1. INTRODUCTION: ARTIFICIAL INTELLIGENCE – QUO VADIS, LATIN AMERICA?
The fourth industrial revolution1 or second machine age2 are expressions used to refer to a new socioeconomic paradigm resulting from a confluence of technologies that rely on automated exchanges of data by intelligent systems capable of replicating human cognitive activities. We refer to these systems, whether based solely on software or on a combination of software and hardware, as artificial intelligence (AI) or “the science of making computers do things that require intelligence when done by humans.”3
AI is no longer only a subject circumscribed to science-fiction films, such as Metropolis, Terminator, Ex Machina, or Westworld, but is already impacting our daily lives. For example, virtual assistants such as Amazon’s Alexa, Apple’s Siri, Microsoft’s Cortana, and Samsung’s Bixby, to name a few, all rely on different AI technologies. Moreover, an AI system, BlueDot, was the first to ring the alarm on the Covid19 pandemic, way before the World Health Organization was aware of it.4
AI technology has been characterized by periods of wax and wane, called summers and winters. It seems a new summer period has recently begun, led by technologies such as natural language processing (the technology behind most chatbots), machine learning, neural networks, and deep learning.
AI-related inventions, an indicator of investments in the area, have constantly increased since 2013.5 For instance, patent applications related to machine learning have grown 26 percent annually between 2011 and 2016. In addition, the ratio of scientific papers to inventions in the field of AI fell from 8:1 in 2010 to 3:1 in 2016, suggesting a shift from basic science to scientific applications, usually a sign of a mature technology.
Technological change is in the DNA of capitalism. Schumpeter wrote, “the process of creative destruction is the essential fact about capitalism.”6 New technologies make existing ones obsolete. The internal engine automobile makes the horse-drawn coach obsolete; the word processor makes the typewriter an item suitable for a museum, not an office. AI can potentially unleash a Schumpeterian wave of creative destruction, creating winners (those who adapt) and losers (those who do not).
Governments must prepare their countries for the AI revolution. To that end, public policy is key, also becoming a tool to increase a nation’s competitive advantage from a comparative perspective. Good laws provide legal certainty while bad ones increase uncertainty. Entrepreneurs invest in countries with higher quality legal systems and institutions in a global economy where investments can be quickly delocalized (regulatory arbitrage).
Reports about the impact of AI are either optimistic or pessimistic. What can Latin American countries expect to win or lose from AI? On the optimistic side, a report published by Accenture emphasizes AI’s potential to solve two of the region’s chronic economic problems, namely, overreliance on the export of commodities in international trade and a productivity deficit, i.e., total factor productivity increases in Latin America have been lower than in other regions.7 A report published by an Argentinian think tank, CIPPEC, suggests that the early adoption of AI technologies may increase economic growth by up to 4.4% in the next decade. 8
On a more negative side, a report by PricewaterhouseCooper suggests that Latin America will benefit less than other regions of the world from AI-related initiatives.9 A more recent report by PricewaterhouseCooper holds the same pessimistic tone.10 In addition, Oxford Insights11 has published a report on the state of governmental AI-readiness in Latin America, concluding that the region’s countries are mostly unprepared. Only two Latin American countries have devised comprehensive strategies to promote AI, Mexico in 201812 and Uruguay in 2019.13
2. THE IMPACT OF COPYRIGHT LAW ON AI-DEVELOPMENTS
Intellectual property law favors investments and exploitation by granting temporary legal monopolies over intangible assets. I will not refer here to the potential and challenges of using the patent system to incentivize research and development in AI. Instead, I shall delve into the risk of copyright law blocking or chilling developments in AI in Latin American countries. It should not be forgotten that copyright law plays a fundamental role in enabling the training or education of an AI system.14
Like a newborn human baby, a new AI system is in its origin, a tabula rasa. Intelligence, whether human or not, needs inputs to generate outputs. Parents buy their children books to help them learn to read, write, and do basic arithmetic operations. AI also needs to access materials to learn. AI systems need to access vast data pools; the more significant the pool, the more accurate the system becomes. Some data may consist of works in the public domain and are thus free to use, but others may be copyrighted works, especially those in cutting-edge science and technology. AI systems must perform actions on works protected by copyright, such as reproduction, text, and data mining (TDM)15 or the creation of derivative works.
