By L. Anisha
Anisha is a student at Hidayatullah National Law University, Raipur. This piece is one of the winning entries from the Copyright Law Blog Writing Competition, 2023.
“One of the most inspiring aspects of AI art generation is that the software can learn from old images and breathe new life into classics.”
Electra Nandu[1]
“AI image generators are 21st-century collage tools that violate the rights of millions of artists.”
Class Action Complaint against Stability Diffusion[2]
Artificial Intelligence (“AI”) image generators are experiencing a remarkable boom with their ability to create stunning and realistic visuals. These models generate artworks after receiving training from millions of pictures gathered from the web. Sometimes, the developers do not obtain proper consent from the copyright owners.[3] Recently, a class-action lawsuit was filed by a trio of artists against popular AI Image Generators like Midjourney, DeviantArt and Stability AI for infringing their copyright.[4]
AI is constantly evolving and pushing the boundaries of Copyright Law. From the Indian copyright infringement perspective, two major questions arise with respect to AI Image Generators: (1) Whether images generated by AI infringe on existing copyrights? (2) Whether accessing, analysing, reading and mining of data to train AI constitute copyright infringement?
AI Artworks: Navigating the Copyright Minefield
Popular AI generators like DALL E, Midjourney etc., use the diffusion model for creating images. These models take an image and corrupt it by adding noise such that it is no longer recognisable. The learning process takes place when the AI figures out how to reconstruct the original image by reversing the corruption process.[5] (Exhibit I).
Establishing infringement is complex when AI is involved as it generates art which is similar, but not identical to the copyrighted work. However, just similarity is not enough to prove infringement as it might be mere coincidence or works being derived from common sources.[6]
In order to establish copyright infringement in AI artworks, three ingredients need to be established as per Indian law:
- The existence of similar work protected under copyright
- The infringing act is touching upon the exclusivity granted by copyright
- The act is “actually infringing” in nature[7]
If the defendant makes use of a substantial portion of the features in a copyrighted work, he is unlawfully exercising the sole right conferred only upon the plaintiff.[8] To satisfy the substantial similarity test, some part of the output should resemble the input. However, with billions of input data, chances are that there might not even be a trace of copyrighted work in the output. AI Image Generators do not merely ‘cut and paste’ existing images to create artwork. Exact replication of input data is rare considering the huge amounts of training datasets fed to AI. The odds of such an occurrence are merely 1.88%, as reported by a study.[9]
AI Image Generators have also become adept at producing artwork “in the style of” a certain artist. However, as per the Supreme Court in R.G. Anand v. M/S Deluxe Films[10], copyright owner has no locus standi to sue the defendant in such instances as artistic styles are not copyrightable. Only expressions of ideas are copyrightable.
Although AI outputs can escape infringement to some extent, the same cannot be assured for data mining activities to train AI.
AI Training Data: Walking The Tightrope of Copyright Infringement
Admittedly, when large amount of data is accessed, analysed and mined for the purposes of training AI, some amount of copyright infringement is bound to occur. In India, Section 14 of the Copyright Act, 1957 grants exclusive rights to copyright owners to reproduce, issue copies, make translations, adaptations etc. of the work. Section 14 is compromised unless such use of data is covered by exceptions to escape infringement allegations.
At the same time, Section 52 of the Copyright Act, 1957 lists out certain acts that do not constitute infringement. If the work is used for “private or personal use, criticism or reporting of current events,” it comes under fair dealing and there is no infringement. The Delhi High Court, in a case, analysed both fair use as prevalent in the US and fair dealing doctrine in India. Ultimately, both the doctrines were used interchangeably in the judgement as they were conceptually similar. So, in case of AI art, the legal validity of training data is judged using fair use parameters which constitutes:
- “The purpose and character of Use
- The nature of the copyrighted work
- The substantiality of the portion used in relation to the copyrighted work as a whole
- The effect on the potential market for or value of the copyrighted work”
The Delhi High court emphasised on the first factor. If the character of the work is transformative i.e., the work is not merely a substitute but materially different from the original work, then it is fair use and it is immaterial if the copying is substantial.[11]
The character of human generated art is fundamentally different from AI generated art. Art is a medium of communication for artists. While humans create art to express, AI uses digitized art to train its algorithm. The character is certainly transformative in nature.
