Disclaimer: This is a shortened and updated version of Section 6 of my article Liable and Sustainable by Design: A Toolbox for a Regulatory Compliant and Sustainable Tech

You can read article in full (in open-access) here:

https://www.mdpi.com/2071-1050/16/1/228?fbclid=IwAR1OVwqknbdNrUJxwW4Rt42z45qs0W3RS8cgTwC7qkXH2UBI2al2qPSvo54

#digitaleconomy #Web3 #Web4 #websustainability #digitalecosystem #techlaw #innovation #innovationlaw #startups #digital #cybersecurity #data #privacy #dataprivacy #dataprotection #AI #GDPR #EUAIAct #IP #IntellectualProperty #copyright #patent #databaserights #EPO #WIPO #TDM #inputdata #outputdata #ChatGPT #sourcingtrainingdata #LDMA #NFT #blockchain #token #crypto #cryptoassets #cryptocurrency #Binance #Bitcoin #BTC #Ethereum #ETH #Litecoin #LTC #Doge

Now that I have your attention (food is good, old but gold, right?), let us delve into some tenets of IP law for Web 3.0 and, increasingly, Web 4.0

(A necessary introduction that you will see in every recipe :)

The pandemic has exacerbated the effects of the digital transformation: the extractive economy is steadily giving way to the new economic space—the digital economy. This transformation shakes the very foundations of the existence and purpose of law, i.e., the regulation of social relations. However, today, the consequences of developing tech in an unsustainable manner are becoming obvious. Unsustainable tech development contributes to trust erosion, misinformation, and polarization, leading to such legal/ethical issues as irresponsible practices of all sorts, unsafe and insecure digital market, inequality, lack of transparency, breach of privacy, etc.

These developments occur partly because the algorithms that shape our economies, society, and even public discourse were developed with few legal restrictions or commonly held ethical standards. Consequently, ensuring that technologies align with our shared values and existing legal frameworks is crucial.

This series of blog posts explores existing and prospective legal and regulatory frameworks that make tech not only legal by design but also, and especially, liable and sustainable by design. The key questions include whether new laws are necessary or if existing legal, regulatory, and ethical concepts can be adapted (or both!).

It is argued here rather in favour of the adaptation of pre-existing legal concepts, ethical standards, and policy tools to regulate the digital economy effectively, while paying attention to possible gaps and possibilities to fill those gaps. The objective is to synthesize these concepts, analyse their applicability to Web 3.0 and Web 4.0 regulation, and provide a toolkit (see below) for a regulatory compliant and sustainable tech. The blog series focuses on organizations involved in tech and innovation, particularly Web 3-4 actors, using systems analysis to examine regulatory constructions both functionally and institutionally.)

Figure 1: Toolbox. Source: Anna Aseeva (2023), 'Liable and Sustainable by Design: A Toolbox for a Regulatory Compliant and Sustainable Tech', in Sustainability, Vol.16, https://doi.org/10.3390/su16010228

In Episode 4 of the series, I analyse the most relevant existing concepts, recent practices, and avenues in IP law applicable to tech organizations.

Recipe #4: IP for Web 3.0 and Web 4.0

1. In this episode

As we saw throughout this blog post series, it is becoming increasingly obvious that the technologies that currently shape socio-economic relations must be consistent with both our shared values and the existing legal and regulatory framework. In the last episode of the series, the one on data protection, we saw that data collected via the internet is, among others, an intellectual property (IP) rights issue, via, for example copyright questions related to data and its deployment by tech organizations.

Today, the issues related to the usage of data with various IP rights attached to it are exploding because of the sky-rocketing development and use of generative AI systems in most areas of our socio-economic and, simply, everyday life. In the previous post, we briefly saw just one new regulatory aspect of the interface of IP with AI: namely, the prospective EU AI Act requires detailed summaries of the copyrighted data used for the AI training to be made publicly available. Furthermore, the AI Act’s compromise proposal contains two provisions relevant to IP, namely, to copyright.  

These two provisions arguably further consolidate the existing balanced legislative approach adopted by the EU, providing additional clarity for both copyright holders and AI model providers, as well as requiring to publish sufficiently detailed summaries of the content used for training, making it easier for third parties to understand the sources of training data (for more detail on these two provisions, see here).

Note that there is currently no consolidated text of the EU AI Act yet. Therefore, in this post, I go into detail regarding most relevant existing concepts, recent practices, and avenues in IP law applicable to AI and Web 3-4 more broadly.