For instance, an AI system developed to discover new treatments against Covid19 would need to carry out TDMs of large quantities of copyrighted papers on medical science. The AI system would have to access, read, and make copies (reproduction) of copyrighted scientific papers. Moreover, those works would have to be adapted or compiled into new works or formats (derivative works). Therefore, to operate legally, the developer of an AI system would be required to negotiate and obtain licenses from multiple authors. This can be problematic for many reasons that will be discussed below.
3. THE LAW & ECONOMICS OF EXCEPTIONS AND LIMITATIONS TO COPYRIGHT
In principle, AI developers requiring access to large numbers of copyrighted works could negotiate licenses with every copyright holder involved. The financial costs of these licenses do not concern my analysis. My concern is transaction costs, i.e., the costs of using the market.16
In a frictionless world of zero transaction costs,17 AI developers would be able to negotiate every license they need.18 However, obtaining a license would be unfeasible when the transaction costs exceed the expected benefits in the real world of positive transaction costs. Thus, uncertainty and transaction costs lead to market failure.
The proper corollary of the so-called Coase Theorem19 is that the initial allocation of rights is essential in the real world of positive transaction costs. Thus, in the case of AI systems, having proper exceptions and limitations becomes fundamental.
From a Law and Economics perspective, exceptions and limitations to copyright may be fair and efficient. According to Landes and Posner, transaction costs20 are the reason for most exceptions and limitations.21 For instance, let us consider all possible transaction costs for an AI developer to negotiate ex-ante licenses to reproduce and/or adapt all the copyrighted works necessary to train an algorithm. These costs would be unaffordable for most AI developers and may discourage them from developing the AI system. Moreover, for an AI developer, it is difficult, if not impossible, to determine in advance what works an AI system will need to access.
Let us return to the case of the right of quotation (droit de courte citation), an exception to the exclusive right of reproduction. In principle, copyright law is absolute, i.e., only the author or rightsholder can authorize third parties to reproduce, distribute, or communicate to the public her work. Due to the problem of high transaction costs and low expected benefits from negotiating ad hoc licenses, there will be few or no citations of others at all in the absence of exceptions and limitations. This will affect the quality of the new works and be detrimental to activities with positive social value, such as teaching, research, commentary, or criticism. Thus, granting an ex-lege permission to quote a certain number of paragraphs from copyrighted work becomes an efficient solution. There is no financial loss for the author in this case since there would be no substitutive market transaction in the absence of an exception. Thus, exceptions and limitations may lead to Pareto efficient outcomes, a situation in which nobody is worse off, and somebody at least is better off. Therefore, from an economic perspective, most exceptions and limitations to copyright law create positive externalities at no cost.
Depoorter and Parisi suggest an additional economic rationale to justify exceptions and limitations to copyright law.22 Without dismissing the transaction cost theory discussed above, these authors state that, even in situations of low or zero transaction costs, exceptions and limitations to copyright law would still be necessary to counteract the potential strategic behavior of copyright holders, in particular, to avoid problems of anticommons,23 i.e., a situation in which a proliferation of copyrights may lead to sub-optimal use of the works.
Finally, this time from a sociological perspective, sometimes the absence of reasonable and efficient exceptions and limitations to copyright law may favor widespread and generalized infringement. I mentioned before that those exceptions and limitations mechanisms built into the copyright system are designed to balance the rights of copyright holders and those of the rest of society. The absence of ex-lege escape valves may lead to collective copyright infringing behavior, which may be facilitated in cases where the chance of being detected and punished is low. For instance, online copyright infringement may go undetected for individual users or, even if identified, it will rarely be punished.
Ostensible and generalized disregard for copyright law not only harms copyright holders by depriving them of their fair share but the entire legal system and society at large, as the rule of law is eroded and trust in the law diminishes. According to Nino, anomie, a concept attributed to Durkheim, is the root of the protracted sub-optimal development of Latin American countries.24
The economic and sociological arguments reviewed suggest that incorporating exceptions and limitations would incentivize lawful AI activities in the region.
