Three routes to protecting AI systems and their algorithms under IP law: The good, the bad and the ugly
Katarina Foss-Solbrekk · Journal of Intellectual Property Law & Practice · 2021
Artificial Intelligence (AI) applications are transforming society, promising to help people make better, more informed decisions and therefore promote greater productivity, improved efficiency and raise general well-being.1 As part of the rapid uptake, algorithms now underpin an increasing number of important decisions about individuals’ lives, influencing credit scoring,2 admission to university,3 and even when people should be discharged from hospital4 or how long they should be sentenced to jail.5 However, this raises serious ethical and legal concerns. One of the important questions is what type of intellectual property (IP) protection these algorithmic models currently attract in the EU. And, moreover, what protection they should attract, for the benefit of both industry and society at large. Katarina Foss-Solbrekk is a doctoral (DPhil) candidate in IP law at Oxford University. She holds an Advanced LLM in Law and Digital Technologies from Leiden University and an LLM in EU IP Law from Stockholm University. This article reviews how Artificial Intelligence (AI) systems can be protected under three IP frameworks: copyright, patent and—taking IP in its broader sense—trade secrets laws. In contrast to other contributions, the focus of this work is how the AI system itself is protected, not its output. AI systems are primarily protected as trade secrets, as attempts to protect AI systems under copyright and patent laws encounter difficulties. In copyright law, algorithms are excluded from protection under the EU Software Directive. They also struggle to meet the author’s own intellectual creation criterion, in addition to not necessarily being a creative expression of said creation. Acquiring patents for AI systems is also difficult as these systems may fail to satisfy the technical character and inventive step requirements. However, although trade secrets law is the most common avenue to IP protection, an alternative and increasingly successful route is to patent AI systems as computer-implemented inventions. The impact of trade secret protection for AI systems in terms of transparency and accountability is also discussed. Because trade secret protection subsists for as long as the information remains confidential and requires actors to take steps to ensure confidentiality, trade secret protection facilitates algorithmic opacity. This has serious consequences. The consequences are extremely serious and widespread, even affecting matters of life or death. In the US, Michael Robinson was sentenced to death, based on evidence produced by computer software. Yet he was not allowed to challenge, or even access, the source code: the judge held that trade secrets law would take priority.13 Similar problems are predicted to arise across the pond. As things stand, trade secrets law will grow in strength and scope because copyright and patent law have been drafted specifically to exclude algorithms from protection. Trade secrets law, therefore, which was primarily intended to prevent breaches of confidence by ex-employees, has been largely subsumed into IP law, creating rights and providing a convenient mechanism for companies to fill the gap where classic IP law fails. The result is that individuals are hit with complete opacity and proprietorial unaccountability, without even a fair use exemption. This article aims to shed some light on how algorithmic systems are treated under copyright, patent and trade secret laws, illustrating how EU law is subject to an Italian epic spaghetti western gamut of ‘The Good, the Bad, and the Ugly’. It shows that the current reliance on trade secrets is inappropriate, unnecessary and, in short, a ‘bad’ solution. Copyright is not the answer either, being at best an ‘ugly’ alternative because algorithms’ technical functionality means they are likely to fall short of constituting creative subject-matter, according to EU case law. Rather, the ‘good’ way forward is to recognize that it is patents that are specifically designed to protect technical inventions. Although in its infancy, IP protection for algorithms under patent law is expanding with its growing acceptance of computer-implemented inventions (CIIs). This is the preferable and logical route, providing the dual benefit of avoiding the dangers of trade secrets law and also acknowledging the novelty of AI. Scholars do not agree about the definitional terms of what constitutes an algorithm. Moreover, defining an algorithm based on a single term, such as an abstract state machine (data structures) or a recursor falls short of encapsulating its entire scope, as ‘algorithmic duality seems to be a fundamental principle of computer science’.14 In simple form, however, an algorithm is, as defined by Hill, ‘a finite, abstract, effective, compound control structure, imperatively given, accomplishing a given purpose under given provisions’.15 Algorithms can thus be viewed as a method constituting a sequence of steps used to calculate different data variables to yield outputs. This way, algorithms can ‘tell’ computers how to complete tasks when incorporated into computer programs.16 This process is enabled by code, which is the actual implementation method of algorithms, achieved by delivering the algorithmic instructions to computers in certain programming languages.17 Computer programs thus comprise codified sequences which express algorithms (instructions).18 Algorithms have various control structures and are therefore divided into algorithmic sub-categories, each warranting a new definition depending on their control function,19 eg ‘sorting’ and ‘searching’ algorithms.20 Of particular importance are predictive algorithms. Predictive algorithms are part of predictive modelling, that is an approach, not a fixed process, designed to try and predict future results based on historical data.21 They scour through big datasets searching for patterns or correlations between different variables and the final