2. IP for Web 3-4: What’s in a name?

The classic definition of intellectual property is a set of the intangible products of human creative activity. Notably, unlike real property and personal property, which are often protected by physical security devices, IP is chiefly protected by sets of enforceable legal rights granted to ‘owners’ or ‘holders’.

Web 3, and even more so, the upcoming Web 4, are bringing about the heyday of IP rights in the digital economy and our everyday life through DAOs, AI, NFTs (see further in this post), and today, increasingly, the metaverse. IP law applicable to metaverse is still under construction, but the excesses in the field have already shown how urgent it is to take a clear regulatory and ethical position on the subject-matter.

The AI-related IP questions can roughly be divided into those stemming (i) from the input data  for AI and (ii) from the output data of AI. For obvious reasons, the most relevant type of AI, especially regarding output data, is generative AI, which is defined below.

3. IP aspects relevant for the input data for AI

Along with classification AI, generative AI systems are based on machine learning and thus partly write and adjust themselves: these systems work through an iterative training process that passes vast amounts of data through the system. More specifically, these systems have been trained on vast amounts of data and have mapped patterns across the data, enabling them to generate similar data that is often very difficult to distinguish from content created by a human.

This type of AI is thus able to create content in response to an input prompt, involving generating answers in text form in response to a question, or an image or piece of video in response to a text prompt. The most typical recent example of such systems is ChatGPT.

The dependence of generative AI on datasets cannot be overestimated: it means that the input for the system—sourcing training data—is critical to its output. When the data are personal data, the respect of relevant ethical standards, as well as privacy and data-protection law are crucial, as discussed in the Episode #2 and Episode #3 of this series, respectively.

In terms of IP law, data-containing material is protected by copyright and database rights. The legal issues raised by these two legal constructs for consideration by any AI operator are as follows.

(i) Firstly, data may be copyrighted. Copyright protects works of artistic or author expression (those are broadly defined and include such material and content as books, films, music/video recordings, and even computer software, databases, etc.) against unauthorized reproduction or distribution by third parties. While it does not limit what material might be considered as protectable artistic or author expression, copyright does not extend to functional works or ideas. Using copyrighted data without the consent of the rightsholder can amount to infringement of the copyright owner’s reproduction right—that is, their right to control copying of the work.

(ii) Copyright protection of/in databases can separately and/or additionally be protected by database rights. Databases can encompass websites, among other digital media, and thus, qualifying databases receive protection when there has been substantial investment in obtaining, verifying, or presenting the contents of the database (see, e.g. here and here). If somebody extracts (including by either permanent or temporary transfer) or reuses all or a substantial part of the content (i.e., makes the content available to the public) of a protected database without the rights owner’s consent, the database IP rights are infringed.

There are, however, some exceptions to both copyright and database IP protection that allow the use of the copyrighted work or database in the context of training AI.

(i) Firstly, the exceptions to copyright infringement include

- non-commercial research or private study,

- criticism,

- review and news reporting, or

- caricature, parody or pastiche—all of which are subject to a ‘fair dealing’ restriction, an exception to which is discussed below.

(ii) There are also related exceptions for text and data mining (TDM) and temporary copies.

All these types of IP protection vary across jurisdictions and must be analysed depending on the country of the AI’s operator, its market, etc.

With respect to database IP protection, a ‘fair dealing’ exception applies to databases that have been made available to the public (in any manner). Database rights in publicly available databases are not infringed by fair dealing with a substantial part of their contents provided that

(i) the extraction is carried out by a person who is a lawful user of the database;

(ii) the data are extracted for the purpose of illustration for education or research and not for any commercial purpose; and

(iii) the source is referenced by the user who extracts the protected content.

Defined in this way, however, this exception is unlikely to apply in a commercial context.

In sum, regarding the input data, the tech firms that offer and/or use generative AI should be mindful that when their AI has been—or might have been—trained on data that are protected by copyright or database rights or both, and none of the above exceptions apply, those data must be validly licensed; if they are AI suppliers, they must make sure that they consider any relevant IP rights in full and with up-to-date information; if they are users of such AI, they must seek assurance on this point from the AI supplier.

4. IP aspects relevant for the output data of AI

The two types of IP protection that are the most relevant to the second item in the present analysis, namely, generative AI output, are patents and copyright. Copyright was defined above in the analysis of the input data. The patent, which is the earliest type (together with the trademark) of IP right, is defined as ‘a set of rights granted to the inventor of a product or process which is ‘new’ (or ‘novel’), involves an ‘inventive step’ (or is ‘non-obvious’), and is capable of industrial application (or ‘useful’)’.