4. AI-ENABLING EXCEPTIONS AND LIMITATIONS IN COMPARATIVE LAW
Some countries have added new exceptions and limitations to promote AI in the last few years. TDM refers to automated processes that create further information, patterns, trends, and correlations from existing text and data analysis. As said earlier, AI systems need input (text, data, etc.) to create output, i.e., to learn. Hugenholtz goes as far as to suggest that “much of the current and future development in artificial intelligence, therefore, depends on text and data mining.”25
Contents to be mined may include raw data (in principle, not subject to any form of IPR),26 as well as text, images, pictures, and any other type of creative work that may be protected by copyright.
TDM activities involving copyrighted content will require at least an act of reproduction, and sometimes even adaptation (i.e., the creation of a derivative work) which fall under the sphere of exclusive control by the copyright owner. Therefore, in the absence of an explicit authorization by the copyright holder (such as a license), any reproduction, adaptation, or other action may infringe the right. In addition, the potential threat of litigation may act as a Damocles’ sword regarding law-abiding AI developers, disincentivizing investments, research, and development in the area. To avoid this, some countries have already legislated ad hoc exceptions and limitations to enable the use of copyrighted works by AI.
In 2014, the UK amended the Copyright, Designs and Patent Act and added a new exception for making copies (reproduction) of works for text and data analysis for non-commercial research.27
Copyright, Designs, and Patents Act 1988
29A Copies for text and data analysis for non-commercial research
-
(1) The making of a copy of a work by a person who has lawful access to the work does not infringe copyright in work provided that—
(a) the copy is made so that a person who has lawful access to the work may carry out a computational analysis of anything recorded in the work for the sole purpose of research for a non-commercial purpose, and
(b) the copy is accompanied by a sufficient acknowledgement (unless this would be impossible for reasons of practicality or otherwise).
[Omissis]
The US Fair use doctrine follows a different approach, codified into Section 107 of the US Copyright Act. It provides a judge with a list of four factors to consider, on a case-by-case basis, whether certain conduct infringes copyright law or must be tolerated by the copyright owner. The open texture of the US Act is certainly more flexible than the civil law’s rigid and closed list of exceptions and limitations.
[Omissis], the fair use of a copyrighted work, including such use by reproduction in copies or phonorecords or by any other means specified by that section, for purposes such as criticism, comment, news reporting, teaching (including multiple copies for classroom use), scholarship, or research, is not an infringement of copyright. In determining whether the use made of a work in any particular case is a fair use the factors to be considered shall include—
(1) the purpose and character of the use, including whether such use is commercial or is for nonprofit educational purposes.
(2) the nature of the copyrighted work.
(3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and
(4) the effect of the use upon the potential market for or value of the copyrighted work.
The fact that a work is unpublished shall not itself bar a finding of fair use if such finding is made upon consideration of all the above factors.
Some scholars consider TDM activities to be covered by fair use, particularly after the decision of the US Court of Appeals for the Second Circuit that upheld Google Books’ project of mass digitization of copyrighted works in re Authors Guild v. Google, Inc (804 F.3d 202, 2015).
Japan amended its Copyright Act in 2018.28 Article 30-4 allows access to copyrighted works for extraction, comparison, classification, or other statistical analysis of language, sound, or image data, or other elements of which a large number of works or a large volume of data is composed and computer data processing. Article 47-4 exempts the making of incidental electronic copies of works from copyright infringement. Finally, article 47-5 allows using copyrighted content for data verification purposes connected to the research.
The recent EU Directive 2019/790 of the European Parliament and the Council of April 17th, 2019, on copyright and related rights in the Digital Single Market introduced two new exceptions mandatory for member States. These two exceptions must be transposed into their domestic laws. Article 3 establishes a broad and specific TDM exception for nonprofit scientific research, whereas article 4 provides a more general but narrower TDM exception, subject to some restrictions, for all other purposes.
Article 3 Text and data mining for the purposes of scientific research
1. Member States shall provide for an exception to the rights provided for in Article 5(a) and Article 7(1) of Directive 96/9/EC, Article 2 of Directive 2001/29/EC, and Article 15(1) of this Directive for reproductions and extractions made by research organisations and cultural heritage institutions in order to carry out, for the purposes of scientific research, text and data mining of works or other subject-matter to which they have lawful access.
