TYGlobe

Authors of this Article: Shen Qing, Huang Jiaping

In August 2023, a news report emerged online that Internet giant Google Inc. and Universal Music Group (UMGNF.US) are negotiating over the licensing of artificial intelligence (AI) music copyrights. According to the report, the objective of the negotiations between Google and Universal Music Group is to develop a set of tools that allow fans to legally generate AI music and pay royalties to copyright owners, and artists shall have the right to choose whether to participate throughout the entire process.

This news may be good news for music lovers and music artists. As is known to all, artificial intelligence (AI) is already capable of imitating singers' voices with extremely high fidelity. A well-known case is that the cover of *Hair Like Snow* performed by "AI Stefanie Sun" and released on Bilibili in April this year went viral across the whole network of China. By mid-May, "AI Stefanie Sun" had covered more than 60 classic songs in different languages including Cantonese, Japanese and English. As of the date hereof, works related to "AI Stefanie Sun" have accumulated over 10 million views on the two platforms of Douyin and Bilibili. In the meantime, "AI Jay Chou", "AI Faye Wong" and other similar AI-generated contents have also emerged successively.

Amid the surging popularity of AI celebrities, issues arising from deep imitation leveraging artificial intelligence have also sparked intense controversy. As AI often generates new music using artists' voices and melodies without obtaining prior authorization, such conduct may have constituted a material threat to the music industry. As the General Counsel of Universal Music Group stated, "The voice of a music artist is the most valuable asset in their professional career, and stealing it in any form whatsoever is wrongful."

As a legal practitioner, this author also conducts legal analysis from a professional perspective on the phenomenon of generating songs by imitating singers via AI, so as to enable readers to obtain further understanding thereof.

What is meant by "AI singer"?

"AI singers" are in essence produced through model training and post-processing, which enable AI to cover songs originally performed by other singers with the voice of a specific singer. Put simply, "AI singers" are "trained" with data. Take "AI Stefanie Sun" as an example. To produce songs of "AI Stefanie Sun", the first step is to obtain raw vocals of Stefanie Sun without accompaniment or reverberation. Such materials can be collected from her interview or performance videos via accompaniment removal or dialogue extraction. Afterwards, these materials are input into the model as the target timbre for training, so as to generate a model capable of producing audios with Stefanie Sun's timbre. Next, this model can be used to convert any song regardless of its pitch, such as songs by JJ Lin, Jay Chou, Zhao Lei and other singers. Finally, through further post-processing and tuning, songs that sound as if they are sung by Stefanie Sun can be obtained.

Does the imitation act of an "AI singer" constitute infringement?

First, from the perspective of the "training" phase, artificial intelligence needs to be fed with a large amount of information in the early stage to obtain vocal feature data of real singers. Take "AI Stefanie Sun" as an example again. According to the current working mode of artificial intelligence, publishers mostly extract Stefanie Sun's voice from her speeches (including those delivered in interviews, press conferences, concerts and other occasions). After the extracted voice materials are input into the artificial intelligence learning engine and trained, the vocal feature data of Stefanie Sun can be obtained. These voice files are legally defined as audio or visual recordings, and their producers enjoy the right of producers of audio-visual recordings. As a performer featured in these audio-visual recordings, Stefanie Sun also enjoys the right of performers in respect of the content of the aforesaid voice files.

Pursuant to Article 7 of the *Interim Measures for the Management of Generative Artificial Intelligence Services*, which came into force on August 15, 2023, providers of generative artificial intelligence services shall be responsible for the legality of the sources of pre-training data and optimization training data of generative artificial intelligence products. Pre-training and optimization training data used for generative artificial intelligence products shall not contain content that infringes upon intellectual property rights. Accordingly, in light of the aforesaid provisions, where pre-training data and optimization training data include works in which others enjoy copyright or objects in which others enjoy neighboring rights, permission shall be obtained and remuneration shall be paid in accordance with the provisions of the *Copyright Law of China*. Failure to do so may constitute infringement of others' copyright or neighboring rights. This is also the premise for the ongoing negotiations between Google Inc. and Universal Music Group on the development of an AI music generation tool, under which model, fees will be paid to copyright holders and music artists to legalize the "bionic" singers generated by such artificial intelligence.

Secondly, from the perspective of the audio generation stage, the author holds that "AI singers" may be subject to the following infringement risks.

Take personality rights as an example. China's Civil Code stipulates that the personality rights of civil subjects are protected by law, and no organization or individual may infringe upon such rights. The right to name, the right to portrait and the right to voice are all part of personality rights.

