Copyrighting AI-Generated Works: Determining the Originality of AI-Generated Work

Copyright is the exclusive right granted to the creator or author of literary, musical, artistic, audio-visual, or other protected works, allowing them to control reproduction, distribution, adaptation, communication, or public performance of those works. Section 2 of the Copyright Act defines copyright as "the exclusive right to control the reproduction and exploitation of literary, musical, artistic works, audio-visual works, sound recordings and broadcasts in the manner prescribed by the Act.”

For copyright to subsist, the work must be original. This statement correctly captures the general definition and nature of copyright under Kenyan law, particularly as provided under Section 2 of the Copyright Act. However, the requirement of originality becomes legally complex when applied to AI-generated works.

Under copyright law, originality is a foundational requirement for copyright protection. In Kenya, although the Copyright Act does not provide an exhaustive definition of "originality", the concept traditionally refers to a work that originates from an author and reflects a degree of human skill, labour, judgement, creativity, or intellectual effort.

Copyright protects the expression of ideas rather than the ideas themselves. And this brings in the foundational question, which remains unanswered when dealing with AI-generated work. In the context of artificial intelligence, there are complex legal and philosophical questions regarding authorship, originality, and ownership of expression.

Traditionally, copyright law assumes that a human author conceives an idea and then personally expresses it through intellectual effort. AI disrupts this relationship because a user may generate the underlying idea through prompts, but the actual expression of the idea itself is often produced autonomously by the AI system.

The challenge with AI-generated works is determining whether a work produced wholly or substantially through artificial intelligence can meet the originality threshold required under copyright protection. Traditional copyright principles are premised on human authorship. All over the world, the language of the law repeatedly refers to an "author", "creator", or "owner", terms historically understood to contemplate a natural or legal person exercising intellectual creativity.

Under prevailing international copyright doctrine, originality generally requires meaning and often sufficient human contribution. While purely machine-generated works will generally be excluded, AI-generated work with the prompts and inputs of humans creates a fundamental challenge. First of all, are prompts and inputs sufficient to make the work generated by AI original? Secondly, to what extent do these prompts and inputs make the generated work be classified as the original ideas of the user?

There is a strong argument in cases where AI independently determines the manner of expression; the originality requirement becomes weakened from the perspective of the user keying in prompts and inputs. The user may have conceptualised the subject matter but did not personally create the expressive elements protected by copyright. Unlike a human author who directly selects words, composition, tone, or artistic arrangement of ideas and words, the AI system generates those expressive choices through learned models and algorithms.

There is a persistent narrative that copyright should be granted for works that involve human input. This supposition creates some sort of reprieve in the registration, ownership, and protection of AI-generated work. If work is generated partly by AI and partly by a human, what has been generated by the human can be protected as long as it meets the criteria set under the law.

This principle is best exemplified by Zarya of Dawn (2023), the US Copyright Office rule that the text and layout of a graphic novel were protected under copyright; however, individual AI-generated images were not. ⁽¹⁾ This is a clear case where the images were fully AI-generated.

In the US, human authorship is a mandatory requirement; this leaves AI-generated content unprotected ⁽²⁾. Work that is generated by mere text prompts is not worth protection. This supposition shows that there is a need for sufficient input in order for the content to be protected. The same requirement for originality and human effort is mandated under the Kenyan copyright law. ⁽³⁾

This brings in the challenge of determining and ascertaining the level of human input which would make AI work worth protecting, i.e., sufficient human input. Is there a percentage, or should such input be based on the quality or quantity of the prompts and inputs? If that is the case, who should determine this sufficiency given various subject matters (artistic work; literary works; musical works; audio-visual works and film; dramatic and performing arts; and others like advertisements, broadcast, and adaptation)?

Secondly, if AI work protection will be based on the sufficiency of human input and prompts, does that mean that all such work must be accompanied by a list of all the inputs and prompts to determine the quality or quantity of such inputs before protection? If so, there is a challenge of updating the registration of copyright to align with such criteria. This would most likely need a complete overhaul or reform of how copyright is recognised and protected. And practically, it will be a nightmare to implement such a system, given the ability of AI models to churn out millions of AI-generated works.

If it is the sufficiency (level and extent) of input or prompts that determines copyrightability. Then it is also correct to suppose that inputs and prompts will be key indicators of the originality of the AI-generated work. If inputs or prompts are used to determine copyrightability, then such inputs or prompts must meet the originality criteria, too. Whether this will solve the problem remains unclear. Determining originality based on inputs and prompts is indeed complicated to conceive in principle, let alone practically.

