II. ChatGPT: the AI language model revolutionising content creation

You may have heard about ChatGPT, which attracted millions of users quickly after its launch. OpenAI released ChatGPT as a chatbotFootnote ¹⁷ designed primarily to assist with online customer care. Yet, the AI technology underpinning this program has outperformed all forecasts and grown in popularity.

What is the definition of ChatGPT? We have chosen to ask the question directly to the chatbot:

The ChatGPT software application was designed to facilitate and support interactive communication by generating automatic text-based responses. The methodology involves integrating information from various sources, such as books, journals, websites and articles, resulting in the development of original and engaging discourse. The field of natural language processing (NLP) involves the use of computational methods to facilitate the understanding and analysis of large amounts of data in a way that mimics human language. Computer systems are capable of acquiring significant amounts of data or executing commands by utilising NLP functionalities.Footnote ¹⁸ It is worth noting that the operational functionalities of ChatGPT are limited to the understanding of syntax. The employed methodology involves a comprehensive analysis of the sentence’s syntactic structure, along with the identification of the interdependencies among its constituent elements. Although ChatGPT has undergone comprehensive training in handling vast datasets, it continues to depend on statistical techniques for analysis and comprehension.Footnote ¹⁹ The aforementioned statement implies that the system possesses the ability to identify patterns in the utilisation of specific terminologies and idiomatic expressions and employs these patterns to generate informed assumptions regarding the intended interpretation of a provided sentence or phrase. It is crucial to recognise that this statement precludes any inferences about ChatGPT’s understanding of the aforementioned principles. In the event that an individual were to inquire of ChatGPT regarding the essence of existence, the response provided may appear to be lucid and perceptive; however, it may lack a comprehensive comprehension of the philosophical principles and hypotheses that underlie the inquiry. ChatGPT is just an advanced chatbot that employs NLP to comprehend vast quantities of information and produce responses that closely resemble human language. Its capabilities are currently limited to syntactic understanding, meaning that it can analyse the grammatical structure of sentences and comprehend how words and phrases relate to each other within a sentence. Although ChatGPT is able to detect patterns in the use of words and phrases in context and to use this information to make educated guesses about the meaning of a sentence or phrase, it still approaches semantic understanding through statistical analysis.Footnote ²⁰ For this reason, ChatGPT’s responses do not always reflect true understanding of the underlying concepts or theories related to a particular question. ChatGPT is capable of independently stimulating dialogues and thus has the potential to produce content protected by intellectual property, including but not limited to articles, music lyrics, programming codes and text translations. The results produced by ChatGPT depend on the data it was programmed with and the computational techniques used and may not always be suitable for all targets. While acknowledging the commendable nature of ChatGPT, it is important to point out that its AI-powered functions are created without human intervention. Even though AI has remarkable precision, it is not free of limitations. Therefore, it is important that individuals – in order to prevent potential problems or errors – review and modify the designs to ensure that they meet established standards for accuracy and efficiency for specific usage scenarios.

III. An analysis of the ChatGPT case study: the question of the output

The emergence and widespread application of AI systems in the creative sectors have raised concerns about the rightful ownership of intellectual property and the protection of copyright. To gain a more comprehensive understanding of the above issues, our research has specifically focused on ChatGPT, a language acquisition model developed by OpenAI. ChatGPT – as mentioned earlier – is a generative AI tool. A full understanding of the complexity of the evolution and ownership of the ideas created by AI can be achieved by evaluating the results and sharing observations about the interaction between humans and computer systems. This emphasises the importance of a recasting of legal rules in order to adequately deal with this set of challenges. In this perspective, the ChatGPT case study can be used to initiate a debate about the legal and ethical implications of rapid technological advancements, as well as the concerns associated with the application of AI in the creative industries. The use of generative techniques such as ChatGPT raises significant issues regarding intellectual property, authorship and the scope of copyright protection for material created with generative AI systems. These aspects have attracted considerable attention in the legal field.Footnote ²¹ The primary inquiry relates to the need to determine the rightful owner of copyright in content generated by AI, whether a natural person or a legal entity. We attempted to ask ChatGPT directly about this question, and the platform provided us with the following response:

Of course, the question of who owns the content developed by ChatGPT is more complicated and may require further clarification or reference to be fully answered. Because ChatGPT is an AI system that generates results based on training data and user input, it is difficult to identify specific authors. Authorship of AI-generated content may depend on a variety of factors, including the purpose of the content, the intent of the user and the legal framework at the time.

In the case of ChatGPT, the author of the content may be based around the individual who created the prompt or input for the response. If the user provides the input, they can take ownership of the output. If the input comes from ChatGPT’s training data or other sources, it may be more difficult to identify the owner. Usually, the copyright owner of texts created with tools such as ChatGPT is the person or organisation that provided the original ideas and data on which the system is based or the person who creatively implemented the instructions in the prompt.

