The rise of AIGC-generated songs has brought many challenges to existing laws. In the process of generating AIGC songs, the right of reproduction will be infringed in the data storage and model training stages, and will be subject to the performer's image protection right. For the original song, the live transmission of the song generated by AIGC is a public mechanical performance, and the permission of the copyright owner of the musical work should be obtained. If the transmission is carried out on radio, television or webcast, it is a non-interactive transmission, and if it is provided on demand or downloaded on the Internet, it is an interactive transmission, and the information network transmission right is applicable. For the performers who are imitated, there is no performance in the process of generating AIGC songs, and the performers' rights are not violated. On the issue of personality rights, the generation of songs allows the use of other people's voice models, which does not constitute an insult or scandalization of other people's personalities, and does not belong to confusing and deceiving the public on the premise that it has been declared to be created by AI, but the unauthorized use of other people's voice models will damage the property value of the voice and infringe on personality rights.
I. Formulation of the problem
After the explosion of ChatGPT, a text-generating artificial intelligence, "AI singers" used other people's voice models to cover songs to attract everyone's attention again. AIGC generated songs are produced through an open-source project called "So-VITS-SVC". However, this project requires a large amount of training data to generate songs, and the application threshold is extremely high. "AI Singer" uses the "So-VITS-SVC" model to train a voice model with the timbre of the target singer on the basis of a large amount of data, and covers the songs of other singers. In practice, the production of an AIGC song mainly involves the following three steps: first, obtaining the song audio file of the target singer, separating it into pure vocals, and forming the audio dataset of the target singer; Thirdly, the songs that need to be covered are selected, and the timbre of the target singer is converted by using the trained target singer's voice model to obtain AIGC generated songs. Article 7 of the Measures for the Administration of Generative AI Services (Draft for Comments) issued by the Cyberspace Administration of China on April 11, 2023 stipulates that providers shall be responsible for the legality of the pre-training data and the source of optimized training data for generative AI products. The pre-training and optimization training data used for generative AI products should not contain content that infringes intellectual property rights. Judging from this provision, it seems that if the pre-training data or optimized training data contain the copyrighted works of others or the objects with neighboring rights, they should obtain a license and pay remuneration in accordance with the provisions of the Copyright Law, otherwise they may infringe the copyrights or neighboring rights of others. Songs generated by sound models extracted and processed by artificial intelligence are not included in the adjustment and protection system of current intellectual property laws and regulations, but there is no doubt that they should be regulated accordingly. Does the act of covering a song by AI infringe upon the copyright of the original song? Does the AI simulating human voice and using another person's voice model for singing and performing infringe on the performer's rights? In addition, it is also held that the misuse of another person's voice model to interpret musical works does not fall within the scope of the Copyright Law, but should be regulated by the relevant laws on the personality rights of civil law. Questions such as these have sparked a lot of controversy. In order to further develop the positive value of AI songs and give full play to the positive role of AIGC sound model technology, the above problems must be reasonably explained.
2. Copyright risks in the process of generating AIGC songs
From the perspective of the song generation process, the copyright dispute that may be involved in AIGC technology is whether there is an infringement risk of the data content used for training. The training data of "AI singer" and the "SO-VITS-SVC" model behind it are concentrated in the field of music copyright and neighboring rights, and the subject matter and rights involved are relatively complex.
At present, the voice file data involved in the training of the "AI singer" timbre model mainly includes two categories: one is the "music audio recording" of the target singer, and the other is the "ordinary voice product" of the target singer, including the daily dialogue of the singer, as well as the recording files of the dialogue of the interview meeting and press conference. In the first category, if the target singer's "music recordings" are used for model training, the "music recordings" are the subject matter of the copyright law, involving three parties and their rights. The first is the right holder of the lyrics and music work corresponding to a particular song and the "copyright" it enjoys; the second is the singer who sings the song and the "performer's right" he enjoys, and the third is the sound recording producer who makes the "sound recording product" and the "sound recording producer's right". In the second case, if the target singer's "ordinary voice file" is used for model training, in principle, only the authorization of the recording producer corresponding to the voice file needs to be obtained. However, it should be noted that if the target singer's voice dialogue is recorded without permission, it may involve the risk of infringement of personality rights such as privacy and voice. If the above-mentioned audio recordings involve the display of content with certain performance significance, such as speeches and talk shows by the target singers, then the same situation as the first category still involves the rights of the three parties and related subjects.
