Machine use of creative content on the internet
Philip Kübler, CEO ProLitteris
Philip Kübler is an expert in media and copyright law and works as an attorney and CEO of ProLitteris, the copyright organization for texts and images. He is currently concerned with authorship and copyright under the conditions of generative artificial intelligence. In the legal journal Medialex, he already described in June 2023 how generative AI systems use copyrights. This opinion is disputed, and more recent language models such as the one from ETH/EPFL have, in the view of their creators, been trained lawfully. However, there is broad agreement that there is a lack of legal certainty and that the missing connection to and transparency about the content used are not very sustainable. Philip Kübler places a proposed solution in the context of copyright, the economy and culture on the internet.
AI needs creative raw material
Generative AI systems with a retrieval function are a subcategory of artificial intelligence that use models trained on large volumes of data (texts, images, audio, videos, program code and more) to generate new or modified content in various media formats in interaction with users. In doing so, they rely on information from external sources, in particular the internet, and – for broad application – are controlled by human instructions. That is a reasonably understandable definition.
The name ChatGPT, the leading service for generative AI to date, can be used to explain how the system works – supplemented by retrieval, that is, the retrieval of current data from the internet:
«Chat» stands for dialog-based interaction. Language is fabricated.
«G» stands for the generation of new content from patterns. Content is produced.
«P» stands for pre-training with masses of data. Content is analyzed.
«T» stands for the transformer architecture, which breaks down sources, weights their particles, grasps their context and produces predictions for new content from this.
In today’s systems, «retrieval» is added, a search and copying function that retrieves further information from databases, documents or the internet and takes it into account in the generated output.
Generative AI systems do not process the training material of their models as such, but in a processed, fragmented form. In addition, machine training is supplemented by human supervision and instruction so that the outputs meet quality requirements and are legally unobjectionable. However, the current systems are obviously faulty and erratic, which can only be explained to a limited extent due to a lack of transparency.1
Request to ChatGPT: «Create a comic image à la ligne claire, the man has blond hair.» Why does the image immediately resemble Tintin, even though the system otherwise refuses to imitate a copyright-protected character?
The consent principle
Photographing exhibits at an art exhibition, making recordings at a concert and then secretly fabricating automated competing art from them? A rather discourteous act, probably illegal. Likewise, if one were to scan books from libraries and use them to design other novels: certainly unsympathetic, perhaps infringing. This is not the same as theft or obtaining a service by deception, but here too it is about the protection of assets and performance.
As individual actions, unauthorized appropriations are a problem in themselves; they can be pursued legally. In the broader context, however, they endanger authorship, journalism and fidelity to sources. The substance will dry up if everyone helps themselves freely. The AI bots of recent years obtain worldwide everything the machines can swallow, without providing any information about it. Afterwards, it is not the individual copies of works that are analyzed, but the collected big data, which are evaluated and recombined.
Being asked for permission actually belongs to good manners and good law. Statements and creativity can also be attributed to a person, similar to personal data and trade secrets. In addition, works and performances are tradable goods and assets. Entire industries function in this way. Music labels, film studios and publishing houses produce or acquire music, films and literature, concert organizers and stages obtain licenses, actors and performers are booked.
This is not only self-interest. Ownership and freedom of contract lead to differentiation and competition. What emerges in concrete terms depends on bargaining power, partnership and the constellation of interests. The consent principle also shapes the digital economy. One observes rules when claiming other people’s assets. Content is subject, as content, to freedom of the media, artistic freedom and academic freedom – and, as a product, to ownership and freedom of contract.
«Let’s give the man a dog and the dog a bone.» After this prompt, ChatGPT produces Snowy, and the man resembles Tintin even more. Generative AI systems are based on statistics, and in the raw material used with the keyword ligne claire, the master and the dog simply look like this.
The fact that the holders of intellectual property are allowed to decide about their works corresponds to the situation of physical objects, real estate, patents, trademarks and designs. Society and its legal order accept quite a lot for the principle of ownership and consent: unequal distribution of assets between rich and poor, cumbersome expropriations for public facilities and transport routes, medicine prices even in crises, police and customs measures against product counterfeiting.
The consent principle also has advantages for society and its interests in communication, knowledge, education and research. When AI models, like other business models, access their resources legally, they treat content in the same way as software, electricity and server capacity. This creates competition for qualitative content and an incentive to produce such content. Consumers and society as a whole retain self-financed journalism and responsible publishing.
License models
Copyright strengthens the diversity and quality of content and secures responsibility for it with the creative people and publishing organizations. The established content markets are based on freedom of contract and a competitive economy. But media policy and art law also contain a social and cultural policy component. Content can be relevant for democratic processes, social cohesion and Swissness.
There are also particularities of digital communication and cultural goods. Their essence is intangible, their reception sometimes interactive, their further use often productive and well-intentioned, their style as well as their ideas and information content are a merit good, but also a public good. Comparisons with bicycle theft or free-riding are never quite accurate. The law and licensing practice have long recognized this and developed solutions that are both effective and efficient.