As an AI language model, the text generated is not, per se, protected by copyright law, as copyright law generally recognises the human creator of an original work as the copyright owner. However, in some cases, the text generated may be considered original enough to be protected by copyright if it was created with sufficient human input or intervention. For example, the resulting work could be deemed sufficiently unique to be protected by copyright law if someone uses the replies as a starting point and then adds significant creative or original content, such as editing, adding commentary or analysis or merging it into a bigger work. The individual who added the extra creative or original content in this scenario would normally own the copyright for the final product.

AI’s inability to hold copyright stems from its legal identity or status as a non-human entity.Footnote ²³ While the Berne Convention and other international copyright regulations do not require human authorship, many countries, such as the USA and those in the EU, place importance on the presence of a human being as the creator of a work.Footnote ²⁴ In addition, copyright law itself adopts a predominantly anthropocentric approach, as exemplified by the copyright term “70 years after the calendar year in which the author of the work died”. This term inherently assumes that the author is a human being, subject to mortality.

Original pieces of art produced with AI assistance or by automated means are not novel occurrences. Some might argue that what we are witnessing with AI systems is simply the repetition of history. After all, copyright laws have always had to evolve and keep pace with emerging technologies and their effects on society. This pattern can be observed, for instance, with the arrival of photography, motion pictures, computer programs and various other novel forms of creative expression.Footnote ²⁵ A historical example of this is the well-known case in Burrow-Giles v. Sarony in 1885,Footnote ²⁶ which was argued before the US Supreme Court. At issue was whether a photograph could be considered a copyrighted work, given that the image was created by a camera rather than a human being. In its ruling, the Court held that the photographer, who was the person behind the camera, was the author of the photograph and therefore had exclusive copyright over it. This rationale persisted even in cases where the machine was responsible for most of the work, since it was recognised that human input was necessary and indispensable for the creation of the work.Footnote ²⁷ The UK is the only country where the concept of “computer-generated works” is recognised in domestic law, and it seeks to address the issue in a practical way by expanding the concept of authorship. Section 9(3) of the Copyright, Designs and Patents Act 1988 (CDPA) provides that the person who makes the necessary arrangements for the creation of the work shall be deemed to be the author.Footnote ²⁸ However, determining who is the “arranger” is not always easy and often needs to be determined on a case-by-case basis. In particular, according to contemporary standards for defining “computer-generated” works, the UK approach can be considered quite outdated.Footnote ²⁹ These provisions became law in 1988, and the AI systems available today are vastly different from the computer systems that existed at that time. Moreover, given the complexity of modern AI programming, there is considerable uncertainty in determining which party is responsible for the “arrangements necessary for the creation of the work”.Footnote ³⁰

Being an AI language model, ChatGPT also lacks legal identity and the ability to possess property or assets in the conventional sense because it is not a human. Even if the content created by an AI language model is original and creative enough to be protected by copyright law, the AI will not own it. According to the various jurisdictions mentioned, the copyright for the material could belong to the individual or entity that has legal authority over the AI, such as the AI system’s developer or owner. In some instances, the content’s copyright may belong to the human users who contributed to or edited the AI-generated work.

The practical approach, then, is to grant copyright to the people behind the machines, namely the programmer, the user and the owner. These key actors are the human or human-owned entities behind the process of AI production and, accordingly, the actors at the centre of the legal discussion about copyright in AI-generated works. On the other hand, if we consider only the statutory and common law understanding of the doctrine of originality and the requirement of human authorship, there is indeed no copyright in works created by AI, and copyright-free works naturally belong to the public domain.

IV. An analysis of the ChatGPT case study: the question of originality

The issue of authenticity of the output of AI generative systems presents another challenge and variation, especially with respect to ChatGPT. While chatbots are excellent at generating responses that engage humans in conversation, these responses run the risk of being unoriginal, completely invented and simply repeating information from the past. The use of chatbots and generative tools for content creation can lead to problems, especially in cases where the resulting output requires distinctiveness and appeal. Intellectual property ownership is closely related to the uniqueness of AI chatbots. As AI production processes become more sophisticated and generate content that closely resembles human-created content, it is important to establish clear guidelines and rules for creation and submission. The significance of this matter is particularly salient in domains such as journalism and creative writing, given that the utilisation of AI-generated content can engender ethical and legal dilemmas. Typically, the degree to which creative works are safeguarded is contingent upon their level of uniqueness. While the Berne Convention does not expressly mandate the condition of “originality” for copyrighted works, several nations enforce this prerequisite.Footnote ³¹ As a consequence, the originality requirement is a prerequisite for the granting of copyright protection to literary, dramatic, musical and artistic works. Currently, the prevailing approach to determining whether a work is original is by evaluating whether it is “the author’s own intellectual creation”.Footnote ³² This means that the work must have an intellectual content that goes beyond the mere combination of its individual parts, taking into account the overall impression. However, the standard for originality varies across different jurisdictions. For instance, US law adopts the “minimal degree of creativity” test, which was established in the Feist v. Rural case,Footnote ³³ while the EU requires that the work be an author’s own “intellectual creation”.Footnote ³⁴ But the issue of originality becomes more nuanced when considering content created by AI. Copyright laws may apply to certain AI-generated material if it is made with sufficient human input or participation. An AI chatbot may be regarded to have created an original work under copyright law if a human offers input or instruction to the bot to create a particular work, such as a tale or a song, and the bot then develops the final output based on that input. Some AI-generated content, however, may be less unique than expected and more derivative or based on previously published works. For example, if AI-generated systems merely reproduce data or existing information without adding significant ideas or original content, the output they create cannot be considered unique enough to be copyright protected. By applying the US “minimal degree of creativity” test for originality, which sets a low bar for copyright protection, one could argue that ChatGPT’s output meets this standard. This is because ChatGPT utilises sophisticated NLP techniques to generate text that is not merely a repetition of its input data, indicating some level of creativity. However, under the EU’s standard for originality, AI-generated works may not qualify, as they lack the creative choices and personal expression of a human author.