The Copyright Law stipulates the specific circumstances of "fair use", and the part involving music generally includes "personal use", "appropriate quotation", "study and research use", and "free performance". However, AIGC-generated songs are widely disseminated on audio-visual platforms, which not only highly overlap with the potential market of original lyrics, songs, and audio products, but also have various revenue channels such as traffic, advertising, and rewards. Therefore, at present, it is difficult to classify the use of works and products involved in the training of the "AI singer" model into the specific situation of "fair use". Although the Copyright Law amended in 2021 adds "general requirements" and "catch-all clauses" to the "fair use" provisions. However, the "catch-all clause" is a semi-open content provision - "other circumstances stipulated by laws and administrative regulations", which cannot be directly applied by the court in judicial practice according to the facts of the case. Therefore, the "fair use" exemption is not yet applicable to the "AI singer" timbre model training, and relevant authorization is still required.
(1) Copying during the data storage phase
In the process of making the "AI singer" voice model, a large number of target singer "music products" and "voice files" are required as training materials. Most of these materials need to be stored in the AI model server for pure vocal extraction and processing. Undoubtedly, this act first involved the licensing of the right of reproduction by a large number of rights subjects. In the copyright law of the People's Republic of China, songwriters, producers of audio and video recordings, and performers all enjoy the relevant right of reproduction, so in principle, it is necessary to obtain the authorization of the three rights subjects before it can be implemented. However, the application of artificial intelligence generation technology and new media communication technology, this right of reproduction is rarely taken seriously. This is because "copying" is largely implemented in preparation for the manufacturing of the target product. If there is no public exploitation of a work regulated by copyright laws such as distribution, broadcasting, or information network dissemination after reproduction, there is practically no way to talk about infringement damage, and there is no way for the right holder to find out whether his work has been exploited. However, there is no doubt that the number of copies in the songs generated by AIGC has increased, so that many materials are relatively stable and fixed, and they are carried by the material carrier of the AI model server, so in the data storage stage, unauthorized copying is an infringement of the right of reproduction.
(2) The behavior of "adaptation" in the training stage of the sound model
In the training stage, the training data containing the timbre characteristics of the target singer is processed, so as to obtain the recognizable and distinctive timbre and voiceprint of the target singer and store them as model parameters for subsequent scheduling. On the issue of data processing and processing within the model, due to the processing and adaptation of the original sound, it may fall into the scope of adaptation right and catch-all right regulation. According to the provisions of the Copyright Law of the People's Republic of China, there are works that enjoy the right of adaptation and catch-all rights, but the rights of producers and performers of sound recordings in the adjacent rights do not cover the scope of the above-mentioned rights. Therefore, according to the above viewpoint, it is obviously unreasonable for the author of the musical work to obtain the corresponding rights protection in terms of the timbre model training process, while the sound recording products and performers lose the rights protection. It is also argued that the above-mentioned acts do not fall within the scope of the Copyright Law. The data processing behavior inside the model is not for the purpose of using the expressive elements in the work, but is a kind of "non-expressive machine utilization", which belongs to the use of the work in the sense of non-copyright law. Therefore, unauthorized internal training and processing of the model is not an infringement under the copyright law and is not regulated by the copyright law. However, although the timbre, voiceprint and other factors of the target singer do not fully belong to the object of copyright protection, the law gives the performer the right to protect the performance image from distortion. Arbitrary adaptation or scandalization may still constitute a distortion of the performer's performance image and constitute infringement.
3. The legitimacy of the AICG-generated song in the song copyright is proven
Does the act of "covering" another person's song by simulating the voice of a singer through AIGC technology constitute an infringement under the Copyright Law? To solve this problem, we must first clarify the idea of copyright infringement. Copyright is the right enjoyed by civil subjects in accordance with the law on works and related objects. An infringement under copyright law must be an act that is subject to exclusive rights. Therefore, when discussing whether an act is infringing, it is necessary to first consider whether the act is subject to exclusive rights. If you answer in the affirmative, you can continue the discussion, otherwise that's it. Similarly, the fact that an AI song is simply produced without being publicly "covered" and disseminated obviously does not constitute an infringement controlled by the Copyright Law, so it is of significance to discuss whether an AIGC song is disseminated. The communication right system in the Continental Copyright Law can be divided into two categories according to the place where the audience is located, the first category is the "on-site communication right", which includes the right of exhibition, the right of performance, the right of screening, and the second sub-right of the broadcasting right - the right to publicly broadcast the received work after initial transmission, and the second category is the right of remote communication, including the first sub-right of the broadcasting right (the right of non-interactive communication) and the right of information network communication (the right of interactive communication). Whether the AI cover behavior is controlled by two types of rights determines its legitimacy.
