In light of the recent copyright dispute between music labels and AI platforms- SunoAI and Udio AI, we are pleased to bring to you this guest post by Akanksha Badika, discussing some copyright concerns around AI generated music. Akanksha works as the Legal Manager at Bhansali Productions and Bhansali Music, Mumbai. Her practice predominantly revolves around copyright law, litigation and advising on all matters related to films and music . Though the author principally deals with copyright and music related matters, the thoughts and opinions expressed in this post are personal.
Suno & Udio AI in Copyright Tussles: Music Plagiarism or Excusable Overfitting?
By Akanksha Badika
The RIAA (Recording Industry Association of America), a collective representation from music labels of Universal Music Group, Sony Music Entertainment and Warner Music has sued the AI Generative music websites of Suno AI and Udio AI recently on 24th June 2024 in the US District Court of Massachusetts and New York, bringing in the much trending AI copyright tussle to the music industry. The music labels intend to seek reliefs of (a) declaration that these platforms are infringing music labels’ recordings (b) injunct the platforms from using their copyright recordings and (c) seek damages to the tune of $ 150,000 per work infringed place or amount proven at trial from profits generated by the platforms. Allegedly, the Plaintiffs in the suit have contended that these platforms generate digital synthetic music files based on user prompts, by using the copyrighted recording catalogues of the plaintiff music labels and are thereby infringing the copyright of these music labels. The crucial questions arise as to whether these platforms are actually infringing upon the copyright of these music labels or is fair use a good defence for these commercially dwelling AI platforms? This post shall delve into the algorithm used by these AI platforms and potentiality of copyright, publicity rights and intermediary liability claims for these AI platforms.
Musical ChatGPT: How Do You Operate?
In a rapidly advancing world, one cannot turn a blind eye to the perils associated with AI, which these days are also producing music, with no requirement of any lyricist being engaged, recording studio being hired or even a composer, as these AI platforms have been allegedly trained on catalogues of these music labels. The output generated out of user prompts is too closely adapted from the training corpus, which allegedly are recordings owned by these labels. The model of these AI platforms, the output of which is similar to the training data sets is referred to as “Overfitting.” Before even testing the waters of these platforms against copyright infringement, it is crucial to discern the model and the algorithm which platforms use to generate overfitted music outcomes. The step starts with scraping of massive training corpus of pre-existing songs from publicly available domains as available datasets, after which it cleans the allegedly copied recordings which it wishes to remove, much like audio diffusion techniques, processes the corpus to establish parameter values from the AI model and ultimately fine tunes them.
AI: Are you Infringement Resistant?
In the entire process, it is crucial to identify whether the training corpus is being used as a reference or inspiration to AI generative music or is it actually overfitting and rendering results which are substantially similar to the original copyrighted music? As a matter of fact, it is not the process which determines copyright infringement but the outcome which needs to be compared with the infringed portion of the copyrighted segment of music. The complaint filed by RIAA also has adduced comparative music sheet examples of prompts from Suno AI and Udio AI and the copyrighted songs of the Plaintiffs, however as a matter of fact, it is not just the music sheet which could decisively state infringement of copyright but an overall impression of an ordinary man and sensation to the ears and the eye. A small fragment being copied can result in infringement if the said fragment is vital, novel, the heart of the sound record and the original expression of intellectual creation of the author, therefore, even if these platforms contend a minuscule usage of these sound recording, it shall still constitute infringement. Representatives of Suno AI have claimed to use “state of the art” filters to avoid usage of copyrighted recordings, if such is the case, it can clearly be inferred that the training corpus for these platforms is the publicly available copyrighted music from digital sources, otherwise the need for using filters wouldn’t have arose in first place. The entire need of using these filters along with the pop up of Suno AI which cautions users if they use any artists name in their prompts to the platform, amply suggests that all attempts are directed towards disguising the world that the content generated is not infringing of these copyrighted recordings, leading to an inference that the same is a kind of colourable imitation/alteration of the original copyrighted recordings.
Even in the Indian context, Mr. A. R Rahman was recently sued in Delhi High Court by Ustad Faiyaz Wasifuddin Dagar, who claimed that the film PS2’s song “Veer Raja Veer” has the same composition as that of “Shiv Stuti” a composition composed by his father and the Court has asked for musical notations for comparison. However, in this complaint, RIAA has already adduced musical note extraction from the AI generated song to depict similarities in the composition below:
Originality Overfitting: An Un-Fair Use?