2. Copies of works or other subject-matter made in compliance with paragraph 1 shall be stored with an appropriate level of security and may be retained for the purposes of scientific research, including for the verification of research results.
3. Rightsholders shall be allowed to apply measures to ensure the security and integrity of the networks and databases where the works or other subject matter are hosted. Such measures shall not go beyond what is necessary to achieve that objective.
4. Member States shall encourage rightsholders, research organisations, and cultural heritage institutions to define commonly agreed best practices concerning the application of the obligation and of the measures referred to in paragraphs 2 and 3 respectively.
Article 4 Exception or limitation for text and data mining
1. Member States shall provide for an exception or limitation to the rights provided for in Article 5(a) and Article 7(1) of Directive 96/9/EC, Article 2 of Directive 2001/29/EC, Article 4(1)(a) and (b) of Directive 2009/24/EC and Article 15(1) of this Directive for reproductions and extractions of lawfully accessible works and other subject matter for the purposes of text and data mining.
2. Reproductions and extractions made pursuant to paragraph 1 may be retained for as long as is necessary for the purposes of text and data mining.
3. The exception or limitation provided for in paragraph 1 shall apply on condition that the use of works and other subject matter referred to in that paragraph has not been expressly reserved by their rightsholders in an appropriate manner, such as machine-readable means in the case of content made publicly available online.
4. This Article shall not affect the application of Article 3 of this Directive.
This is the second time an EU Directive has established a mandatory exception, which suggests the strategic importance AI represents for the EU legislator.29
Because TDM exceptions refer to non-expressive uses of copyrighted work, they do not, in principle, compete with the original. In other words, this exception does not create a substitute good for the original work in economic terms. Moreover, the EU exception makes it a mandatory prerequisite for text and data mining activities to have lawfully acquired a copy of the work. Thus, in principle, it does not unduly prejudice the economic interests of the copyright holder. Moreover, the power of the rightsholder is reinforced in the case of TDM activities carried out by for-profit institutions, as the copyright holder retains the right to deny authorization in this case. However, this provision has been criticized as unnecessarily restrictive.
However, according to some scholars, the two new EU mandatory exceptions may probably be insufficient, if not useless. Article 3 has been criticized for leaving out the scope of the broad exception for-profit activities, negatively affecting AI entrepreneurs, SMEs, MNCs, and even media companies. Article 4, in turn, can be foreclosed, and its purpose may be entirely defeated, as the copyright owner could opt out of the TDM exception, in which case the only lawful way to perform TDM over copyrighted works would be to negotiate a license.30 This leaves the door open for copyright holders to behave strategically to the detriment of AI developers, a risk already mentioned by Depoorter and Parisi in their paper. The risk of strategic behavior may require a broader exception than the one included in the EU’s legislation.31 Last but not least, being a Directive, it must be transposed to the domestic legal systems of Member States. Different national transpositions may create a disharmonizing effect, which is the opposite of what the Directive intended.32
Isolated and disarticulated national and regional exceptions to enable AI technologies, like the ones undertaken by the UK, Japan, or the EU, may be welcome but may not be adequate in a global market where AI investments are investments in AI research and development and are becoming increasingly transnational. Therefore, some scholars have suggested that the time may be ripe for discussing a multilateral treaty on exceptions and limitations to enable AI activities, perhaps under the auspices of WIPO.33
Last but not least, exceptions enabling AI technologies, such as TDM, may incentivize private investments and have a broader social impact. As mentioned earlier, AI output may be biased if the training data set (including works) is not sufficiently large.34 Thus, there is another angle to advocate AI exceptions and limitations, as they act as an antidote against algorithmic discrimination due to limited training datasets. Unlike human discrimination, algorithmic discrimination is opaque and difficult to detect. Trade secrets, contract law, and/or technical protection measures usually protect private AI algorithms. Broad enabling exceptions and limitations for AI may avert this risk by allowing access to a broader data pool. The broader the training dataset, the less likely an algorithm would produce an unwanted biased output.