(1) The songs generated by AIGC are controlled by the on-site transmission right
Obviously, the songs generated by AIGC do not involve the content of the exhibition right, the screening right and the second sub-right of the broadcasting right, and the performance right of the original song may be controlled by the performance right of the original song in the form of using AIGC technology to interpret the song with the voice of other singers. The copyright law of the People's Republic of China defines the right of performance as "the right to perform a work in public, as well as the performance of a work to be publicly broadcast by various means", which mainly includes two aspects: live performance and mechanical performance. Live performance is the performance of the work by the actor to the live public, and its main regulation is the performance of the person. Mechanical performances, on the other hand, use mechanical devices to broadcast performances of works to the public. Playing other people's music as background music in shopping malls, restaurants, bars, etc. is a typical mechanical performance. First of all, in the case of mechanical performances, the work is performed through mechanical equipment. For example, a film containing a performer's video is played in a cinema using a mechanical projector. However, mechanical performances do not necessarily have to be performed by humans and then reproduced through facilities, and mechanical performances can be performed without involving human beings. For example, the popular player piano in the 20s of the 20th century works by arranging the score into a perforated paper tape, and controlling the sound of the keys through the movement of the tape. No human performance is a prerequisite for this process, but as long as the music being played is protected and played to the public in a public place, it should naturally be permitted and controlled by the owner of the music copyright. Similarly, although artificial intelligence is far more complex than common music players, projectors and other machines, it uses advanced algorithms and software programming techniques. Although there are also calls for AI to be given an independent personality, at this stage AI can only be regarded as a mechanical device. The essence of the songs generated by AIGC is nothing more than a musical performance performed by a complex machine compared to a perforated paper tape. The second is the issue of "communication to the public". The World Intellectual Property Organization's Guide to Copyright and Related Rights Treaties Administered by WIPO clearly stipulates "communication to the public": "...... It is ...... that it can be perceived by people outside the usual family circle or outside the close social relationship of the family. This means that two specific situations are excluded from the definition of "communication to the public", namely, the family situation, such as transmission among family members, and the case of close friends, such as sharing between close friends. If someone downloads a song from an "AI singer" on the Internet and plays it at home for self-appreciation. This act is not a public performance and is not infringing. Even if the music played is pirated, it is still not under the category of communication to the public and is not an infringement. The performance right only regulates public performances for the public, but the performance of non-public-facing individuals is also a performance, just because it does not constitute a "public performance" regulated by the performance right. Therefore, it is a non-public performance to be played at home. This is not subject to exclusive rights, even if pirated audio is played, nor is it subject to copyright law. However, if the broadcast of AIGC-generated songs in the business premises is a public mechanical performance, the permission of the copyright holder of the musical work shall be obtained, regardless of whether the music is an AI cover, and the music being performed is protected by the Copyright Law, and the performance of music, regardless of whether the human performance or the mechanical performance is a public performance, shall be subject to the permission of the copyright owner.