Regarding patents, in the EU, for example, ‘inventions’ generated by AI are not patentable. In 2022, the European Patent Office (EPO) decided in the DABUS Case that a generative AI system cannot be regarded as an inventor: that is, on a patent application, the inventors have to be persons with legal capacity. Neither an invention generated solely by an AI machine without human involvement, nor the owner of such an AI system, is entitled to a patent: AI is rather considered a computer-implemented invention.

The European Patent Convention (EPC) does not provide patent protection for computer programmes; however, it can offer such protection for inventions involving software if the invention produces a technical effect serving a technical purpose (EPC, Guidelines for Examination; Part G, Chapter II, 3.3.1 Artificial Intelligence and Machine Learning). One example might be an automated system processing physiological measurements to provide a medical diagnosis, such as a heart-monitoring apparatus using a neural network to identify irregular heartbeats.

Regarding the interface of copyright with AI output data, copyright protects (i) the so-called entrepreneurial works (films, sound recordings, broadcasts, and published editions) and (ii) original LDMA works (literary, dramatic, musical, and artistic works).

Unlike patents, copyright does not require novelty (see, e.g., Bently, Sherman, Gangjee, Johnson, Intellectual Property Law (OUP), 2022, pp.126-127). Regarding entrepreneurial works, there is no threshold of originality either: for instance, sound recording protections extend only to a specific recording of a song and last for 70 years from the recording’s creation. It is argued that these rights could apply to generative AI outputs: for example, if AI generates a song and records that song in the process, the person who took the necessary steps to have the AI generate and record the song is likely to be the producer and would hence hold the copyright to the recording.

In contrast to entrepreneurial works, the copyrightability of LDMA works requires originality—i.e., the work has to be the author’s own intellectual creation (see, e.g., here and here). When an LDMA work is created by a human with the use of AI, if the work expresses original human creativity, the AI will be treated as a tool (see, e.g., here). The work will receive copyright protection much like any other LDMA work, and the rights will belong to the human author.

The copyright’s threshold of originality is important here because later in this post, it will also be relevant in a discussion of NFTs. Thus, originality refers to a particular type of relationship between the author of the work and the work itself. The authors should still be mindful that the nature of the originality condition differs across jurisdictions.

In the EU, the necessary relationship exists only if the work is the author’s own intellectual creation (see ECR, Infopaq International A/S v Danske Dagblades Forening [2009] ECR I-6569). Regarding more specifically AI-generated creative works, it has recently been decided that EU copyright law does not grant copyright protection to such works. More precisely, the CJEU has ruled that copyrightability does not extend to computer-generated works: copyright protection requires some form of human input because it must reflect the author’s personality. Thus, a creative work that involves assistance from AI is protected, provided that the human input meets the originality requirement.

The baseline of the EU copyright law’s applicability to computer-generated output is that purely AI-generated works or works generated by other automated processes lack any form of human input and are, as such, not eligible for copyright protection: an AI system or any other type of computer programme may be copyrightable, but any output they autonomously create would not be (see, e.g., here, here and here).

In contrast, current UK copyright law protects LDMA works generated entirely by computers, provided that the originality condition is met. Generative AI may thus put the originality requirement under pressure. Recall that, as outlined earlier in this post, machine learning systems are dependent on their training data: for example, an image-making generative AI might create a digital image of an object, but the output would be shaped by images of that object that were created by somebody else and that already existed in its training data.

As the analysis of copyrightability of computer-generated LDMA outputs has shown earlier in this section, at the time of writing it remains untested whether such works generated by AI would undermine claims to originality. After more than two decades of freedom of the internet, with free and open access to large numbers of pictures, photos, videos, and other visual, sound, and/or artistic content (the latter is fully classified earlier in this blog post in regard to LDMA and entrepreneurial works), non-fungible tokens (NFTs) are radically transforming the way these kinds of digital content are understood and regulated.

5. IP and non-fungible tokens (NFTs) 

‘Non-fungible’ refers to anything that is unique and cannot be replaced. ‘Non-fungible tokens’ are defined as assets that have been tokenized via a blockchain by assigning to them unique identification codes and metadata that distinguish them from other tokens. Unique or collectible, original art works themselves or visual representations thereof, photos taken with one’s camera or life stories, they have a value (actually, contrary to fungible tokens, which have value properly speaking, the value of NFTs, similarly to the AI input discussed above, lies in their stored content, and, hence, the data), and thus can be bought and/or exchanged.

For example, one could even tokenize their own image through the process described above (by assigning to it a unique identification code and metadata via a blockchain), and if anyone finds it worth spending their money on it, they will pay for such an NFT. Once rights are attached to that kind of content, NFTs can be traded and exchanged for fiat money, cryptocurrency, or even other NFTs—it will all depend on the value their owners and the market assign to them.