Arguments for complex prompting have been suggested. The argument states that a person who directs the AI and makes arrangements for the necessary creation may be eligible for copyright ownership. This supposition is also vague, as it does not define what complex prompting and simple prompting mean. Without delineation of complex and simple prompting, it will be difficult to determine copyrightability.

Without a certain metric of measuring human inputs or prompts, it is very challenging to determine (a) to what extent, (b) what type of human input, and (c) what quality or quantity of inputs or prompts would make the AI-generated work worthy of protection under copyright.

Where an AI system generates text, images, music, or audio-visual content without substantial human creative control, questions arise as to: (a) who the author is; (b) whether the work can truly be said to originate from human intellectual effort; and (c) whether originality subsists in the AI-generated output itself. Since AI works in principle through human input and prompts, the question becomes how such actions should be evaluated vis-à-vis the legal requirement of originality and human expression of ideas.

There is a good argument for protection under the Tool Test. Here, AI is treated like a brush or a camera, or any other tools used by humans for creation, such as Photoshop. This test would support the supposition that AI-generated work is worth protection only when AI is used as an assistive tool. However, applying the principle may be difficult in practice. An example is Suryast (2023), where an AI-generated work based on the work of Vincent Van Gogh's famous 'The Starry Night' was denied copyright in the US because the final composition lacked sufficient human creativity. The AI regenerated work blended a digital photo with Van Gogh's "The Starry Night". This work was rejected because it had no human authorship required for copyright protection.

Also, the difficulty is that copyright law was developed in an era where tools merely assisted in human expression rather than in independently generating expressive content and ideas. A typewriter, camera, or word processor does not independently create expression. Generative AI systems, however, actively participate in producing the expressive form itself. This blurs the traditional distinction between idea and expression. And that is the crux of the matter: reconciling the traditional distinction of idea and expression with what is generated by AI.

Others hold the position that when the user exercises substantial creative control over the process, then copyright should subsist. The argument is that if a user develops highly detailed prompts, iteratively edits the outputs, selects expressive elements, rearranges the content, and significantly modifies the final work, the human contribution may become sufficiently original to attract copyright protection. In such cases, AI functions more as an assistive tool rather than an autonomous creator. While this argument may make sense, implementing it creates practical challenges.

The argument does not consider the whole idea of originality vis-à-vis the ability of AI to create expression. AI-generated work, even when created under extensive control, will retain expression and ideas purely generated by the model itself, despite the inputs and prompts. There is the conundrum of determining what sufficient human input, prompts or efforts are enough to trigger protection. If heavy editing is what is required for AI-generated work, the standard for copyright protection would be lowered across the board.

In Théâtre D'opéra Spatial 2023 ⁽⁴⁾, an artwork which was generated by AI and then heavily modified by a human could not be registered because the AI-generated portions were not adequately disclaimed and disclosed. The Théâtre D'opéra Spatial speaks of a de minimis amount of AI-generated content. This is very difficult to achieve in most AI systems, which would merge AI elements and human modifications. In this case, the user had heavily used Gigapixel and Photoshop to modify an AI-generated image. One may say that the problem in this case arises from the fact that the original image was AI-generated and, hence, the core of the work was AI-generated. It is not that sample. A comparison of Théâtre D'opéra Spatial with the Suryast shows a different story. In Suryast, the original was a digital photo blended with Vincent van Gogh's famous 'The Starry Night' painting. If there were consistency, Suryast would have had a higher percentage of being recognised.

This just shows the complexities of copyrighting AI-generated work.

Questions arise as to what extent AI can be used as a tool and for the final output. Is there a percentage of final composition which would allow combining human creation with AI? Further, in the above cases, there is less clarity on the level and extent, or the percentage of input by humans. In any case, the quantity and quality of input must be properly addressed.

There needs to be in-depth clarification as to the extent to which the user may use AI. Previous decisions have not clarified such a minimum threshold. This clarification would help establish sufficient input or prompts to meet the human authorship criteria. At the moment, no law seems to address these challenges.

Therefore, while the principle that “copyright protects expression and not ideas” remains legally valid, AI-generated works expose a significant gap in traditional copyright theory because, in many cases, the user provides only the idea while the AI produces the protected expression.

In Kenya (and in many jurisdictions like the US), the copyright law does not expressly address AI-generated works or define the extent of human involvement necessary for originality in AI-assisted creations. Copyright law still largely operates on assumptions of human authorship. As AI systems become more autonomous, courts and legislatures may eventually need to reconsider whether originality should continue to depend exclusively on human intellectual creation or whether new forms of authorship and ownership should emerge.