ChatGPT, as an AI language model, has undergone extensive training using vast amounts of textual data gathered from diverse online and offline sources. By leveraging these training data, it generates responses to user queries encompassing a broad spectrum of subjects, ranging from common knowledge to specialised fields. When presented with a question, ChatGPT examines the overall context and relevant keywords to formulate a response based on learned relationships and patterns derived from the training data. These responses are algorithmically generated and do not rely on the respondent’s personal opinions or experiences. It is important to note again that ChatGPT establishes statistical correlations between words without genuine comprehension of the underlying meaning. The tool excels at producing high-quality written content across various domains, saving considerable time compared to human effort, thanks to its extensive database and syntactic correlation capabilities. However, due to the absence of human authorship, ChatGPT lacks the necessary human creative input required to substantiate a copyright claim.Footnote ³⁵ In copyright law, in fact, the act of creating a copyrightable work is typically associated with human creativity and authorship. Therefore, if there is no human involvement in the creation process, there is a lack of originality, and consequently the work may not be eligible for copyright protection. The concept of personhood is crucial in this context, as it distinguishes between entities with naturalistic dimensions of life and self-awareness and those that do not possess these attributes. Robots or other AI technologies, regardless of their level of autonomy, cannot be considered as persons under ethical and legal frameworks. The qualification of personhood plays a significant role in copyright law as it serves as a boundary for attributing creative authorship and the associated rights. This distinction is based on the understanding that copyright protection is intended to incentivise and reward the unique and subjective contributions of human creators.

While robots and AI systems can generate content or imitate human-like behaviours, they lack the essential qualities that define personhood, such as consciousness, intentionality and the capacity for subjective experience. These intrinsic limitations prevent us from assimilating them into the category of persons within ethical and legal contexts.

Therefore, it remains firmly established that copyright protection requires a human element, where the creative efforts and expression originate from individuals possessing the characteristics and attributes inherent to personhood.

In light of this analysis, it can be deduced that in order for an AI system to fully replace a human author, it would require the capacity to independently conceptualise and complete a creative work without relying on explicit training or pre-programmed instructions.Footnote ³⁶ As technology continues to progress, it is reasonable to envisage a gradual reduction in human involvement in the creative process, leading to the emergence of new artistic creations that cannot be attributed to a specific or recognised artist.Footnote ³⁷ While the current capabilities of ChatGPT may not align with this vision, a completely revolutionary future appears to be within reach.

V. An analysis of the ChatGPT case study: the question of the input

At present, the intellectual property discourse pertaining to AI predominantly revolves around the issues regarding the authorship and creative ownership of the outcomes generated by AI systems. Despite the ongoing discourse, there appears to be a significant gap in the examination of the legal questions that emerge in the management of intellectual property rights pertaining to the inputs, namely the data employed in the training of these AI systems.Footnote ³⁸ The second and more fundamental question to be addressed here is whether the use of copyrighted material to train generative AI programs represents an infringement of copyright. Indeed, machine learning heavily relies on vast amounts of training data to achieve accurate results, including in facial recognition, stop sign recognition, natural language recognition and translation generation. This is especially important when it comes to ChatGPT because it relies on large amounts of training data being fed into the system.Footnote ³⁹ In order to create interactive and authentic articles, ChatGPT needs to ingest information, including text, images and other content, from publicly available websites on the Internet.

