO has found the same method patentable when used to simulate a circuit in a particular manner. The EPO guidelines provide other examples of patentable AI technical applications: use of a neural network in a heart-monitoring apparatus to identify irregular heartbeats, or the classification of data based on low-level features (e.g., edges or pixel attributes in images). AI may also be protectable where it is specifically adapted to run on a computer, where “the design of the mathematical method is motivated by technical considerations of the internal functioning of the computer.”
As is the case for all patents, a CII needs to meet general patentability criteria: the invention must be novel, provide an inventive step, a “technical solution to a technical problem,” and have industrial applicability.
The US also permits certain software patents. Before the AI era, the United States Supreme Court found specific software patents to be patent-ineligible “abstract ideas” implemented on a computer. Subsequently, lower-level courts have upheld a limited number of software patents, but there is no clear rule distinguishing the patentable from the unpatentable. The US Patent and Trademark Office guidance states that in order to refuse a CII as an abstract idea, patent examiners must factually prove that claims are “well-understood, routine, and conventional".
When successfully obtained, or sometimes even as pending applications, patents can offer significant value to businesses. That said, due to the relatively high cost of obtaining patents, protection is usually sought selectively in key jurisdictions. This means that patent holders can only prevent third parties from using their inventions in certain jurisdictions. Even for large enterprises, “worldwide” patent protection is usually both cost prohibitive and unnecessary. At the moment, considering the countries that are racing to become the global leader in AI, the most important jurisdictions are the US, France, Germany, the United Kingdom, Japan, Republic of Korea and China.
Can I use copyright to protect my AI model or system?
Copyright is a critical source of protection for software, which is protected as a “literary work” under Article 4 of the WIPO Copyright Treaty (WCT) and Article 10 of the Agreement on Trade-Related Aspects of Intellectual Property Rights (TRIPS). When it comes to machine learning and neural networks, copyright only protects the code. The model will be protected by a combination of copyright and patents law, and possibly by trade secrets.
Exactly what qualifies for copyright protection varies by jurisdiction. In the UK, preparatory design work qualifies for protection if a computer program can result from it at a later stage.
In the EU, the protection of software is regulated under the Software Directive, which establishes that protection subsists whenever the work is the “author’s own intellectual creation.”
Copyright protects expression and does not extend to ideas, procedures, methods of operations or mathematical concepts. This means copyright does not prevent competitors from creating their own code that can behave like the protected code. If a competitor can access protected code, they may be able to rewrite the code so that it no longer infringes but has the same functionality. Even without direct access to the code, it may be possible for a competitor to generate a similar program from observing how software behaves. If they write their own code, and it does not include the protected source or object code, there will generally be no infringement. In fact, a competitor’s code will then be protected by its own copyright.
Can I use trade secrets to protect my AI innovation?
Software may benefit from trade secret or confidential information protection. If a competitor cannot access the source code or otherwise determine how to recreate the software’s functionality, trade secret protection may be effective. For AI, this could be particularly effective if third parties do not have direct access to the AI system and only the AI’s output is shared outside an organization. For example, weights and parameters of an AI model could be protected as confidential information.
Trade secret protection requires constant vigilance since it relies on secrecy. To constitute a “trade secret” under IP (with the caveat that different jurisdictions have different rules):
the information may need to have economic value by virtue of the fact it is not generally known; and
the owner must take reasonable steps to keep the information secret.
This is one reason why non disclosure agreements (NDAs) are common and essential, particularly as part of contracts with employees and independent contractors. If a company fails to take reasonable steps to protect information (including technical measures) and discloses it without any confidentiality agreement in place, then that information generally does not obtain trade secret protection from the outset or loses that protection.
If a company takes reasonable steps to restrict disclosure of information that is lawfully in its control, and the information has value due to its confidentiality, then it may be possible to prevent that information from being disclosed, acquired or used without their consent in a manner contrary to honest commercial practices. In practice, this means that if the information is leaked through, for example, a hacker attack or a breach of contract, there may be legal remedies available to restrict use of the information or seek damages.
What is the role of open-source licensing in IP and AI?
Special considerations apply to the use of open-source software (OSS). OSS usually refers to source code that is publicly available, and that may be used by third parties without cost. However, while OSS is usually free of monetary cost, it generally comes with restrictions on how the code may be used.
Using OSS may require a license that prevents users from making modified versions of the source code or from using it for commercial purposes. OSS licenses might also restrict anyone incorporating the source code into other software unless they agree to make the entire program open source, effectively eliminating its proprietary status.
Thousands of different types of OSS licenses exist, many of which limit the IP protections available to any product incorporating OSS. This applies both to the use of code libraries and to snippets that may be incorporated into a broader codebase.
The discussion around open-source AI is still in its infancy. According to the open-source Initiative, an Open Source AI is a system made available under terms that grant users the freedoms to use, study, modify and share the system.