The Complaint filed schematically highlights as to how transformative Fair Use set forth in § 107 cannot be a defence for these AI platforms, the table below lists down the factors and its inapplicability in the present case:
| Four Factors for Transformative Fair Use | Inapplicability for Suno AI and Udio AI |
| Purpose and Character of the Use | No functional purpose for use of the copyrighted recordings, Suno AI and Udio use these records not only as training corpus but also resemblance can be marked in outcome as it earns revenue in proportion to the music files. |
| Nature of copyrighted work | The copyrighted recordings of these music label are essentially expressions of intellectual creation. |
| Amount and substantiality of portion used from the original work | The comparative charts of the sound recordings generated by these platforms are substantially similar to the original copyrighted works, generated as a result of user prompts. The essential parts/catchlines and hook parts of the song were shown to be matched. |
| Effect and use upon the potential market of copyrighted work | Suno AI works on a twin model- (a) paid subscription and (b) free music generation. It is quintessential to note that only paid subscribers get the rights to commercialise their music, which they can upload on music streaming apps like- Spotify, etc., which indicates that these songs are in direct competition with the copyrighted recordings potentially diminishing the value of these recordings. |
With this explanation and the way these AI platforms are functioning, nowhere could Fair Use lend a helping hand or act as a great defence to their case. The position from an Indian standpoint shall also remain the same, considering the factors remain similar as per Civic Chandran v. Ammini Amma– (a) quantum and value of the substance taken, (b) purpose of usage and (c) likelihood of competition. Even in India TV News Service Pvt. Ltd. v. Yashraj Films, Delhi High Court has placed reliance on the above four factor theory. Considering- (a) Suno AI, as per analysts does at times take the heart of the song (as illustrated in music notation comparison), (b) purpose is also to make profits out of users generating songs and (c) the platform allows paid subscribers to publish their recordings, there exists likelihood of competition. In light of the above, the position in India too shall remain similar and fair dealing defence shall not hold good.
Potentiality of Ancillary Claims?
Analysts, for instance Ed Newton Rex vide his MWB article, have identified striking similarities in the AI generated musical work and the composition, melody, harmony, chord sequences of the copyrighted songs like Bohemian Rhapsody, Tracks of My Tears along with usage of producer tags (small piece of music indicating a signature style of an artist) and sound alike versions of famous artists like ABBA, Eminem, Ed Sheeran to name a few. This suggests that there exists minimal evidence to at least likely infer that these AI generative music are a result of independent design, but to the contrary have overfitted on the colossal training corpus of plaintiff owned copyrighted content. The usage of these producer tags bearing even the names of these artists like that of ‘Geo Beats’ and soundalike voice clones of artists indeed points towards a publicity rights issue in this matter, which essentially protects a person’s attributes like name, image, voice, style, etc against unauthorized and unlicensed commercial use.
One more aspect from which Suno AI and Udio AI could be viewed from is that of an intermediary (governed under the provisions of Sec. 230 of the Communication Decency Act and does not extend the safe harbour to case of IP infringements), possibly an active one which generates music after prompts given by its users. Suno AI uses Chat GPT to generate lyrics, which itself is facing a tune of legal battles for copyright infringement, one such instance is with New York Times. In such circumstances, even if the test of actual knowledge is applied (a test to identify whether intermediary has specific knowledge of using copyrighted content), Suno and Udio can still be liable as the element of actual knowledge is much present due to massive training corpus of copyrighted recordings, even the factor of red flag knowledge (a test to comprehend whether intermediary had sufficient knowledge that the act will lead to infringement and they will not come within the safe harbour provision, one may read here to know more) comes into play, as representatives of these AI platforms have repeatedly confessed that they are investing in the company despite potential of receiving copyright infringement claims from music labels.
It is crucial to note as to how India is yet to embrace the AI litigation as yet, considering most of these companies operate in the US or other foreign nations. However, from an Indian viewpoint, there exist many challenges which are to be unveiled in the near future. The Copyright Act , 1956 does not cover instances of machine generated works which work on data sets of third party copyrighted works like overfitting. In any circumstance referencing does not lead to copyright infringement, however if the output is substantially similar or sounds similar to an ordinary prudent man (RG Anand v. Deluxe Films) or the substance of the song is getting replicated (Ram Sampath v. Rajesh Roshan), then troubles in India are bound to come. Yet one midway could be to opt for AI Generative Licensing of songs, which protects interest of both parties and lets AI harmonise with human composed songs, which seems difficult due to plausible risk to marketability of the original music which music labels would never pass as excusable overfitting at any cost, everyone wishes to enjoy exclusivity, agreed?