(2) The songs generated by AIGC are controlled by the right of remote transmission
If a radio station, television station or online anchor plays an AIGC-generated song in a live broadcast, what rights will be infringed by the copyright owner of the musical work......? rights. Second, the right of performance: "the right to perform a work in public, as well as the right to publicly transmit the performance of the work by various means". The law stipulates that the right of broadcasting is "the dissemination of a work", while the right of performance is "the performance of the transmission of the work", and the broadcast of songs generated by AIGC by a radio station, television station or anchor is obviously a "performance of the work" rather than a "work". This understanding of the "communication work" in the broadcasting right is only considered to be "the original form of the communication work without performance" is too narrow. Statutory permission for the broadcast of sound recordings in the Copyright Law: "Radio and television stations may broadcast the published works of others without the permission of the copyright owner, but shall pay remuneration in accordance with the provisions." "Television stations broadcasting other people's audio-visual works or video products shall obtain permission from the copyright owner of the audio-visual works or the producer of the video and pay remuneration; "Radio and television stations usually broadcast sound recordings of audio works. Article 1 of the Interim Measures for the Payment of Remuneration for the Broadcast of Sound Recordings by Radio and Television Stations (hereinafter referred to as the Interim Measures) reads: "In order to protect the copyright owner from exercising the right of broadcasting in accordance with the law,...... Article 2 of the "Formulation of these Measures": "Radio and television stations may pay remuneration to the copyright owner for the broadcast of published musical works, ......the amount and so on" It can be seen from the above provisions combined with common sense that the musical works broadcast by radio and television stations refer to the musical products of the performance of musical works, and no matter how beautiful the music is, it is only through the performance that people can hear, see, and feel the beauty, otherwise it is only a musical symbol listed on paper. In fact, the "work" referred to in the broadcasting right is a general concept, which generally refers to works in any form, including works that are performed. The dissemination of "performance of a work" in the performance right is a small concept that cannot cover the dissemination of the original form of the work that has not been performed. Thus, the first sub-right of the broadcasting right could cover the non-interactive remote dissemination of works in any form, including dancing and singing in live webcasts, etc. The performance right in mainland law is the right of on-site transmission, which includes human performance and mechanical performance, and the remote transmission of "performance of the work" has nothing to do with the performance right. For the remote transmission of songs generated by AIGC, if the audio of "AI singing" is played on radio, television or online live broadcast, it is non-interactive transmission and is controlled by the broadcasting right, and if someone provides the on-demand or download of the "AI singing" audio on the Internet, it is interactive transmission and is controlled by the information network transmission right.
4. Performers' rights regulate the songs generated by AIGC
(1) There is no performance in the songs generated by AIGC
The performer's right is an important part of the neighboring rights system. According to the provisions of the WIPO Performances and Phonograms Treaty and the Beijing Treaty on Audiovisual Performances, a "performer" refers to a person who performs the expression of literary, artistic and folklore works in various ways. The performer's right belongs to the performer, with the performer as the main body, and is generated based on specific performance behavior. The so-called "specific performance": first of all, each performance of the performer can generate an independent performer's right, even if the content of each performance is different. Second, the performer has to choose to perform an act according to his or her free will in order to form a performance right. There is no way to talk about performers' rights without performative acts.
(2) AICG generates songs without the right of performers
Songs are valuable and used to disseminate. Since the emergence of AI-generated works in the world, the academic community has been discussing the issue of copyrighting AI-generated works. The proponents argue that AI is copyrightable as long as the requirements for originality and reproducibility of the work are met. This view asserts that the subject of natural person authorship is not a necessary factor in identifying the work. The opposing side argues that AI is the result of running algorithms, rules, and templates, and cannot reflect the unique personality of the creator, and at the same time, it is not an expression of the emotions of human subjects. Although the generated content has the appearance of a normal work, it lacks an internal personality foundation. The validity of any point of view must be universal. The selfie of the macaque in "Macaque Selfie Case" cannot be a work, and the author of the work can only be a natural person. From the perspective of the creation of copyright laws in various countries, copyright laws encourage people's creation, and people get paid after creation, so as to encourage people to recreate, and promote the continuous development and evolution of human intellectual achievements. Therefore, no matter how artificial intelligence develops, we must not lose the original intention of the establishment of the Copyright Law: to encourage human creation and respect human value. Objections are clearly able to meet this requirement. This view can also explain the AIGC sound synthesis technology. Monkeys can dance and even dance when they are drunk, but we cannot say that monkeys have the right to perform; can parrots learn to recite a poem after training, is a parrot a performer? Can they thus obtain performer's rights? The answer is no. Copyright encourages people to create, and performers' rights also encourage people to engage in performances. In addition to the above two points of view, there is also an artificial intelligence tool that says: This view holds that AI is a tool for humans to create works, and that AI creations are actually human creations, and that the creations are ultimately judged to be works based on copyright standards. According to Article 11 of the Copyright Law, the legal person or unincorporated organization behind the AI is deemed to be the author. However, the premise of considering the unit as the author is that there is a work created by a natural person, and then whether the work meets the constitutive requirements of a legal person's work is determined. Article 11 of the Copyright Law only addresses the issue of ownership of rights, but does not resolve the issue of identification of works. The problem of generating songs for AIGC is simple, and there is no need to create the "people" behind it. Article 37 of the Copyright Law of 2010 includes performers and performers, but after the 2020 amendment, the situation of "performers" as performers has been abolished, and performers can only be natural persons participating in performances, and no other person else. Therefore, AIGC does not generate performer rights for songs generated by AIGC, and even if others use AI technology to synthesize songs, the performers behind the scenes cannot generate performer rights.