5. SHOULD LATIN AMERICAN COUNTRIES INCORPORATE AI-ENABLING EXCEPTIONS AND LIMITATIONS TO THEIR COPYRIGHT LAWS?
Copyright law is sticky; all rights are reserved to the copyright holder. Therefore, in the absence of an exception or limitation, any act of reproduction or the making of derivative works must be authorized by the rightsholder. Thus, in the absence of a license or authorization from the rightsholder, the developer of the AI system should refrain from using copyrighted works or do it and risk being sued for copyright infringement.
Exceptions and limitations are an endogenous stabilizing mechanism built into the copyright system, designed to balance the rights of copyright holders, users, and other stakeholders. The terms exceptions and limitations usually appear together; however, each word has a specific meaning. For example, limitations refer to types of works excluded from copyright protection (e.g., articles 2.4, 2.8, and 2bis.1 of the Berne Convention). In contrast, exceptions refer to uses of copyrighted works that do not require authorization from the copyright holder (e.g., articles 2bis.2, 9.2, 10, and 10bis of the Berne Convention).
Existing exceptions and limitations are insufficient to provide legal certainty to essential AI activities like TDM. For instance, the quotation right (droit de courte citation) recognized in every Latin American copyright law may not be sufficient for AI systems. Furthermore, this exception usually only allows quoting or citing a small part of a copyrighted work, while AI systems, to improve, need to access, reproduce, and even modify substantial amounts of complete works. AI systems require new ad hoc exceptions and limitations to securely reproduce or exploit copyrighted works to improve their algorithmic capacities.
6. STATUS QUAESTIONIS: ASYMMETRIC TREATMENT OF EXCEPTIONS AND LIMITATIONS ACROSS LATIN AMERICAN COPYRIGHT LAWS
The rights of copyright holders are mostly consistent and similar across Latin American countries. However, the quantity and scope of exceptions and limitations to copyright law are asymmetric. Some countries have incorporated long and updated catalogs of exceptions and rules, while others have only legislated a handful, suitable for an analogical but not a digital paradigm.
For instance, the Argentinian Copyright Act no. 11.723, which dates back to 1933, has a limited and outdated set of exceptions and limitations,35 the Brazilian Copyright Act no. 9.610 has a slightly broader set,36 and the Chilean Copyright Act no. 17.336 of 1970 has probably the largest one in the region, after its amendment by Act no. 20.435 of 2010.37
A report by Crews concerning copyright limitations and exceptions for libraries and archives, first published in 2015 and updated in 2017, seems to confirm the disparity38 According to this study, Argentina and Brazil belong to a minoritarian group of 28 countries (from a study pool of 191) with no specific exceptions or limitations for libraries or archives. Chile and Peru, on the contrary, have incorporated many to allow libraries and archives to perform their functions without infringing copyright law.
To the knowledge of this author, after a brief but not necessarily exhaustive analysis of the WIPO Lex Database Search,39 no Latin American country has yet incorporated any specific exception or limitation to copyright law to enable or facilitate the training or education of AI systems, such as EU’s recent TDM exceptions. At this point, it is worth noting how quickly technology evolves, which makes it even more challenging for the legislation to catch up, even when it is willing to do so. For instance, the Chilean reform of 2010 included many exceptions to enable the (then) new or recent technologies, such as reverse engineering of computer programs (Art. 71 “o” b) and transitional or accessory reproduction of works for the purpose of transmission in a network (Art. 71 “p”). However, none of these, nor other exceptions and limitations included in the Chilean law, serve today the needs of training an AI algorithm that requires access to creative works.
From this perspective, a flexible and open-ended system, like the US-style fair use, may seem a better regulatory option. Some scholars have pointed out that, in principle, there may be no normative or institutional constraint for a country belonging to the droit d’ auteur-civil law tradition to adopt a system like US fair use. However, in practice, due to other institutional and social constraints, including the preparation and role of judges, to include a US-style fair use clause in Latin American copyright legislation seem chimeric, to say the least. With a certain degree of pragmatism, other authors have suggested that it may still be possible to reach some middle ground between the inflexibility of the civilian model of exceptions and limitations and the open texture of the US fair use.40 This is an avenue a reasonable legislator may want to consider.
