IP protection of deep learning systems

On 15 June 2018, the Dutch Association for Copyright (VvA) organised a study session on the interaction between intellectual property and artificial intelligence. While this issue is often approached in terms of IP protection of artificial works and inventions, the first session considered the IP protection of deep learning systems themselves. Jean-Marc Deltorn of CEIPI (University of Strasbourg) gave a detailed presentation. I delivered a brief response, which contained a number of specific questions about Mr. Deltorn’s presentation, and also raised some broader issues about the extent to which further IP protection of deep learning systems is necessary and the impact of deep learning creations on the market for human-created works.

Both papers have been published in the Dutch copyright journal AMI, and a pdf of mine is available here. The English version follows below.


Artificial Intelligence (AI) has been the subject of debate in technological and legal circles for decades. Indeed, Timothy Butler’s article on Copyright Aspects of Artificial Intelligence in the Hastings Communications and Entertainment Law Journal dates from… 1982. Nonetheless, as Mr. Deltorn’s presentation shows, AI systems have recently been undergoing a significant growth spurt, measured in sophistication, scope and scale. This applies not so much to general artificial intelligence systems – a computer system that can learn to perform any human task – but rather to more narrow systems that learn a specific task, such as recognising a face or driving a car.

A machine learning system is trained, using example data, to develop rules that enable it to perform a task. The most advanced systems, based on neural networks, are often described as deep learning systems. Such systems are rapidly becoming more advanced, and have gained the potential to effect fundamental change to our society, culture, and eventually even our biology. Of all the technological hypes that lawyers – generally with some delay – get excited about, machine learning stands as a good a chance as any of proving to be genuinely transformational.

At the same time, we lawyers should be realistic about our role: the most urgent and essential questions will probably be fodder for engineers, politicians, and philosophers, with the lawyers once again being asked to clean up the mess afterwards. In the meantime, though, machine learning provides a useful lens to study the function and development of intellectual property law.

In the coming years, more and more deep learning systems will be deployed that can assist, replace or surpass humans in the performance of everyday tasks and creative endeavours. They will therefore inevitably encounter the same kinds of IP questions as humans do. Can something made by or using a deep learning system be protected by IP rights and, if so, who gets the rights? Can an deep learning system’s operation or output infringe IP rights and, if so, who or what is liable? Can a deep learning system invoke exceptions and limitations and, if so, whose use and expression rights should the law balance against the exclusive rights of the rightholder? When does copyright protection lapse if the author is an immortal machine?

There is already a long-standing and lively debate about whether the advent of AI systems challenges the fundamental assumptions, structures and concepts of copyright and other IP laws, or whether current laws will suffice as long as they are applied with a modicum of technical understanding. In part, these mirror familiar philosophical debates about the justifications for IP protection. If copyright, for example, is there to promote the progress of science and useful arts, it may be possible to make room for non-human authors, users and infringers. However, those to whom copyright is a fundamental, moral right derived from human intellectual creation, will find it hard to admit algorithms into the pantheon of protection, even as they learn to create stories, music and images that are indistinguishable from human works. There are also more practical questions around the relevance and application, in the age of AI, of classic copyright concepts such as reproduction, distribution, display and communication.

These questions are important, but not the subject of Jean-Marc Deltorn’s presentation, and therefore of this response. The question that he addresses is less philosophical but at least as topical: to what extent, and on the basis of which intellectual property rights, can deep learning systems themselves be protected? Although these systems will increasingly be developed by other deep learning systems (thus attracting the same questions described above), we are currently still talking about at least partly human creations. Analysing the IP protection of deep learning systems thus involves applying existing IP laws to these new technologies and highlighting any problems that arise.

Specific questions

The first question that Jean-Marc Deltorn’s presentation raises is to what extent deep learning is unique as an IP problem. Deltorn analyses the potential IP protection of deep learning systems as a new form of technology, which suggests that this raises entirely new problems. I wonder whether this is the case, or whether this is actually a species or the broader, long-running question of IP protection of software.

Secondly, it is clear that there are not only gaps in the IP protection of deep learning systems, but also a degree of overlap. For example, Deltorn shows that training processes, weights and network architecture can, at least to some extent, be protected both as trade secrets and through patent law. That raises the question of how companies should decide which protection to seek. On the one hand, patent protection might be preferable because there is no exception for independent creation. On the other hand, trade secrets may be easier to protect in cross-border situations.

Thirdly, Deltorn describes that trade secret law does not protect against reverse engineering. However, article 3(1)(b) of the EU Trade Secrets Directive 2016/943/EU does allow for a contractual exclusion of reverse engineering, in which case reverse engineering would be an unlawful acquisition of a trade secret. This suggests that trade secrets law may well prove to be a viable way to protect deep learning systems, at least in some situations.

There is also, fourthly, an inherent tension between patent protection and trade-secret protection, specifically as regards disclosure. A patent applicant runs the risk that its application will be rejected, or that its patent will later be invalidated, in which case its technology will be both unprotected and public. This risk is obviously not unique to deep learning patents, but will nonetheless have to be factored into a company’s protection strategy. It may be an argument to rely on trade-secret protection in more cases, although the life cycle of the deep learning system will also be a relevant consideration.