As the landscape currently stands, open-source AI exists on a spectrum from fully open to fully closed, depending on which components are publicly available under open-source licenses. Such components can include documentation, software code, training data, weighting factors and model architecture. The EU AI Act defines "free and open-source AI" as AI components made accessible under open-source licenses, though this definition has limitations and doesn't specify how many components must be open. This is important, because the AI Act provides lighter requirements for stakeholders releasing “open-source AI” (though this excludes high-risk systems and monetized services). Some critics have argued this creates incentives for "open-source washing" to avoid transparency obligations.
To summarize, open-source AI requires different concepts than traditional open-source software due to its complexity and multicomponent nature. Clearer definitions and legal frameworks for what constitutes “open-source AI” are still emerging. In the future, reliance on open-source frameworks could provide a key source of transparency around AI, provided that the entire open-source system is considered and open standards are adopted horizontally. Users should be aware that different AI OSS licenses may come with their own conditions and restrictions.
Can I use copyrighted materials to train an AI model?
As outlined in Chapter 1, recent rapid advancements in machine learning have been made possible thanks to the widespread availability of large quantities of digital data, often referred to as “input material”. This can be in various forms: from statistical information such as numbers, where no copyright concerns would arise, to text, image, video and voice recordings, where copyright issues have become a key point of contention and debate.
In some cases, input materials are collected as training material for an AI system through text and data mining (TDM), which is the highly automated processing of large amounts of data to reach new knowledge. (Again, think of the example from Chapter 1: tens of millions of purchase records from a retail chain are unintelligible to humans due to the sheer scale of the data, but machine learning can be used to accurately predict sales trends.)
Generally, a TDM operation involves the following stages:
Extracting information from accessible online sources using bots or other automated systems. This is often called “crawling” or “scraping.”
Formatting the collected data.
Scanning the input materials to find correlations, trends and patterns.
Processing and analyzing data.
Storing data for validation purposes.
The key point of contention here is that each of these stages may involve a different kind of reproduction, which is one of the exclusive economic rights reserved for copyright holders in all jurisdictions, and the principal right at stake in these processes.
That said, many jurisdictions see significant societal and economic value in TDM activities, and so legislators are seeking to balance copyright holders' rights and users' interests by allowing exceptions for TDM. As noted in Chapter 2, Japan was the first country in the world to explicitly exempt TDM from copyright liability in both commercial and non-commercial cases.
In the US, there are a few decided cases (and several pending) on whether “fair use” is a defense against copyright infringement where copyright-protected material has been used for training AI systems.
When deciding if “fair use” is at play, the four factors of section 107 of the US Copyright Act have to be considered:
the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes;
the nature of the copyrighted work;
the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
the effect of the use upon the potential market for or value of the copyrighted work.
In two recent decisions, Bartz v. Anthropic and Kadrey v. Meta, the courts held that there was fair use in the use of the copyright-protected material in training AI models. The decisions were based largely on the first factor, “purpose and character of use”, where the key issue is analyzing whether or not the use of the copyrighted material was “transformative.” In both these cases, the courts decided that AI training is indeed transformative, and so counts as “fair use.” In Kadrey v. Meta, the court also looked at the fourth factor: the impact on the market and the value of the copyrighted works. The conclusion was that because the AI system could not fully reproduce the copyrighted works or replace them in the market, the training was fair use.
However, it is worth noting that these decisions are specific to the facts of the cases and for now the US Copyright Office continues to recommend that fair use is applied case-by-case. Without further clarity, it is best to avoid using copyrighted content to train AI models, or obtain permission from the right holder(s) where necessary.
Other jurisdictions, such as Israel, have a similar, open-ended, fair use doctrine. Israel’s Ministry of Justice has been vocal in the AI/copyright intersection, stating openly that TDM training falls within either the fair use exception or the temporary copying exception.
At the EU level, TDM exceptions have been introduced through two mandatory exceptions in Articles 3 and 4 of the Directive on Copyright and Related Rights in the Digital Single Market Directive (the CDSM Directive): one addressing scientific purposes, and a second for the purposes of TDM, including AI training. The EU legal framework states that for the second exception, rights holders are allowed to place contractual restrictions limiting TDM through their Terms and Conditions or Terms of Service. All member states had to implement both exceptions by June 7, 2021 and would be held liable for lack of compliance if they failed to introduce them into their national copyright law.
In conclusion, there is no simple answer to the general question of whether copyrighted materials can be used to train AI models. Representatives of creative industries have engaged in extensive lobbying to defend their position and ensure that authors, artists, musicians, photographers, and other rights holders are remunerated for the use of their copyrighted works. They maintain that when TDM is used on copyrighted works to train AI models this could amount to copyright infringement, and that rights holders should be approached for licenses and able opt out altogether, if they wish.