With regard to the right of name, the Civil Code of the People's Republic of China stipulates that no person may infringe upon another person's right of name by means of interference, usurpation, counterfeiting or other improper means; nor may any person use a name or stage name that enjoys certain social popularity and causes public "confusion". For example, the use of "Stefanie Sun" in the concept of "AI Stefanie Sun" may be intended to clarify that the relevant song is produced by synthesizing a voice highly similar to that of Stefanie Sun through AI training. However, such use of "Stefanie Sun" does not exclude the suspicion of "confusion" and "free-riding". As Stefanie Sun is a well-known singer, the public tends to believe that there is a certain degree of confusion or association between "AI Stefanie Sun" and Stefanie Sun herself. Although the aforesaid act may not necessarily devalue the commercial value attached to the name "Stefanie Sun", in practice, if the publisher uses the relevant content for commercial purposes or to attract followers and divert traffic, the act involved in "AI Stefanie Sun" may be determined as an infringement upon the right of name.

With regard to portrait rights, the Civil Code of the People's Republic of China stipulates that no one may create, use or publicize the portrait of a portrait right holder without the consent of the holder. If an "AI singer" uses a singer's portrait in its released song videos, such act is suspected of infringing upon portrait rights. Although the Civil Code stipulates that "using the publicly disclosed portrait of a portrait right holder within the necessary scope for the purpose of personal learning, art appreciation, classroom teaching or scientific research" constitutes fair use, the act of releasing songs online falls neither under personal learning nor art appreciation or other eligible scenarios, and thus shall not be deemed as falling within the scope of fair use.

With respect to the right of voice, the Civil Code of the People's Republic of China provides that the protection of such right shall apply mutatis mutandis the relevant provisions on the protection of the right of portrait. Namely, without the consent of the right holder of the right of voice, no entity or individual may produce, use or publicize the voice of the right holder, nor defame, deface or forge the voice thereof. For instance, in the case of "AI Stefanie Sun", where AI technology is employed to imitate Stefanie Sun's voice to produce, use and publicize the right holder's voice without authorization, the publishers of the relevant content may be found to have committed infringement.

In addition, the author points out that from the perspective of the Anti-Unfair Competition Law of the People's Republic of China, such act of using Stefanie Sun's voice will mislead the audience into believing that the performance is rendered by Stefanie Sun herself, which is highly likely to constitute an act of confusion as prescribed in the Anti-Unfair Competition Law.

Finally, publishing cover songs performed by "AI singers" online may infringe upon the relevant intellectual property rights of lyricists, composers or other intellectual property right holders.

Take the most popular AI-generated Stefanie Sun cover of Jay Chou's song *Hair Like Snow* on Bilibili as an example. The publisher shall obtain the copyright authorization for the lyrics and melody of *Hair Like Snow*, as well as the authorization from the producer of the sound recording of the accompaniment audio of *Hair Like Snow*. If any of the aforesaid authorizations is missing, there is a legal risk of infringing upon the intellectual property rights of relevant right holders. Without the permission of the copyright holder of the original musical work, where sound recordings generated by AI cover performances are used for mechanical performance or dissemination in specific business premises, the right of performance of the musical work shall be infringed; where such recordings are broadcast by radio and television stations, the right of broadcasting of the musical work shall be infringed; where such recordings are disseminated online via interactive communication, the right of communication through information networks of the musical work shall be infringed.

Are cover performances by "AI singers" entitled to legal protection?

As far as current laws are concerned, Article 2 of the Copyright Law of China restricts the subjects entitled to copyright to "Chinese citizens, legal persons or other organizations" and eligible "foreigners and stateless persons". In other words, only intellectual creations produced by human beings are protected under the current legal framework. While generative artificial intelligence is not granted copyright protection under current laws, relevant authorities have recently been actively rolling out relevant legislative documents, requiring all providers of generative artificial intelligence technologies to be subject to supervision and explicitly mark that their products are produced by artificial intelligence. Otherwise, relevant competent state authorities shall have the right to pursue the legal liabilities of such technology providers.

Articles 16 and 17 of the *Regulations on the Administration of Deep Synthesis Internet Information Services*, jointly issued by the Cyberspace Administration of China, the Ministry of Industry and Information Technology of the People's Republic of China and the Ministry of Public Security of the People's Republic of China in November 2022, stipulate the information disclosure obligations of service providers. Among them, Article 16 provides that for information content generated or edited through the use of its services, a deep synthesis service provider shall adopt technical measures to add markings that do not affect users' use, and keep log information in accordance with laws, administrative regulations and relevant state provisions (implicit disclosure). Article 17 provides that where a deep synthesis service provider provides voice generation services such as synthetic human voice and voice imitation, or editing services that significantly alter personal identity characteristics, which may cause public confusion or misidentification, it shall place conspicuous markings at reasonable positions and areas of the generated or edited information content to alert the public of the deep synthesis situation.

With the rapid development of generative artificial intelligence, strong artificial intelligence is gradually replacing weak artificial intelligence and permeating all aspects of people's daily lives. It has not only opened up a new horizon for the development of artificial intelligence, but also opened Pandora's Box. Risks including AI-generated content that is indistinguishable from authentic content, data leakage and infringement have gradually come to light. In terms of the future development and regulation of generative artificial intelligence, it is imperative to keep pace with the scientific and technological revolution, and continuously improve relevant legislation on the basis of complying with sci-tech ethics, people-oriented principles and the principle of the rule of law.