AI has been the subject of science fiction for some time. The science fiction of yesterday is the reality of today. Virtual agents, sapient algorithms, and robots have invaded our lives. Today, AI can write news and novels, generate artwork, and write and perform music. AI and machine learning challenge the traditional legal notions of intellectual property (IP), such as “copy,” “originality,” “creator,” “author,” or “inventiveness.”
Can an AI be an author? Can AI be an inventor? Can AI co-author a work with humans? Who is going to own works and inventions generated by AI? Should AI’s inventions be considered prior art? Who owns the dataset from which an AI learns? Who is going to be liable for AI-generated works and inventions if they encroach upon others’ rights?
The regulation of AI’s activities is set to become a primary policy and governance issue. In particular, AI’s disruptive effects on traditional business and governance models will force a reconsideration of the existing IP policy and governance framework. This article explores legal, policy, and ethical issues concerning AI’s impact on creativity and innovation.
AI and big data will impact IP systems across several fronts.
1. AI as a creator and inventor
Unlike previous technologies, AI can easily replicate labor activities at a much greater scale and speed and even perform some tasks beyond human capabilities. Currently, most jurisdictions will protect the expression of the algorithm and the software through copyright and/or patent law. However, things get more complicated where the AI continues to “learn” and make changes to its own software structure without any human involvement. How should regulators respond in this regard: should they recognize copyright in works created by AI and the ownership of these works. Who will own the IP rights in inventions/works of authorship that are created by AI and machine learning?
AI and copyright are intertwined in a variety of ways. AI systems learn from data that can be protected by copyright and database rights. For instance, an AI system that composes music can be trained using musical works. These musical works are protected by copyright, which may be infringed. Some AI systems can create copyrightable works without any human intervention. This creates challenges to the traditional legal concept of human authorship. For example, the Next Rembrandt Project trained AI to develop a new painting in Rembrandt’s style and used data from 346 of Rembrandt’s works to do so. Google’s AI Quick, Draw!, where AI attempts to identify doodles drawn by users, has been used more than 50 million times. The AI improves as it learns from each submitted doodle.
The example of Philyra: an AI system that creates fragrances by sifting through data and combining chemical formulae
The ability to craft a fragrance is something that takes master perfumers years of experience to develop. A group of IBM researchers and skilled perfumers at Symrise, a global producer of flavors and fragrances, got together to explore how to use AI to do just that. Mixing artistic and scientific thought into a big pot resulted in the creation of Philyra — an AI product composition system that can learn about formulas, raw materials, historical success data, and industry trends. Philyra uses new, advanced machine learning algorithms to sift through hundreds of thousands of formulas and thousands of raw materials, identifying patterns and novel combinations. As Philyra explores the entire landscape of fragrance combinations, it can detect gaps in the global fragrance market for which entirely new fragrance formulas can be designed. The first two fragrances to be produced using Philyra will be for Brazilian cosmetics company O Boticário.
AI presents several novel challenges to traditional patent law. For instance, the standard of the “person skilled in the art” used to evaluate inventive step and serve as a measure of the patentability of an invention is challenged by the role of AI in inventing new patentable products and methods. Most countries require patent applications to disclose a natural person as an inventor. This legal requirement is designed to protect and acknowledge the rights of human inventors. Yet, inventors do not necessarily own their patents; in fact, most patents are owned by businesses. Ownership rights can pass from an individual to a company by contractual assignment or otherwise under law. For example, in many jurisdictions, ownership passes automatically to an employer if an invention is created within the scope of employment (work for hire). Even when an inventor does not own a patent, laws requiring a natural person to be listed as an inventor ensure that people receive due credit. However, these laws were created without regard to the future possibility of inventive activity by machines. Recently, the United Kingdom Patent Office (UKIPO), the European Patent Office (EPO), and the USPTO have stated that inventors must be human and do not allow AI tools to be named as an inventor. For instance, the applicant for two UK patents had asserted that the inventor was the AI machine named DABUS. This raised the issues of patent inventorship and ownership. The hearing decision concluded that a non-human inventor cannot be regarded as an inventor under the Act. This decision has been appealed to the High Court. The EPO has similarly rejected patent applications from the same applicant where the AI system is named as the inventor.
Some would argue that the approach taken by the EPO, UKIPO, and the USPTO is wrong: “allowing people to take credit for work they have not done would devalue human inventorship. It would put the work of someone who merely asks an AI to solve a problem on an equal footing with someone who is legitimately inventing something new”. On the other hand, some have said that current AI systems are not “inventors” since they are no different from other tools used by people to invent, such as combinatorial chemistry used to produce novel drug compounds.
2. IP licensing and assignment
Under normal circumstances, in a typical IP license agreement, the licensor will usually own the right to improvements. In the context of AI and machine learning, the software’s function is to improve its analytical ability by harvesting the data generated and provided by the licensee. It isn’t easy to disentangle the licensee’s data and its underlying know-how from improvements to the licensed AI system itself. Therefore, the common contractual practice that each party retains the rights to improvements to its IP may not apply. To deal effectively with these challenging issues, organizations should use the growing number of open source and Creative Commons licenses. Open source licenses often are used for making software freely available, while Creative Commons licenses often are used to make other copyrighted works and databases available on a no‑cost basis. Organizations should assess the various forms of these licenses and consider how they might be used on an in‑bound and out‑bound basis to further their business objectives.
3. Data ownership and protection
As AI collects more data, there will be disputes over data ownership and data protection. Training an AI system entails processing a large dataset used by the system to test and enhance its decision-making abilities. An important question in this regard is who owns the IP rights in the dataset. Under U.S. law, data is not copyrightable because “facts” are not considered original works of authorship. Limited copyright protection might be available for how the data is selected, coordinated, or arranged. Similarly, EU law affords copyright protection to databases that are “original” in the selection or arrangement of their contents. Europe also provides for sui generis database rights, which provide limited protection to databases if significant investments have been made to obtain, verify, or present their contents. Organizations can also rely on trade secret laws to protect data, as long as adequate measures are implemented to protect data confidentiality. AI will also open up novel challenges concerning “data licensing” — especially in health and biotechnology. Health/biotech data is sensitive and protected under privacy laws; this adds up another complexity layer. Healthtech and biotech companies will struggle to find ways of proper valuation of their datasets for licensing purposes. These companies will also need to make sure the data is properly de-identified and that they have adequate human consent for all the patients. Maintaining a proper data inventory and mapping the types of data controlled and processed by companies would become increasingly important for IP management and valuation of the companies’ intellectual assets.
Emerging technologies, such as AI and machine learning transcend boundaries. This implicates jurisdictional issues. In the context of IP rights management, an important issue would be the location where the AI and the data generated by the system will be “used.” This can be further complicated in the context of cloud-based computing. This can also have tax implications. Data localization laws should be taken into account as well.
The license agreement for the AI system should address liability for IP infringement “committed” by the AI system. The AI system cannot by itself infringe on a third party’s rights, but the person most closely connected with the infringement (e.g., the operator of the system) is likely to be held liable. For instance, under normal circumstances, when copyright is infringed, the copyright owner has the right to take action against an infringer. However, if an AI system infringes copyright, a natural person or legal entity must ultimately be legally responsible. This would be the person who has control over the infringement, the ability to stop future infringement, and who is able to compensate the copyright owner.