To facilitate the training of AI algorithms, various techniques are used, including text and data mining (TDM)Footnote ⁴⁰ as well as generative deep learning techniques.Footnote ⁴¹ TDM processes involve the extraction and analysis of vast amounts of data to identify meaningful insights and patterns, which can then be leveraged to improve the performance of AI models.Footnote ⁴² TDM has become an essential tool in the field of AI, enabling researchers and data scientists to explore vast amounts of unstructured data and extract valuable information that would otherwise be impossible to obtain manually. By analysing these vast amounts of data, AI algorithms can learn from these patterns and make predictions with a high degree of accuracy, facilitating the creation of content, discoveries and innovations. So, without access to large volumes of data, AI algorithms would struggle to “learn” and improve their performance. Therefore, it is clear that the future of AI hinges on TDM and its capacity to extract and analyse data on a large scale. However, a significant challenge lies in the fact that AI systems cannot learn from art in the same way humans do, since they require an exact copy of the artwork in their training dataset.Footnote ⁴³ This necessitates the creation of a training set of millions of examples by making copies of copyrighted images, videos, audio or text-based works. Consequently, the question of whether machine copying should fall under fair use or other copyright exceptions arises. On the other hand, we have generative deep learning, a specialised branch of deep learning that focuses primarily on the task of generating novel data.Footnote ⁴⁴ Generative models are crucial in this domain as they provide a probabilistic framework for describing the data generation process.Footnote ⁴⁵ By harnessing these models, it becomes possible to generate new data samples through the process of sampling. These techniques employ deep neural networks, which are artificial neural networks with multiple layers, to learn and replicate the patterns, structures and statistical properties present in the training data. ChatGPT is precisely a form of generative deep learning technique that harnesses the power of deep learning models, particularly the GPT (Generative Pre-trained Transformer) architecture.

The problem here is that established companies such as Google, Facebook, Amazon and OpenAI have access to large collections of language and image data, which they can use for AI purposes.Footnote ⁴⁶ Access to large collections of language and image data can be also considered a competitive advantage in the field of AI.Footnote ⁴⁷ As a consequence, these companies can leverage their existing datasets to train and develop more advanced AI models, which in turn can improve their products and services. This can create a legal problem for new entrants because the ownership and licensing of datasets can be complex and subject to intellectual property rights, privacy regulations and other legal considerations.Footnote ⁴⁸ Additionally, the cost of building or licensing a dataset from scratch can be prohibitive, making it difficult for smaller companies to compete with established players.

Moreover, there may also be antitrust concerns if the dominant players in the market control access to the datasets needed to develop AI models, as this could potentially stifle innovation and competition. Therefore, ensuring fair and open access to training data is a critical legal issue in the development and deployment of AI technology.

Another issue with input data is that while some large datasets are merely informational and not protectable, the majority of training datasets consist of copyrighted works. For instance, the corpus of works used to develop AI algorithms for text, facial recognition and image recognition all include copyrighted works. Thus, the question arises as to whether using these works is lawful and under what circumstances.

Currently, data collection for TDM has been considered fair use in the USA,Footnote ⁴⁹ and there are exceptions and limitations under EU copyright law.Footnote ⁵⁰ Specifically, in the USA, Google Books was granted permission to search entire libraries to provide search functions and excerpts from books.Footnote ⁵¹ However, it is unclear whether these conclusions apply to data collection and input for machine learning, as there is no copyrightable output. Indeed, it cannot be guaranteed that courts will apply this precedent to comparable technologies.Footnote ⁵² In the USA, data collection for TDM may be permissible if it is a transformative use,Footnote ⁵³ but it is not immediately clear that a copyrighted work is being transformed into another copyrighted work. In addition, in the Google Books case, the court recognised that Google’s digitisation of copyrighted books, undertaken for the purpose of creating an extensive index and facilitating search functionality, constituted fair use.Footnote ⁵⁴ This digitisation process was specifically designed to enhance users’ ability to locate and access copyright owners’ books, providing an invaluable tool for researchers, scholars and the general public. In this context, it is important to note that Google did not intend to compete with or replace the original works, but rather to improve their discoverability and enable consumers to make informed decisions about purchasing or accessing the entire works. On the other hand, when we examine generative AI technology, we encounter a contrasting scenario. Generative AI systems have the potential to empower users to easily produce content that may directly compete with the original ingested material. These systems utilise algorithms and machine learning techniques to generate new works, such as texts, images or music, based on the patterns and information gathered from existing content. Unlike Google’s indexing and search functionality, which primarily served as a tool for information retrieval, generative AI opens the door for the creation of derivative works that could potentially overshadow or undermine the market for the original content. So, while Google’s efforts in the Google Books case were found to align with fair use principles, the ease and accessibility of generative AI introduce complexities and challenges regarding copyright protection. Exactly for this reason, numerous court cases are currently underway in the USA seeking to clarify the definition of a “derivative work” and “transformative use” under intellectual property law, particularly with respect to copyrighted material used to train AI systems.Footnote ⁵⁵ In particular, OpenAI and other prominent generative AI platforms are currently facing lawsuits alleging copyright infringement for training AI systems with illegally acquired datasets.Footnote ⁵⁶ Specifically, in the legal case of Tremblay v. OpenAI Inc.,Footnote ⁵⁷ the plaintiffs assert that OpenAI employed their copyrighted books without obtaining proper authorisation in order to train ChatGPT. The assertion is made that ChatGPT possesses the ability to effectively condense the content of various books, thereby implying that the chatbot has comprehensively engaged with and assimilated the information contained within said literary works. In the case of Silverman et al. v. OpenAI Inc., the plaintiffs assert that OpenAI engaged in unauthorised utilisation of copyrighted work, specifically the book titled The Bedwetter, for the purpose of training ChatGPT.Footnote ⁵⁸ Specifically, the authors of this class action claim that ChatGPT is capable of producing summaries of their novels when provided with a suitable prompt. They base this claim on the fact that the AI tool has been trained using their copyrighted material, thereby establishing its familiarity with the content. Finally, in the dispute Getty Images Inc. v. Stability AI, the famous photo agency alleges that the software developer responsible for the AI art tool known as Stable Diffusion engaged in the unauthorised scraping of a substantial number of its images.Footnote ⁵⁹ This act was purportedly carried out for the purpose of training the aforementioned system without obtaining proper permission or providing compensation to Getty Images. In addition, the AI tool Stable Diffusion generated a modified rendition of Getty’s watermark, with the purpose of promoting, facilitating or concealing the infringement of Getty Images’ copyright. This action – according to the plaintiff – also constitutes a violation of the Digital Millennium Copyright Act (DMCA) regulations regarding copyright management information.Footnote ⁶⁰ Getty Images has also filed a similar complaint in the UK, requesting the High Court of London to issue an injunction barring Stability AI from selling its AI image generation technology in the country.Footnote ⁶¹