output to engender predictions.22 Machine learning, a method by which computer systems learn from data and then act autonomously with little to no human intervention, usually aids this process. There are two main categories of machine learning algorithms: supervised and unsupervised. Unsupervised learning algorithms find hidden structures from unlabelled datasets, ‘as the aim is for the system to group data that is similar’.23 Supervised learning algorithms detect structures based on labelled inputs, which is ‘tagged data’, and desired outputs.24 For example, if the data comprises images, then such tags may include gender, elephants, or watermelons. Supervised learning algorithms are trained on labelled data to form algorithmic models, ie AI systems, which may be applied to ‘unseen points’ to make predictions.25 Predictive algorithms are those most commonly employed for the AI systems introduced into society. Indeed, their influence is widespread. Their calculations impact current events, as well as future ones, by determining criminal sentences, whether individuals are to be granted loans or admitted to university, our ‘newsfeeds’ and furthermore. these systems are IP protection to protect their The of copyright is and if copyright to AI systems the of and that the and of copyright protection to AI across EU protect the algorithms also with algorithms eg protected in not in a of the current this on the copyright of the method or of algorithms the more Because algorithms can be as computers programs and form part of are two by which they attract copyright as or as computer Although of the protection to and no of computer the the to for of that computer programs be under of the of the Copyright that computer programs fall the of the of these computer of the that and are Algorithms to fall its The EU has and in the to the into its legal In the EU codified the Software in with the of of the EU in that it is to the the under the Software those in the when the protected subject falls the of the algorithms fall under EU copyright law is therefore under the Software Directive. of the Software that computer programs are protected under the as also that of or are not that computer programs include in form, those which are incorporated into that a computer expression is protected the and of programs and its are not protected, is the algorithms and programming those and Algorithms This is by case law. In the that the copyright protection of computer programs to form of a computer which the in different computer such as the source and the The that do not to a computer as under of the Software However, the that may attract copyright protection under the The that a programming and the of data used in said fail to a computer in that they may fall under the This is in greater that the to the definition by the Software Directive. to and the source of an algorithm copyright protection, the algorithm itself constitutes a computer As that the not protection on algorithms. Algorithms therefore fail to copyright protection as part of the computer programs they underpin to the Software Directive. under the algorithms satisfy two they be by being the author’s own intellectual In its case law, the has that this may be of creative or sequences and express in and the that for subject to be is both and that the subject the of its as an expression of and creative Moreover, the held that subject is not the of a subject has been by technical or other which have no for creative an author’s own intellectual creation some form of creative human to be as the work an expression of said expression not be in form, it be with and and not be by technical an example, in the held that the is not if the of the different or of the the are by their technical In however, where the whether attract copyright protection, it that copyright protection may be on subject even if its has been by technical that its being has not the from in that subject as an expression of and creative the held that is in part to a technical may be subject to copyright, the work is and an expression of creative the of other that may to the technical result not necessarily this technical functionality if the of subject by their technical and are by said then copyright protection is not according to the different of an are that the and the expression This with where the that copyright is not granted of or of the is more by protection to whether or the or form of its expression may This therefore a to algorithms, although the has it in its the EU is to with the under the as by the may a of as will be algorithms are to according to case law. these two are on the type of algorithm. Unsupervised learning algorithms and models are to be as an author’s own intellectual creation because the a in the algorithms’ algorithms and learn without human work the author’s own intellectual creation if the not as they are on unlabelled the not the data in an creative Unsupervised learning algorithms therefore fall short of to under the Directive. Supervised learning algorithms struggle to satisfy these two as they are thus not with technical For example, algorithms and algorithms of to a single output to in the of such and their of these algorithms fall on to algorithmic models, the most common and As hidden from the tasks without and human control its they with the author’s own intellectual creation also encounter with this as they As a and are for copyright protection. and models, as well as comprise of algorithms on labelled labelled data are algorithms and data to for a purpose the an intellectual the the expression a greater and models, as well as can be in different such as or Their expression is likely it by technical as In however, the a by that the technical may be if the technical not prevent the from their in the work through and creative This may an for these supervised algorithmic the and fail to recognize is that a and creative when and algorithms, and said algorithms on data by that in to a The in this the and as the said their it is the that to that and no technical prevent from should the of by the in as the the technical and as