Finally, as regards the protection under the sui generis database right, it is not clear whether a collection of training data would meet the requirement that there is a collection of “independent works, data or other materials” as defined in article 1(2) of the EU Database Directive 96/9/EC. In its recent Pearson decision, the Dutch Supreme Court concluded from the CJEU’s Freistaat Bayern decision that this refers to “elements that can be separated from each other without affecting their informative, literary, artistic, musical or other content. The value of the informative content of an element of a collection is not affected if that element, after being removed from that collection, retains an independent informative value. The independent informative value of an element that has been extracted from a collection must be assessed in the light of the value that the information has for any third party that shows interest in that element, and not in the light of the value that this information has for a typical user of that collection.” In many cases, individual training data (this photograph does not contain a cat) will not possess such independent informative value.

Is there a problem?

Jean-Marc Deltorn’s presentation focuses on how deep learning systems might be protected by IP, and highlights a number of gaps in that protection. This conclusion seems to be correct: current IP laws do not provide full IP protection for deep learning systems. However, this discussion might suggest that there should be full IP protection for deep learning systems. In my view, such a conclusion would be premature. If current IP laws do not fully protect deep learning systems, it does not follow that there is a problem, or that rights should be expanded to ensure that deep learning systems are fully protected.

If one takes a deontological, human-rights approach to IP: to the extent that no IP protection is available because there is no human creation or invention, there is no human whose human or moral rights are at stake. Why should we worry about a “value gap for algorithms”? No one’s rights are being infringed. And to the extent that a human author or inventor is involved, then his or her intellectual creation can be protected by existing IP laws.

If ones take a utilitarian approach: it seems that investments in deep learning technology are increasing exponentially. In other words, these deep learning systems are being created anyway, so without the lack of full protection seeming to inhibit economic incentives to create. Why create exclusive rights if the object of potential protection is being made even without the prospect of obtaining that (additional) protection? There seems to be little evidence to suggest that more extensive IP protection of deep learning systems would lead to the development of more or better deep learning systems.

In other words, it is clear that there are IP questions around the protection of deep learning systems, but is there actually an IP problem?

A related point is about freedom to operate. Debates in intellectual property circles often revolve around whether there should be more IP protection for a particular category of creation or technology. This is also frequently how clients initially approach their lawyers: how can I prevent my competitor from appropriating my technology? As often as not, another question turns out to be equally pertinent: to what extent does their wonderful product infringe the (alleged) rights of others? This may also be a significant risk in the development of deep learning systems.

For both business objectives – maintaining freedom to operate without infringing third-party rights, and protecting investments in deep learning technology – it is important that rules about IP protection are coherent and predictable. So that extent at least, the study of “AI and IP” is useful and important. However, there is always a risk that overprotection of deep learning systems could chill further innovation.

The real IP impact of deep learning systems

If we focus on the IP protection of deep learning systems and their creations, we risk missing a related, and perhaps more fundamental IP question. This pertains to the impact of deep learning systems and creations on human creations, and more specifically on the market for human creations.

Regardless of the extent to which deep learning creations are protected by IP rights, they stand fundamentally to change the market for human-created works. If a deep learning system can create a new novel, a new Van Gogh, or a new antibiotic, then the maker or owner of that deep learning system has the chance to make money. But what of all the writers, painters, and scientists, whose entire career output can just as easily be produced by a smartphone app? Everyone is entitled to their own opinion as to how quickly algorithms will write better symphonies than Sibelius, and those who cherish the supremacy of human creation are probably statistically over-represented in IP circles. However, few will deny that deep learning systems will learn to match an ever-increasing range of human creation.

Of course, there will always be human winners at the top end of the copyright market: buyers are human (for now, at least), and there will probably always be a certain snob value to buying ‘real’ real human art or literature, even as it increasingly takes an expert to decide if the work is ‘really’ human-made. Just as a niche market will probably remain for human-operated cars, which  becomes all the more exclusive as insurance premiums for non-automated driving skyrocket.

For makers of bread-and-butter work, however, the competition with artificial creations may well be painful. Although George Eastman opened photography to the general public in 1888 (with the Kodak Box Camera: “You press the button, we do the rest“), the market for professional photography remained relatively stable in the century that followed. In more recent years, professional photographers have been significantly disrupted by the rise of cheap digital cameras and internet communication, which have led to to the availability of a mass of amateur photographs that are (almost ) free and “good enough”. Licensors are increasingly content with “iPhoneography”, and reluctant to pay for real artistry. Yet so far, professional photographers are only competing with other humans.

A significant amount of legal, legislative and lobbying energy is currently being expended on improving the enforcement of IP rights online. This effort is not just intended to protect the kind of top-tier human creativity that no mere algorithm could ever hope to match, but also – especially – the mass of ‘average’ works whose day of algorithmic reckoning is fast approaching. Ultimately, what good is better enforcement of IP rights if there is no longer a market for human creations, because equivalent works can be obtained more cheaply from a machine that has no family to feed? This is obviously only a small part of the broader question of how artificial intelligence technology will impact employment – where, again, the engineers, philosophers and politicians probably have the more important role to play.

In other words, the question of IP protection for artificial intelligence systems is interesting and important, but probably less important, in the longer term at least, than the broader issue of how artificial creations will stand utterly to disrupt the market for human creations.


Leave a Reply