The resolutions of all of these cases remain pending, and the manner in which they will be resolved remains uncertain at present. Nevertheless, these cases mark the initial significant legal confrontations regarding the utilisation of AI in relation to copyright violation. If the plaintiffs achieve a favourable outcome, they have the potential to exert a substantial influence on the advancement of AI technology.

However, the US Supreme Court’s recent ruling in a non-technological case has already raised concerns about potential adverse implications on the intellectual property rights of works generated by AI.Footnote ⁶² This case seems to have shifted the focus of the transformative use assessment. The controversy pertains to a conflict concerning copyright infringement, specifically regarding the utilisation of a photograph featuring the musician Prince that was taken in 1981.Footnote ⁶³ The photograph was subsequently incorporated by the artist Andy Warhol in a series of prints and illustrations without obtaining the photographer’s authorisation.Footnote ⁶⁴ The fair use doctrine was invoked by the Andy Warhol Foundation for the Visual Arts to justify the creation of derivative works. In this context, the US Supreme Court ruled that the Foundation lacked a fair use defence to license a derivative rendition of the photograph for commercial purposes.Footnote ⁶⁵ This recent decision could potentially result in a significant restriction of the transformative use doctrine, given that the Supreme Court appears to have effectively limited its scope.Footnote ⁶⁶ So, it will be interesting to see what happens when US courts have to use the rules set up in this case to judge the licensing of AI training input. In the event that a court determines that data ingestion – which involves acquiring unprocessed data from one or more sources and modifying them to render them appropriate for the purpose of training AI machines – constitutes an act of infringement, the entire AI system may encounter significant legal difficulties. In fact, the vast majority of data that generative AI systems have assimilated – including both textual and visual content – have been de facto obtained without the express authorisation of the rights holders. Consequently, here the question at hand pertains to the potential copyright infringement that may arise from utilising copyrighted works as training data. Specifically, it is necessary to determine whether such usage automatically constitutes copyright infringement or whether the distinct purpose of training data sufficiently diverges from that of the original copyrighted works, thereby warranting a fair use defence.

In contrast to the USA, the EU adopts a protectionist stance and has established a degree of accountability for the utilisation of training data. Specifically, the Directive on Copyright in the Digital Single Market (CDSM Directive)Footnote ⁶⁷ includes Article 4(1), which provides a broad exception for TDM. Under this provision, individuals such as commercial AI system developers and educators may make copies of works or databases for the purpose of extracting information from text and data. They may retain these copies for as long as they are needed for the AI training process.Footnote ⁶⁸ However, rights holders have the option to exclude TDM exemptions from their contracts with miners (ie entities or individuals that engage in TDM activities) in order to safeguard their commercial interests.Footnote ⁶⁹ This particular provision has met with considerable criticism for providing a copyright exception that is perceived as being too restrictive. In contrast to the traditional understanding of copyright, which generally focuses on the protection of original expression, this provision appears to include factual information and data, and this aspect has drawn much criticism.Footnote ⁷⁰ However, the manner in which this opt-out option can be implemented and the extent to which AI developers will adhere to it are still to be determined.

An additional issue associated with data aggregation pertains to the implementation of EU data protection legislation.Footnote ⁷¹ Indeed, the process of data aggregation is of paramount importance in the training and refinement of generative AI models. This entails the gathering and merging of substantial quantities of data from diverse origins to augment a model’s proficiency and functionalities. The processing of personal data within the EU is subject to stringent requirements and limitations, as stipulated by the General Data Protection Regulation (GDPR).Footnote ⁷² These challenges remain unexplored in both doctrine and policy and need to be further explored and resolved.