algorithms such they are likely to for copyright protection if decisions are As it copyright the source code, not the that the source is based ie the algorithms, algorithmic models, their data and to the result of the of the not its The work in its should be protected, as is the case for other that not make the of that expression It and creative and should be as This to to algorithmic models shows a the technical is an part of the process not the creative work of the to recognize creative on the that algorithms technical seems in light of the of and copyright protection to and technical drafted by a the in which can with the law it is how an new way for example, creating do data copyright protection under the Indeed, copyright may to data if they are by or they they likely fail to satisfy the data protection under the which is to from copyright, is The of a not has to that big data and data fall its also protection to the or of has been a or The has to on these two the that the of big and data and the of as such is or machine learning data may therefore protection under the Directive. are from of data from the would be general may be to the use and a of would also if algorithms for copyright protection, as copyright may be to protect confidential In contrast to the of that copyright to protect confidential is to the that the to of information and of the to a from the rights under copyright the and in This would if copyright to algorithms, the should algorithms fall under copyright One benefit of this in the are the under copyright, algorithmic information to be without IP protection. The that subject to computer may or said the source code, without copyright in moreover, the algorithm by as under of the algorithms and AI systems under copyright on other which is on the that should be for the they Yet to satisfy the the of copyright in this new at It is not copyright has not to for such as with the of the it that algorithms largely fall the scope of EU copyright law. protection has largely been to AI However, as it will be it is to patents for these systems if as although this route is not as used in as it information from the patent also the and patent to on the of AI systems, the that they not have or that such data was not for rights that from their and are as a method or use protection, an satisfy three main it be which requires the to not form part of the state of the ie that it is to a patent it have an inventive step it is to a in the it have technical by creating a technical which a technical or through being to a technical to algorithms, it is important to that whether computer programs and are at remains a of patents whether or in of of the the legal for patents in and computer programs under the as to ‘a sequence of that ‘a method the method from Yet the has more patents with the also that computer are not excluded from in to algorithms, the that models and algorithms AI and machine learning are if the patent of a method the use of technical as a technical character is on the as a patent Because of the of computer programs and the of a computer to technical character on models, have as a computer or other at is by means of a computer As the algorithms to this a computer to be into they may be as part of the method of a computer or as Computer programs are the of computer algorithms are to patent protection as part of the computer they In the the of a patent in ‘as it was to a the that computer programs are if they a technical that between a and the to the However, in the a patent for a computer data across different by that the technical character as it a and the The has computer those that the of a computer with a patent on its technical the type of and its algorithms The predictive algorithms and models to this are to the of as well as applications on the not a functionality a technical between and software. predictive algorithms therefore fail to satisfy the technical by the from the are not and the of said are to where it the or of a As most of the are it is that may the of computer programs The for computer The Italian and patent do not computer patents for that these inventions are and as The patent of and the computer As it is that algorithms are as computer Algorithms are more likely to be as The that are their method computer or a certain functionality when by a on a computer a computer The has also that models in patent not their patent However, the technical character of inventions by the what is for other when the inventive those the that to the technical character of the are to be In the technical from a computer algorithm to some However, in the a patent for a data algorithm because it and a technical In the that a machine learning algorithm to information according to an inventive step and technical In the that the algorithm to technical character because the was in a In the the of patent for a algorithm as it an inventive step and technical In the a patent for a based on an algorithm because of the of an inventive although the the case to the of and In the of that an algorithm to the technical by the on and by the and the data and be in the inventive step It is in the two patent applications because they an AI machine as the the requires a which most The and also a patent because a machine was as the in It is thus not to patent inventions where the AI system is as the is a in the AI system and AI produced output. As this on the the not the algorithms in the type of algorithm may a in whether a Unsupervised algorithms of a as is the case with on the be more difficult to protect as they can without human transforming it into a system not by a This the human is for the and may its and the process. this would be no the not from The to supervised algorithms they even more human As certain AI systems are ‘a of a human these the of patents for algorithms. an algorithm in a computer to complete tasks is not technical if the algorithm the technical criterion, they an inventive and to be as algorithms are Scholars such as and that algorithms are likely for patent the that it is to patents for algorithms they a technical in a and with