Similarly, the fourth parameter i.e., the effect on potential market is minimized if the work is transformative and not merely a substitute. Most buyers will prefer the original artwork of a renowned artist instead of opting for a computer-generated version. In this scenario, market is not affected.
But, the fourth factor is likely to be against AI as it has potential to create market disruptions especially if the buyers are price sensitive. Art created by humans will be priced comparatively higher to account for the time and effort consumed by the process. In such an event, buyers will prefer AI artworks due to cheap prices. Moreover, AI has become an expert in replicating the style of artists in an inexpensive manner. While popular artists might escape the danger, small scale artists are at risk of substitution. The legal loophole that styles are not copyrightable will certainly cause disruptions in the market in the future.[12]
Final Thoughts: Harmonizing the Pros and Cons
Technological revolutions have always been met with legal strife. It pushes the law to constantly evolve. The current legal position regarding AI outputs is in technology’s favour. It is difficult to identify copyright infringement when the AI generates an entirely new image based on billions of training data. Although replication rates are low, they are not zero. Therefore, it is imperative for developers to take steps to thwart copyright infringement. They must consider using reverse image tools on outputs to determine substantial similarity to existing works. In such a scenario, alternative images must be generated.
Regarding AI inputs, developers are using various arguments to satisfy the four-factor test in the fair use doctrine. Some of these arguments seem valid only because AI Image Generators are at their infancy. While they are transformative, their nature of use might be non-commercial only in the initial stages. Platforms like Dall-E and Midjourney have already begun rolling out subscription models.[13] If these businesses choose to commercialise their platforms in the future, it will weigh against fair use. It is also too early to determine AI’s effect in the potential market. It is currently not disruptive but, in the future, it might undermine the jobs of millions of artists.
So, to answer the question whether AI Artworks are copyright infringement or creative innovation – they are both. While AI outputs fall in the realm of creative innovation, AI inputs are walking a tightrope bordering on copyright infringement.
India’s 1957 legislation is doing a remarkable job in applying traditional copyright concepts to AI developments. However, it is time for the law to evolve. AI is here to stay. The law must seek innovative solutions to legitimise Text and Data Mining (TDM) activities to train AI and balance the rights of copyright holders instead of holding TDM activities as infringing altogether.
[1] Electra Nanou, The Ethical Pros and Cons of AI Art Generation, MakeUseOf (Jan. 18, 2023) https://www.makeuseof.com/ai-art-generation-ethical-pros-cons/. (last visited June 27, 2023).
[2] Stable Diffusion Litigation, https://stablediffusionlitigation.com/ (last visited June 20, 2023).
[3] Gil Appel and others, Generative AI Has an Intellectual Property Problem, Harvard Business Review (Apr. 7, 2023), https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem (last visited June 23, 2023).
[4] Supra note 2,
[5] Andres Guadamuz, A Scanner Darkly: Copyright Liability and Exceptions in Artificial Intelligence Inputs and Outputs, SSRN (Feb. 26, 2023), https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4371204 (last visited July 3, 2023).
[6] R.G. Anand v. M/S Deluxe Films, 1978 AIR 1613.
[7] Adarsh Ramanujan and others, Infringement Analysis in Copyright Law, Lakshmikumaran & Sridharan Newsroom (Aug. 10, 2011), https://www.lakshmisri.com/newsroom/archives/Infringement-Analysis-in-Copyright-Law# (last visited July 1, 2023).
[8] Mohini Mohan Singh v. Sita Nath Basakm, AIR 1931 Cal 233.
[9] Supra note 7.
[10] 1978 AIR 1613.
[11] The Chancellor Masters and Scholars of the University of Oxford v. Narendra Publishing House, CS(OS) 2439/2012.
[12] Jessica Gillotte, Copyright Infringement in AI-Generated Artworks, 53(5) UC Davis Law Review (Aug. 20, 2020), https://lawreview.law.ucdavis.edu/issues/53/5/notes/files/53-5_Gillotte.pdf.
[13] Cala Coffman, Does the Use of Copyrighted Works to Train AI Qualify as Fair Use?, Copyright Alliance (Apr. 11, 2023), https://copyrightalliance.org/copyrighted-works-training-ai-fair-use/ (last visited July 2, 2023).
ScholarshIP 