International copyright law gives countries some room to carve out their own exceptions and limitations, as long as they do not conflict with the so-called three-step test, as described in Article 9(2) of the Berne Convention for the Protection of Literary and Artistic Works (exclusively concerning the right of reproduction), Article 13 of the TRIPS Agreement (encompassing all “exclusive rights”), or Article 10 of the WIPO Copyright Treaty (in relation “to the rights granted to authors of literary and artistic works under this Treaty”). Moreover, contrary to the prevailing judicial interpretation of the three-step test standard in Latin America, a balanced interpretation capable of addressing the interests of copyright holders, users, and other stakeholders is possible.41
Finally, Argentina, Paraguay, and Uruguay still enforce a domaine public payant, which requires the payment of a fee (actually, a tax) to a state agency to publish, translate, adapt, etc., works in the public domain. The domaine public payant may become an additional hurdle to educating AI systems, as in these countries, it may also be required to pay for specific uses of works in the public domain.42
Exceptions and limitations are a supplement, not the antithesis, of copyright law (and neighboring rights). Both are necessary elements of a sustainable knowledge economy, like two sides of the same coin. In this regard, the US may be taken as an example. This country, still the most innovative and creative globally, aggressively advocates vigorous enforcement of US copyright abroad and safeguards the most generous fair use in the world. The flexible US fair law has sometimes enabled disruptive technologies, even at the expense of copyright holders, such as in the famous Betamax case concerning video cassette recording technology43 or, more recently, by validating the ambitious Google Books digitization project. Latin American countries should learn from it.
7. CONCLUDING REMARKS
Latin America is the most unequal continent in the world.44 Unchanneled by a fair and efficient legal framework, AI technologies may aggravate the existing inequality. However, at the same time, AI technologies may contribute to a more equal and sustainable development in the region, increasing the productivity and competitiveness of Latin American firms, and in so doing, creating skilled jobs and promoting economic growth.
Latin American governments must provide clear rules to make the most of the AI promise while minimizing potential risks and negative externalities. In this article, I suggested a cost-effective measure, incorporating exceptions and limitations to balance the rights of copyright holders and users, as well as to provide a signal to attract investments in AI in the region. Ideas are free to copy. Moreover, most countries’ laws are not subject to copyright protection.45 Thus, Latin American countries may profit from the experience of first movers like the UK, the EU, and other developed countries that have already legislated new exceptions and limitations for AI. Of course, we are not suggesting an exercise of copy-pasting. A transplanted law, to succeed, like a transplanted plant, must be adapted to the climate and characteristics of its new soil.46 The role of the local legislator is not minor, to study and learn from other countries’ legal experiences and to adapt, implement and even improve foreign laws if possible.47
Properly incorporating copyright exceptions and limitations for AI may provide the following advantages for the countries willing to implement them, namely,
1) Cost-effectiveness. Changing laws is not a costly endeavor; it requires no big budget or infrastructure. Moreover, introducing legal changes that are necessary for the region’s progress and do not prejudice the interests of any particularly strong interest group seems politically feasible.
2) Foreseeable positive impact on the Latin American AI industry. Clear rules, and legal certainty, provide positive signals to the market, attracting investment and talent.
3) Limited to no negative impact on copyright holders. As explained above, most exceptions and limitations for AI do not preclude potential substitute market transactions. Moreover, they may incentivize them (licensing). In economic terms, such a situation is deemed desirable (Pareto superior).
4) Promoting a culture of abiding by the rule of law. Too rigid copyright laws, without escape valves to accommodate changing social norms, may even be detrimental to those they are intended to protect (the copyright holders). Well-calibrated exceptions and limitations may limit copyright anomie and overall improve the respect of the rule of copyright law.
Developing countries do not need to reinvent the wheel. Legal reforms can create valuable public goods. With limited legislative budgets, this can be achieved on the coattails of developed nations. It is time for Latin American governments to prepare for the AI revolution and adapt their policies, strategies, and legislations accordingly. Incorporating new AI-enabling exceptions and limitations to copyright laws is a cost-effective proposal not only to balance the rights and obligations of rightsholders and users but also to compete on equal terms in a global algorithmic economy.