Authorship of AI-generated works in artistic domain

Размещено на http://www.allbest.ru/

Department of Fine Arts Boris Grinchenko Kyiv University

Boris Grinchenko Kyiv University

Authorship of ai-generated works in artistic domain

Snihur Volodymyr Ihorovich

Postgraduate Student

Scientific advisor: Bratus Ivan Viktorovich Candidate of Philology, Associate Professor, Associate Professor

Ukraine

Summary

The article analyzes the question of authorship of Al-generated works in the artistic domain. As usage of AI is projected to increase in many different industries, sometimes regardless of presence of any tangible benefits to the user, we have to evaluate potential impact of an AI on the artistic domain, mostly from the perspective of it being used as a tool by various artists. Discussion is held from the perspective of the evaluation of actual work done and the quality of the result, accounting for the time spent to achieve that result, along with human-made edits where necessary. Discussions about legal standards and decisions of determining sufficient authorship of the artistic work are largely outside of the scope of this article, however, the decision by the U.S. Copyright Office is used as an example of a sufficiently good middle ground found as an attempt to acknowledge existence and importance of AI-generated art, while avoiding any decisions that could prevent its use in both commercial and non-profit environments.

Keywords: artificial intelligence, generative neural networks, AI art, AI.

Анотація

Назва роботи (українською мовою): Авторство створених штучним інтелектом творів у мистецькій сфері

У статті аналізується питання авторства створених штучним інтелектом творів у художній сфері. Оскільки передбачається, що використання штучного інтелекту зростатиме у багатьох галузях промисловості, іноді незалежно від наявності будь-яких відчутних переваг для користувача, ми маємо оцінити потенційний вплив штучного інтелекту на мистецьку сферу, здебільшого з точки зору його використання як інструменту. різними художниками. Обговорення ведеться з точки зору оцінки фактично виконаної роботи та якості результату, враховуючи час, витрачений на досягнення цього результату, а також редагування, внесені людиною, де це необхідно. Дискусії щодо правових стандартів і рішень щодо визначення достатнього авторства художнього твору значною мірою виходять за рамки цієї статті, однак рішення Бюро авторських прав США використовується як приклад достатньо хорошої середини, знайденої як спроба визнати існування та важливість створеного штучним інтелектом мистецтва, уникаючи при цьому будь-яких рішень, які могли б перешкодити його використанню як у комерційному, так і в некомерційному середовищі.

Ключові слова: штучний інтелект, генеративні нейронні мережі, AI art, AI.

Introduction

In recent years, particularly a period from 2020 to 2023, AI and AI-based generative solutions reached the state where they can produce convincing images, stories and sometimes function as tech support agents. This promptly meant that the traditional ways of artistic expressions and tools used for them changed very rapidly and significantly. While the author of this paper believes in these tools being merely an extension and significant improvement to productivity of artists, there is a question that should be answered, and in some ways already has been answered, albeit not to the degree that would satisfy all parties interested in one or another outcome. This is the seemingly simple question of authorship and copyright.

Problem statement

Using neural networks, particularly generative, in professional activity or research is not a problem in itself (or at least not a problem yet), and material produced by them requires human oversight, corrections, selection and improvement. The same process has to be applied to creative works, all to receive something that is close enough to what the artist wanted. AIs also tend to "hallucinate" or in other words - blatantly lie and produce unusable data with full confidence of their output being true or at least usable in some form. Luckily, artistic world is not as much concerned about correctness and perfection of the AIs work (for all we know, we would want them to be slightly off and incorrect, for multiple aesthetic and personal reasons), but we are more concerned with maybe not insignificant, if perhaps more philosophical in nature, question of authorship and copyright. To put bluntly: where lies the limit of what is considered an artist's own authorship and what is not?

The purpose of this article is to discuss the questions of authorship of artistic works in the age of more active and widespread use of AI and generative algorithms.

Generative network as a writer. AI as a writer is, as odd as it may sound, both very, very close to producing convincing human-like textual works, and at the same time its output is way too rigid and "dry". For the purposes of this article, we will use ChatGPT [1] as an example, in particular because it is fairly proficient in natural language processing and understanding and is a product that everyone can test out. artificial intelligence artist

It will have two tasks to complete:

Write a review paragraph about a game "Portal 2". This game is likely familiar to a wide audience, old enough to be known to AI and simple to describe.

Write a brief (no more than 5 sentences) outline of a story set in the 31st century future.

First task tests its factual knowledge of popular entertainment media and ability to express its own judgment about it in clear, concise and short review-like text. Second task is given as an evaluation of completely unprompted (beyond simple theme-setting sentence) creative expressions. Of course AI is free to, and will utilize previous knowledge and experiences to form any of the responses, however we will not "prepare" it for these tasks beforehand, except for asking it to keep answers as factual as possible, choose the most fitting response style and use only English language. The AI product used in this case is a free access version of ChatGPT, its results will be compared to human-authored content.

Table 1 Comparison of AI- and human-generated texts.

Both examples are copied "as is" and while first can be credited to ChatGPT directly, second was written by this article's author using the same prompts given to an AI:

Please write a short review for a game called "Portal 2", please keep information related to the game as factually correct as your training data set allows. Average length of actual human reviews that are useful and comfortable to read is around 10-12 sentences, so your response ideally would be somewhere around that length.

Please write a brief (no more than 5 sentences) outline of a story set in the 31st century future, humanity still exists and is capable of FTL travel, include any kind of questlines if you want to.