VI. Source attribution and other copyright challenges in language models

ChatGPT’s input as a language model comes from a variety of sources, including books, essays, websites and social media posts. These sources may contain copyrighted works that are used to train the language processing algorithms in ChatGPT. Given the legal concerns surrounding copyright and the use of training data for machine learning, it is likely that ChatGPT faces similar issues. The dilemma, as with other AI systems, is whether using copyrighted material to train ChatGPT’s language processing algorithms is legal and under what conditions. Because ChatGPT and other generative AI systems rely heavily on large amounts of training data, which may include copyrighted works, this presents a significant legal hurdle. The input data utilised by ChatGPT are produced via a method referred to as “training”. During the training phase, the model is presented with a vast corpus of textual data, which are employed to instruct the speech-processing algorithms. The corpus under consideration exhibits the capacity to encompass a diverse range of text-based sources, including but not limited to books, articles, websites, social media posts and analogous materials. The type of data employed to furnish instructions to ChatGPT is dependent on the particular task or use case for which the model has been trained. If ChatGPT is directed to address customer service inquiries in a particular language, the training data corpus employed could be sourced from transcriptions of customer conversations or online evaluations. To guarantee adherence to copyright regulations, it is important to acquire any information utilised for the dissemination of ChatGPT through lawful channels. This may involve obtaining permission to use copyrighted materials or accessing publicly available information. Under specific circumstances, fair use or other legal exemptions may be relevant. However, it is essential to note that this is a multifaceted and dynamic field of law that necessitates meticulous examination on a per-case basis.

It is relevant to bear in mind that the programmers accountable for the development and training of ChatGPT hold the responsibility for ensuring that the training data remain free from any copyright violations. The provision of a comprehensive list of data sources may not be practical; however, OpenAI could explore more transparent options for disclosing the origins of the training data it employs. This could involve specifying the sources utilised or outlining the methodologies employed to gather and evaluate the data. This could reduce concerns about potential copyright infringement and improve transparency in the creation of AI models.

An intelligent individual writing creative content must provide a list of sources to prove the validity of their work and to avoid plagiarism. Why are ChatGPT language models exempt from this requirement? The answer is quite evident: language models lack personal convictions and the capacity to generate authentic ideas. Their capabilities stem from extensive training on diverse data sources, enabling them to generate texts. Nonetheless, it is crucial for any writer, be it human or machine, to acknowledge and reference their sources appropriately. This practice not only ensures the accuracy and reliability of the text, but also prevents some cases of plagiarism. However, there are significant differences in the way human writing and language models (eg ChatGPT) go about proofreading and citing sources. Human writers often view the source as a moral responsibility and express their accountability. They are responsible for the truth of their claims, and citing sources is one way to support their evidence.

In contrast, the issue of citing sources is somewhat more complicated for language models such as ChatGPT. Because these models generate text by using patterns and structures from training data, they do not inherently “support” particular claims or ideas. Instead, their answers are formulated based on statistical probabilities and patterns in the data. Consequently, it should be noted that ChatGPT’s responses are not always accurate or reflective of reality, even though they come from an extensive data corpus. Therefore, it is important to emphasise the importance of source citations for language models. Incorporating this practice would effectively maintain the accuracy and credibility of the text, curb the spread of misinformation and provide transparency regarding the legitimacy of data sources. Sometimes sources can be provided automatically by using training data or contextual cues in the input text. In essence, the issue of source citation is equally important for human authors and language models. Although the approach and timing of source citation may be different for these two groups, the basic principles of maintaining accuracy and credibility and avoiding plagiarism remain unchanged. It is also worth noting that there are different perspectives that challenge copyright holders’ concerns about the use of their intellectual property in generative AI systems. Differing views arise from the fact that developers prioritise data encapsulated in copyrighted works over actual expression.Footnote ⁷³ From the developers’ perspective, documents and creative works are fundamentally viewed as collections of textual content, visual elements or auditory components that serve as unprocessed inputs to computational goals. The main goal of their research is to use the above raw material to train and extend generative AI models. This, in turn, facilitates the development of novel content by leveraging patterns and insights gained from existing works.

On the other hand, copyright law focuses on protecting the unique manifestation of a creation, commonly referred to as “original expression”. This refers to the distinctive and innovative approach authors use to convey their concepts or create visual representations, melodies or other forms of artistic expression. It should be noted that copyright law is not able to cover the basic data, facts and concepts contained in copyrighted materials.Footnote ⁷⁴

Proponents of using copyrighted materials for the purpose of training generative AI systems contend that copyright laws do not protect the basic data and concepts, and therefore it should be considered acceptable to use such works for computational purposes. The argument is that the focus is not on reproducing the exact form of the source material, but on using the information and structures present in that material to achieve novel and inventive results.