However, NFTs have no value without the rights associated with the content: that is, without a licensing agreement, assignment of IP and/or image rights, and, specifically in the case of NFTs, a contract for the transfer of rights. In a nutshell, NFTs’ essential features are indivisibility, irreplaceability, and uniqueness. Aside from acting as IP, their primary real-life purposes are to serve as academic title, artwork, music composition, gaming, utility, access to a service, e.g., a subscription, and even assets, such as stocks or shares.

Therefore, it is important for startups and tech organizations of any size, as well as innovation project leaders and authors of LDMA and entrepreneurial works, to know and understand their rights, to register and protect any intellectual property as soon as possible, and to design an IP strategy that includes NFTs. Indeed, recall that, in the case of an LDMA work created entirely by an AI, the originality requirement would be limited by the AI’s training data: an AI might create a digital image of a bicycle, but it would be shaped by the images of bicycles created by somebody else and included in its training data. Protecting one’s images of a bicycle or of anything else, particularly in the form of NFTs, involves assigning to them a unique ID and metadata, and, eventually, value.

Note that any crypto token, including cryptocurrency, is built upon the same underlying blockchain technology as are NFTs. Some NFTs are traded on cryptocurrency exchanges and are thus also part of cryptocurrency exchange (the most famous of them being the Doge NFT), but not the other way around. Cryptocurrencies are payment coins that have their own blockchains: Bitcoin (BTC), Ether (ETH), and Litecoin (LTC) are examples of cryptocurrencies that function on their own blockchains. Crypto tokens, on the other hand, are created on blockchains developed by another entity: for example, Chainlink (LINK) and Uniswap (UNI) are tokens developed on the ETH blockchain.

Many NFTs can be purchased only with cryptocurrency, most of them with ETH, which makes the former a kind of a good and the latter a payment method for that good. As was said previously, cryptocurrencies are built using the same kind of programming as NFTs, but that is where the similarity ends. Like fiat (physical) money, cryptocurrencies are fungible because they store value and act as a medium to buy or sell goods. That is already the topic of the next (and last) episode of this blog series: monetary and financial aspects of the regulation of cryptoassets.

6. Summing up

Regarding IP applying to Web 3-4, as discussed in the previous post, the recent agreement on EU AI Act contains prospects for far-reaching rules on IP-related issues. Until the Act is crystallized and becomes effective, the existing rules applying to IP in AI-related business and other concerned elements of the tech world (NFTs, metaverse) have been analysed in this blog post. As outlined, like metaverse itself, IP law applicable to this issue is still in the nascent, if not embryonic, stage, and more scholarly, regulatory, and policy developments are needed, thus representing rich opportunities for development in these regards.

The AI-related IP analysis was divided into ‘hottest’ questions on (i) input data for generative AI and (ii) those on its output data. With respect to the input data, it was established that the most relevant IP rights are those pertaining to copyright and database rights. Organizations offering and/or using generative AI should be mindful that when their AI has been—or might have been—trained on data that is protected by copyright or database rights or both and no exceptions apply, those data must be validly licensed; if they are AI suppliers, they must ensure that actions taken to respect any IP rights concerned are comprehensive and up-to-date; if they are users of such AI, they must seek assurance on this point from the AI supplier.

As to data output by generative AI systems, the two types of IP rights identified as the most relevant are copyright and patents. Regarding the latter, in the EU, for example, inventions generated by AI are not patentable. In 2022, in a related dispute, the European Patent Office decided that a generative AI system cannot be regarded as an inventor. Neither a computer programme that generates an invention without human involvement, nor the owner of such a system, is entitled to a patent. Regarding copyright of the output data, under EU copyright law, works generated purely by AI- or other automated processes lack any form of human input and are, as such, not eligible for copyright protection: AI or any other type of computer programme is copyrightable, but any output they autonomously create is not.

In contrast, for instance, current UK copyright law protects some types of artistic works that are generated entirely by computers, provided that the originality condition is met. Therefore, generative AI may put the copyright originality requirement under pressure across jurisdictions, and this question is an excellent avenue for future research.

Not only does it remain untested whether such works generated by AI would undermine claims to originality, but many of them now also may constitute NFTs—the concept that is radically transforming how IP is understood and regulated and how other legal questions apply to unique, irreplaceable digital content. NFTs are tokenized via blockchain. However, NFTs have no value without the rights attached to the content. Consequently, it is vital for enterprises that are the authors of covered works to know and understand relevant IP rights, to register and protect any intellectual property as soon as possible, and to design IP strategies, including one regarding NFTs.