Not only the examples above show radically different approaches to the same tasks, but also demonstrate certain rigidity in approach to the task at hand, evidently in ChatGPT's answers, this content, while technically has no human authorship, beyond initial prompt (and in the eyes of the law it has no authorship at all), will need substantial editing to be fit for specific purpose, at which point not only the authorship will lie mostly with human editor, but may cross the threshold of being an insignificant draft, rather than full expression.

Generative network as a painter. The use of generative AIs for producing pictures can not only expand artists' available toolset and let them carry out rapid prototyping phase on a much faster schedule than was previously possible, but also is fairly enticing for many people as a source for "cheap labor" that can produce works possible to credit as one's own while adding nothing transformative to them. For example, Midjourney [2] is capable of producing images of high quality and in different styles but frequently does not understand geometry, scale, perspective or anatomy, especially in parts of an image that were not specified by the user, like is evident in the above example (img. 1) produced from a prompt: "1960s USA heavy industrial overland train". This can sometimes be alleviated by changing the prompt itself, without specifying extra details, like in the next example (img. 2): "USA overland train, style of 1960s or 1980s magazine illustrations".

Image 1. Image generated by Midjourney, example 1

Image 2. Image generated by Midjourney, example 2

As with the test of generative LLM networks, we can see that both images either need extensive prompt edits, editing using Midjourney regeneration feature, employing other neural networks for img2img generation or substantial editing by human artist, as first-attempt generative results are not always of the desired quality and sometimes the work required to achieve it is much greater than that of employing human labor (in our case drawing the image ourselves).

Similarities of human and machine learning. Learning process of human and AI artists look alike: both are influenced by the works of their predecessors and both can improve on what they have learned, they essentially form their own style using knowledge already gained and utilized by somebody else instead of reinventing the proverbial wheel. The only difference, speaking strictly about the learning process itself, is in AI's ability to process hundreds or thousands of images with greater speed than that of its human counterpart. This becomes the point of contention in any discussion about generative algorithms, and is used as an argument by both parties in the ensuing debate, however, what both parties tend to ignore (at least when it comes to non-scientific discourse) is that unlike humans that can grasp the concept and imagine general look of an object as, for example horse, from a singular photograph, AI has no ability to do that. Any generative neural network requires large amounts of data to be used for training for its output to be usable in any way. Beyond that we have to note that image generation is just one part of the process, the second one being natural language understanding and processing, which is in itself an exceptionally hard task for the computers to do. Again, humans naturally excel at this task, as well as the ability to convey complex concepts in simple, short and coherent sentences. If we combine both of these factors together, the fact that there are services like Midjourney, DALL-E [3], DALL-E Mini [4] and an open-source Stable Diffusion [5] with its own different versions, too numerous to count, is a significant achievement, comparable in importance to when computers became capable of real-time image editing. As evident from previous examples, while the use of AIgenerated content is possible, even in "first-attempt" cases, likely it will require substantial human involvement, different from extensive prompt-crafting and generation of multiple different variations, and therefore can be considered merely a draft and as such the authorship partially or wholly lies with the human author as the one finally bringing the work to the required level of quality.

Legal stance. In March of 2023 US Copyright office issued a notice [6] that can be summed up as "if the work was produced by an AI and submitted to us "as is", so that no substantial human involvement, beyond initial prompts, is present, then the work can not be granted copyright protection". This solution is further supported by the notice of UK Intellectual Property Office [7] that concludes that AI can only infringe copyright (and as such for the purposes of this article be considered "author" of infringing piece) only if it is tasked with recreating the original work itself, in which case the infringement lies on the person or entity guiding AI to such an act, however, the fact of learning and remembering about the work is not an infringement in itself as it is not an infringement for a human to learn and remember about a certain piece of media or any other work or event. This stance is further supported and expanded by an article in the "International Review of Intellectual Property and Competition Law" [8] stating that assigning copyright and authorship to an AI will be detrimental to the AI itself as pretty much any service that would dare to offer such restrictive terms of use is doomed to fail from the start. This is a fairly correct stance as doing so would equate the manufacturer of photographic equipment claiming copyright and authorship of photographer's images only because their equipment was used, as such there is a potential of legal issues with very real monetary, reputational and other damages involved for the user and the service provider as well, along with potential of rapid and irreversible loss of customers, both individual and corporate.

Conclusion

While generative algorithms and neural networks are capable of producing content of fairly high quality, they should be seen as tools for rapid prototyping rather than creation of production-ready content (at least not without very specialized use-case specific training involved), their work frequently needs extra touch-ups and improvement done by human authors. This by itself grants a large part of authorship to the human user of an AI, which is further enhanced by partially self-enforced and partially legally-limited inability and ineligibility of AIs as copyright and/or authorship holders. However, if the work is presented "as generated" the work may remain without individual copyright protection, the question of authorship thus remains open. Answering the question posed in the beginning of the article we can conclude that in its strictest definition the human authorship in AI-generated works begins with the act of editing the generated material.

References

1. ChatGPT by OpenAI

2. Midjourney

3. DALL-E

4. DALL-e mini

5. Stable Diffusion

6. Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, US Copyright Office

7. Artificial intelligence call for views: copyright and related rights. (n.d.). GOV.UK.

8. Hugenholtz, P. B., & Quintais, J. P. (2021, October 4). Copyright and Artificial Creation: Does EU Copyright Law Protect AI-Assisted Output? - IIC - International Review of Intellectual Property and Competition Law. SpringerLink.

Размещено на Allbest.ru