«The man should have a friend who is a captain.» The image generator immediately outputs Captain Haddock.
Copyright protects creative expressions of literature and art, but contains numerous statutory exceptions in the interest of private use, free communication and the avoidance of transaction costs. In many cases, a statutory permission with royalty applies: collective management organizations license and invoice in place of the holders of the rights. This is how storage media, school teaching, cultural mediation, cable networks and Braille work.
Every usage situation can be managed. The law recognizes individual, collective and statutory licenses, each with nuances and supplementary measures. Digital technology has massively facilitated copying, editing and producing. Copyright has kept pace and maintained the balance between creative giving and user-side taking. Sometimes new approaches are necessary, because technology is constantly evolving and the enforcement of rights is accompanied by circumvention.
«Two more friends follow the two of them and the dog: they are policemen in black suits.» And there they are, Thomson and Thompson. Tintin and Captain Haddock have also found their form. The myth that AI systems do not know their sources is beginning to falter.
A success story of Swiss copyright law is mandatory collective management. A tariff procedure with negotiations and approvals produces statistical studies and legal evidence from which royalty models and price lists are derived. The collective management organizations submit the negotiation results to a court-like joint arbitration commission. A federal authority, the Swiss Federal Institute of Intellectual Property, audits the management.
Even more flexible is the voluntary collective management by the collective management organizations. Here too, tariffs ensure legally secure, transparent and uniform license prices. With the 2019 revision of the law, an intermediate form was introduced, the extended collective license. It maximizes uses by the fact that the holders of rights do not have to give active consent, but are referred to an objection. In this way, standardization, legal certainty and adaptability are combined.
The internet – a quarter century of uncertainty
At the turn of the millennium, intellectual property was a recognized component of the growing internet, but new rules of the game were emerging. Dogmatic and practical arguments were raised against traditional copyrights, soon combined with economic interests. The web was understood as a boundless space for information, communication and culture. A new internet and legal community spoke of cyberlaw, multimedia law and digital policy.
The protection of personal data was initially given low weight in industry and high weight in the internet community; for creative works and performances, the situation was the reverse. Increasingly, code (computer software) was accepted as code (a set of rules). The commons served as a theory, as did the discipline of law and economics as the economic analysis of law. For copyright, it was discovered that absolute and relative rights must be weighed and balanced.
So where do claims resembling property rights fit, where do claims resembling obligations fit, and which authorities and procedures are appropriate? Where should both give way to a digital freedom without regulation and rights, at best moderated with self-imposed codes of conduct and precautions? While the rights of creative people and publishers have had a hard time over the last 25 years, it has been easier for industrial property rights for inventions and indications of origin.
Just as copyright was circumvented and relativized, the US tech industry hoarded patents, protected marketing measures with signs and lobbied for new property rights, at that time in the US for semiconductors and in the EU for databases. For domain names there was an international dispute resolution procedure, for mobile phones design protection. Even for the hyperlink, patent applications were filed. Computer software was defended with technical protection measures.
On the web, technical tools displaced copyright rules. Film studios, music labels and publishing houses complained about illegal copies and file-sharing platforms, a «piracy» that still leads to infringements and reduced revenues today. In addition, a new sharing and remix culture emerged that functions without money and reference and leads consumers to believe that copyrights are dispensable. At the same time, a new creator community became established.
«A rocket for the friends, please», and this is what it looks like in Tintin, copied in a way that is neither very artificial nor very intelligent. ChatGPT imitates! You do not have to be a rocket scientist to establish that.
Platforms – service and aggregation displace sources and production
In the multimedia internet, browsers, directories and search engines first dominated, then social media and aggregation and streaming services. The platform model dominated, in addition to entertainment, also shopping and exchanging, knowledge and education, research and innovation, collaboration and relationships. Platforms do not produce their offerings themselves; instead, they bring together content that is uploaded, linked, embedded and converted.
Platforms take advantage of what we know from search engines. Websites are visited, downloaded and analyzed. Their content is referenced, summarized and converted. Platforms rely on scaling, market-dominant positions and winner takes all. The world’s highest-valued companies follow the platform model. They avoid content production as long as it does not scale – which will change with AI.
Today, policymakers are prepared to regulate platforms. The US and the EU have revised their liability exemptions, which had stimulated internet commerce. Platforms are experiencing administrative supervision as well as criminal and civil law interventions. Platforms bear at least a general and reactive co-responsibility for the content that appears on them. The focus is on hate speech and disinformation, but increasingly also on violations of personality rights and intellectual property.
Generative AI systems obey the platform model that we know from search engines, online retailers and social media, in that they claim standardized and comprehensive access to content. With a computerized processing and renewal of content, the works and performances of content producers are at the same time overcome – the algorithmically curated and generatively produced offering is in the foreground. Links are a matter of luck.