VII. Exploring alternatives for safeguarding AI training data

As previously discussed, using data for training purposes involves assimilating information obtained from publicly accessible websites on the Internet, including texts, images and other content. This procedure involves reproducing the content, which may violate the exclusive right of reproduction protected by copyright law and jeopardise the rights of authors and performers.Footnote ⁷⁵ In order to advance and deploy generative AI across multiple industries, it is crucial to improve the access to and use of training data in terms of transparency and fairness. This is because AI algorithms rely heavily on enormous amounts of data to acquire knowledge and make accurate predictions. The availability and accessibility of such data are crucial factors in determining the efficacy and performance of an AI system. Keeping this objective in mind, we have considered a number of strategies for achieving this objective.

Establishing explicit data-sharing agreements with data providers is an essential first step.Footnote ⁷⁶ Data-sharing agreements can be used to address the complex issue of using protected content for AI training while ensuring compliance with copyright laws and protecting content owners’ rights.Footnote ⁷⁷ These agreements are essential for delineating the scope of data usage, establishing limitations, specifying required permissions and arranging the necessary licenses for using copyrighted material in AI training processes.Footnote ⁷⁸ Data-sharing agreements also enable AI developers to establish a legally binding framework that regulates the access, utilisation and administration of protected content throughout the AI training process. The implementation of agreements can yield advantages in terms of establishing unambiguous provisions pertaining to the authorised usage of data and guaranteeing that such usage is confined to the domain agreed upon by both parties. In addition, they have the potential to establish criteria for identifying non-viable data and assessing the permissible utilisation of content in the context of AI training. This encompasses determinations such as the permissible categories of AI algorithms or models and the criteria that govern the timing and duration of data utilisation. Furthermore, these agreements may comprise clauses that pertain to limitations on data utilisation and guarantee adherence to limitations enforced by proprietors of content by AI developers. These limitations may entail forbidding the retrieval or repurposing of data beyond the initial AI instruction or the dissemination or monetisation of protected material.

An additional viable measure for ensuring the protection of AI training data is to contemplate the implementation of certain types of remuneration programmes, such as revenue sharing or royalty payments, to guarantee that creators of copyrighted materials utilised in AI systems are duly compensated.Footnote ⁷⁹ This strategy is important to demonstrate recognition of the inherent value of copyrighted content and to ensure that content creators receive a fair share of the benefits resulting from the use of their works by AI systems. AI developers can establish a direct correlation between the financial gains generated by AI systems and the use of copyrighted works by implementing revenue-sharing or royalty structures. The aforementioned scenario presents a persuasive motivation for content producers to provide their works as training data, given that they stand to gain directly from the financial prosperity of AI systems that utilise their creative output. In accordance with a revenue-sharing arrangement, creators of content would be entitled to a pre-established portion of the revenue produced by an AI system that is utilising their copyrighted materials. This could be a proportional arrangement in which the content creator receives a fair share of the revenue in proportion to their contribution to the training data. Such an arrangement ensures that content creators receive fair compensation for the value that their copyrighted works have on the functionality and success of the AI system.

Alternatively, a royalty-based compensation model could be implemented in which content creators receive a set fee for each use of their copyrighted works by the AI system. This fee structure could consist of a fixed amount per use or a percentage of the revenue generated by the AI system. This model guarantees that content creators receive fair compensation for the duration of the AI system’s use of their copyrighted works by linking the fee to their usage.

The implementation of revenue-sharing or royalty structures requires explicit agreements between AI developers and content creators, specifying the exact terms of compensation. It is obviously important that these agreements specify the exact method for calculating revenue sharing or royalties, as well as their periodicity and temporal scope. Implementing transparent and mutually agreed-upon remuneration mechanisms can safeguard the interests of both AI developers and content creators, promoting a fair and sustainable ecosystem for integrating copyrighted works into AI systems. In general, the concept of compensating content producers through revenue sharing or royalties aims to recognise the importance of their copyrighted material in the AI ecosystem and to ensure that they receive adequate compensation for their role in the prosperity of generative AI tools. The scheme fosters a symbiotic connection between AI developers and content creators while maintaining equitable and just practices in the use of copyrighted material in the AI field.