Although most of these algorithms, algorithmic models may fall under the of and on their they are and form part of patent technical which to a technical on the such as data structures for they are as by the Acquiring patents based on technical and and in is this that ‘a of the technical should be and in the patent because of technical in the it is to ensure that the have to be In other the technical and the in a that both are to a technical to a technical an however, legal and is little to The technical character remains and it on the type of technical and at The for state that a a for the purpose of a technical based on likely although steps of the and the may to the technical they help the technical The may therefore be usually and Although the and patent that algorithmic patents are in and the US, patent state that they do patents for algorithms, that are as This is to The and that such patent applications is The the of information or patents to the of this although they and patents for and control in Of the patents by the this and The to patent algorithms and algorithmic models therefore that the of may be to on the part of legal algorithms under patent law may some of transparency as patent are This two patents of with to the exclude of the in patent information would to trade secret protection. the of patent applications to it may actors to of their without of their they would also be an as the of the for which be used for algorithmic in and to to algorithms’ to patents on the with under patent and copyright law, and the of algorithms, actors to trade Trade secrets are the for to protect their algorithms. even this is the best of as in some things that be protected by are best Indeed, this the of information it from the purpose of trade laws. They to such as confidential information a from and breaches of and the trade secrets in and the the which an may or ‘a may and will a that information be Trade secrets are not intended to be used as an to protect are they to be an IP as is in greater systems or in and criminal and, by accountability This is in the their protection. of to ensure that and legal may protect from being or used without their on three that the information is it remains to people commonly with such that its from its that the said information has steps to protect its The EU trade secrets protection through the in The incorporated of with not trade secrets, and of the that they as well as and in addition to data on algorithms’ general and they are for trade secret and algorithms for example, trade As trade secrets of the data and other information to the algorithm falls its data may be also Although the to those in the trade secrets information from and do as under of the rights to or Trade secrets or where each of these are in the on or for the of and for However, a the no such as that trade secrets be or to the way for algorithmic This is when trade secrets for AI systems are The algorithm is not from the for It is hidden to trade secret as this remains for as long as the information remains actors take steps to ensure and that is little for to algorithmic This is by the that AI systems are IP fail to for patent or copyright protection, trade secrets do not or property rights The trade secrets from other such as copyright and which are treated as property in the Indeed, as and ‘as for the the on trade secrets not trade secrets as an The secrets are not which has trade secrets this not necessarily to property EU do not trade secrets as property or AI systems are and should be protected as They should not fall under the of trade laws which are not designed to protect or type of property or are trade secrets IP by most EU and, as AI systems are to recognize this is for both and It means rights for their AI systems as trade and it means algorithmic transparency as trade secrets are rights and information to the algorithm how it or the data its results are based on may be on trade secret Trade secret law, therefore, to the algorithm and to the transparency of the entire as trade secrets are in for as long as the information is it means that AI systems may It is thus in the of both and society to AI systems to a different IP This article has the for an of how AI systems and algorithms should be protected under IP law, in a to and best the current reliance on trade secret laws is given the under copyright and patent laws, the for IP protection be the rights of individuals and society. individuals to of and that impact their lives, is not is the reliance on trade secrets in terms of the aims that aim to AI systems and algorithms under patent or copyright law is therefore not for society, also to the of the IP system at large. IP law a to recognize AI as creative or technical that it for copyright or patent protection. As AI systems and their algorithms are technical for copyright, not technical for patent protection, are currently is Copyright seems however, with the in terms of of algorithms and AI systems under patent law as is more as from the The new for even that machine learning algorithms such as a technical purpose are how and when algorithms the technical character little for to try the patent law route when trade are more IP and their therefore, to of AI systems and algorithms as a not an algorithms under the of patent law not the to algorithmic transparency patent applications and also the of the as technical and the system The of other to the transparency of AI even the to the of for the benefit of society, should be as as for to AI such as the algorithms and how they trained and not to the of trade In short, the IP should that the current use of trade law to protect AI work is a ‘bad’ creating a of transparency and Copyright protection would a ‘ugly’ for systems to be and subject to if it would technical systems under a designed for The best answer is to the scope of patent protection to algorithms, algorithmic models and their this would help with the transparency of systems not of their functionality and not also by greater into which systems In the current of our and of to to it is patents that are the ‘good’ and should be as