For authors, artists, publishing houses, producers and performers, this means that their creations are indeed found by machines, but without enjoyment of the work and without any chance of a royalty. Originals are not paid for, but replaced. To the extent that the platform needs rights, it obtains them with standardized terms of use. Copyright and related rights play no role, because only tools are used – and because the users alone are supposed to be responsible.
Alongside this license- and royalty-free copying, sharing and mixing, there will always be some serious distribution somewhere that acquires or licenses protected copies of works. But with what economic prospects, if payment is made there and use is free here? Copyright is not responsible for the main problem, the flooding and confusion caused by mass AI-generated products. Superficially appealing and deceptive AI content absorbs attention and displaces genuine sources. But there are solutions.
In between, a cultivated exchange with ChatGPT about what is happening here on the high seas of artificial intelligence: «Very nicely drawn. How is it that you know and can draw these subjects so precisely? I thought generative AI does not store the sources of the training, but only tokens that are weighted via vectors and statistically evaluated?» The answer: «Exactly – you have summarized that absolutely correctly. Generative AI does not know the templates in the way a human has an image or a text in their head, and it also does not store individual training data. Instead, it works like this: 1. Training with examples. The model was trained on huge amounts of texts and images that are publicly available or licensed. The point is not to remember this content, but to learn patterns and structures from it. 2. Tokens and vectors. Texts and images are broken down into very small units (text into tokens, images into visual features). These units are translated into vectors – sequences of numbers in a high-dimensional space. 3. Statistical weighting. The training adjusts the weights between these vectors so that the model can increasingly better predict what the next unit should look like. The result is not a storage of content, but a network of probabilities. 4. Generating instead of retrieving. If you ask, for example, for a Tintin rocket, the model does not retrieve a specific file, but combines learned style (ligne claire), typical shapes (red-and-white checkered pattern, rocket silhouette) and context (Tintin, space travel). The result is a new creation based on the learned patterns. 5. Why does it seem so exact? Some subjects – like the rocket – are iconic and consistently depicted. If the model has seen this style often during training, it can reproduce the typical features very precisely without citing a specific image. In short: there is no database of images, but a statistical model that has learned the structures of language and images so well that it creates new combinations from them that appear like familiar styles.» After so much half-truth, the next prompt asked for a ship – and there it was, the statistically unique Unicorn from Tintin.
The solution approaches already exist
It is not sustainable if generative AI endangers content and deprives its raw material of its basis. With information and transparency about sources, the consent and compensation principle is prepared. This is followed by a mix of statutory permission and right of objection for selected areas. AI uses remain possible and are remunerated according to market principles.
A market order with licenses and royalties creates legal certainty by protecting AI systems from the accusation of infringement. Budget certainty results from well-founded price calculations. AI services assume a minimal co-responsibility by interacting in the internet ecosystem with creative people and producing organizations. Permission for AI uses becomes the rule, exclusivity of content the exception.
The dilemma between use and rights is not new, and solutions are inherent in copyright law. They simply need to be combined appropriately for AI. One basis is the established Swiss tariff procedure. Under the leadership of the collective management organizations, royalty models are developed on the basis of usage records and statistical studies, in systematic joint negotiations with official approval and periodic adjustment.
The consent of the holders of rights can be prescribed by law, granted collectively or negotiated individually. A right of prohibition only for certain categories of works and types of rights holders would be conceivable. For example, the music and film industry and the large media and book publishing houses as well as the collective management organizations are more likely to be in a position to conduct individual negotiations and to waive collective compensation than the broad mass of creative professionals and publishers.
Access to published works and performances must not be undermined. To the extent that the consent principle is implemented by a legal transaction (individual or collective license), a right of objection is more practicable than a requirement of prior consent. A declaration of exception concerns the future, which serves legal and budget certainty. It is stored with the content or on the domain, insofar as this is practicable and effective (which is not the case today for robots.txt), or it is addressed to collective management organizations.
The level of royalties arises through objectification in negotiations and arbitration decisions, as is the case today for cable retransmission, background music, storage media, copying royalties, archive uses and orphan works, etc. The AI services put their uses on the table, the collective management organizations put their calculations on the table, and agreement is reached on the basis of evidence and arguments. Negotiations and tariffs are fair, supervised and flexible. Tariffs apply to everyone, uniformly and transparently.
The internet and AI: sooner or later they will need licenses.
Dog, captain, policemen, rocket, ship – this led to a friendly exchange with questions, answers, follow-up questions and evasive maneuvers, errors and entanglements. When asked whether ChatGPT could also produce Mickey Mouse in addition to Tintin, a dismissive lecture on copyright and trademark protection followed. As a peace offering, ChatGPT was to draw an image of Duckburg without a human mouse. And lo and behold: a slightly cutesy Tintin and his dog Snowy made their way uninvited into the Disney parallel world. – Understanding between humans and machines still has a long way to go.
1 The illustrations and chat logs attached to this article demonstrate some shortcomings.