An additional crucial policy element for safeguarding and reinforcing AI training data could potentially involve the creation and maintenance of open-source datasets intended for the purpose of training machine learning models. The majority of AI research is currently being funded by larger corporations. Hence, it is imperative to institute a programme that provides unrestricted or unobstructed entry to AI training data with the aim of fostering ingenuity, promoting collaboration and propelling the field of AI on a more democratic, equitable and transparent trajectory. The provision of such datasets could ensure that scholars, programmers and corporations are able to utilise them to construct and improve AI models. Advocating for the free or open accessibility of AI training data aligns with the principles of knowledge dissemination, accountability and equitable opportunities for both incumbent enterprises and emerging players from a policy perspective.Footnote ⁸⁰ The reason for this is that well-established enterprises, owing to their prevailing market positions, possess extensive repositories of linguistic and visual data that can be leveraged for the advancement of AI. Enabling broad access to data has the potential to promote the progress of AI technology for the collective benefit of society rather than confining its advantages to a select few. This approach can also foster equitable competition by reducing entry barriers and facilitating the participation of smaller entities and marginalised communities in AI research.Footnote ⁸¹ Hence, it is desirable to develop legislative measures that facilitate the promotion and the exchange of data while safeguarding privacy and intellectual property rights, thereby facilitating the utilisation of open-access AI training data. Specifically, these training data should be considered as a public “participatory good” because their production is based on collective efforts and their value results from the collective participation of numerous individuals who offered their creative content for the creation of training datasets.

To ensure the adequacy of data diversity for the training of AI models, it may be advisable for the law to incorporate provisions that incentivise businesses to voluntarily furnish anonymised data to publicly accessible repositories. The proposed framework has the potential to establish benchmarks for the ethical handling of data, safeguarding the privacy of individuals and preventing the deployment of AI applications that may lead to discrimination or harm. By establishing legal frameworks, it is possible to address issues related to data ownership and licensing, as well as defining the rights and responsibilities of both data providers and consumers. The implementation of similar legislative measures has the potential to foster collaboration among the government, industry and academia by means of funding schemes and recognition systems that prioritise initiatives pertaining to open-access AI training data. One potential strategy by which policymakers could foster a culture of openness and collaboration within the AI industry would be to provide grants and other incentives to researchers and organisations that prioritise the sharing of data. Under this perspective, a practical and concrete solution that could address the problem of copyright clearance of input data (training data) and give AI developers some breathing space would be to establish data repositories or clearinghouses for machine learning training datasets.Footnote ⁸² Establishing data repositories or clearinghouses has the potential to make obtaining licenses and approvals much easier while promoting a more efficient and open process. Indeed, these repositories could act as centralised platforms that facilitate the process of obtaining licenses and permissions and enable negotiations between AI developers and content creators to take place more easily. They also play a crucial role in streamlining the complicated process of resolving copyright disputes, ensuring fair compensation and protecting the interests of all parties involved. Content creators have the option to formally register their works in a designated repository, where they must explicitly state the terms of use and compensation they expect to receive for the use of their protected intellectual property. This allows AI developers to easily access these data and ensure that negotiations are based on accurate and transparent information. AI developers and content creators can more effectively manage the complexity of compensation and rights issues through the use of data repositories. The centralised nature of these repositories promotes consistency and fairness in determining compensation and ensures compliance with licensing terms and copyright laws. In addition to the benefits already mentioned, data repositories can also promote fair competition among AI participants and provide good opportunities for new entrants in the field. Namely, these repositories can provide access to valuable datasets that smaller organisations might not have been able to access on their own. This creates new opportunities for innovation and competition, as a broader range of AI model and algorithm developers can leverage high-quality data, reducing the concentration of data ownership and promoting competition in the AI industry.Footnote ⁸³

Ultimately, implementing ethical guidelines and industry standards for AI training could serve as an additional viable means of protecting AI training data.Footnote ⁸⁴ These guidelines address issues such as data collection, consent, fair compensation and the transparency of the training process. Adherence to these guidelines can help to ensure responsible and legal use of copyrighted works, builds trust and foster a culture of compliance and respect for copyright.

Consent is another essential element that can be covered by such ethical principles. Legal use of copyrighted works requires obtaining appropriate consent from content creators or affected individuals. Such ethical guidelines provide guidance on obtaining informed consent that ensures content creators fully understand how their works will be used in AI training processes and allows them to give consent or opt out if they wish. This promotes integrity and respect for the rights of content creators and fosters a mutually beneficial relationship between AI developers and creators of copyrighted works.

Another fundamental aspect addressed in such ethical guidelines is transparency. These guidelines would promote transparency in AI education by ensuring that the sources and uses of copyrighted works are properly documented and disclosed. Transparency increases accountability and trust among stakeholders and allows content creators and AI system users to track the use of copyrighted works and resolve potential concerns or disputes.

Despite significant advances in the field of generative AI, there is still a reluctance to deploy it widely. A significant obstacle that requires attention is the lack of clearly defined standards in critical areas, including but not limited to intellectual property rights. Successful management of copyright protection for training data can be facilitated through the implementation of effective policies by policymakers. This can ensure the smooth incorporation of generative AI systems. The aforementioned proposals seek to highlight the importance of clear regulations in this context. In particular, these proposals could help to enable fair and clear data retrieval, equitable remuneration of creators, simplified licensing procedures and compliance with copyright regulations and ethical principles. When formulating AI regulations and policies, policymakers should consider the above factors.Footnote ⁸⁵ By incorporating these variables into their models, policymakers can create a durable and innovation-driven environment for the advancement of generative AI.Footnote ⁸⁶