(3) The songs generated by AIGC do not damage the image of the performance
Judging from the interpretation of "distorting the image of the performer" in the "interpretation" of the Legislative Affairs Committee of the National People's Congress, as far as the singer is concerned, this mainly refers to "changing the voice image in the recorded record, causing it to lose its original style and be tampered with", thus causing the public to misunderstand the image of the performer or damage the reputation of the performer. However, AI technology is used to simulate the singer's voice model for performance, and in the process, the singing voice of the existing singer is not modified, but only a new singing voice is generated by AIGC technology. From a teleological point of view, content generators initially hope to realistically imitate or approach the singer's original singing voice and genre through technological means, rather than to tamper with and distort them. Moreover, when the publisher released the imitation of the singer's singing voice with AI technology, it had clearly informed the public that the song was not sung by the singer's own voice, but the voice imitated by AI, and the public only appreciated the audio out of novelty or curiosity, and did not have the misunderstanding that this was the singer's own performance, so even if such an imitation is difficult to reach the singer's own singing level or style, it is difficult to constitute a distortion or tampering with the singer's existing performance image, and it does not harm the singer's personality interests as a performer for his performance。
5. Infringement of personality rights by AIGC-generated songs
For the protection of personality rights, there is a separate section on personality rights in the Civil Code to regulate, and there are relevant provisions for the protection of names, portraits, honors, reputations and even voices. However, there is still a boundary of protection—a balance between the free development of other people's personality and the supervision of public opinion. For example, Article 999 of the Civil Code stipulates that: "Where news reporting, public opinion supervision, and other acts are carried out for the public interest, the name, title, portrait, personal information, etc., of the civil subject may be reasonably used." "Specific personality rights, such as portraits, names, etc., are external symbols, which internalize personality interests. The reference or meaning of the symbol is the natural person who is the owner of personality rights, and the application scenarios of the symbol lie in objects and activities. AIGC-generated songs are the product of symbolic application scenarios, and their generation, dissemination and social impact are bound to be regulated by personality rights.
(1) The right to name controls the AIGC songs
When the publisher releases the relevant audio after producing the AIGC song, it declares to the public that it is the song sung and released by the person who imitates it, and falsely claims that the performance work that is not the performer's own performance is the performer's personal performance, which is obviously an act of impersonating the performer, and this move undoubtedly infringes the performer's right to name. There is also an understanding that such an act is an act that infringes the performer's right to "identify the performer" under Article 39 of the Copyright Law. At the same time, Article 53 prohibits "the production and sale of works that counterfeit the signature of others", and the act of passing off also constitutes infringement. However, in "AI Singer", the publisher clearly stated that the song was created through AI technology at the time of release, and stated that the song was not sung and performed by the singer himself, and the "AI Singer" was not the singer himself. What is certain is that such a label obviously does not cause misleading the public, and the public will not mistakenly think that the AI singing is really the singer singing, but the publisher labeling "AI singer" literally compares it with the real "singer" In fact, it still takes advantage of the singer's popularity as a singer among pop song lovers and the commercial value of the singer's name in the music market, so this is a damage to the property interests in the singer's name, and there is still a possibility of infringing on his name rights. If the publisher does not label the "AI singer" when releasing such a singing video, and does not use such a "gimmick" to attract attention, how much attraction will it have to the public to appreciate? This is where the commercial value of a celebrity's name lies, and the property interests of the name right should also be protected by law.
(2) Protection of voice rights in AIGC songs
AIGC generated songs are made using other people's voice models and are essentially sounds capable of producing specific connections. Sounds can be protected by trademark law by registering a sound trademark. At the same time, the Civil Code of the Mainland China stipulates that "the protection of the voice of a natural person shall be governed by reference to the relevant provisions on the protection of portrait rights", and as a voice model of personality characteristics, it is also an external image that can be identified by a specific natural person, so it enjoys the protection of personality rights. Referring to the provisions on the protection of portrait rights, no one is allowed to make, use or disclose his or her voice without his or her consent, and it is not allowed to defame, deface or falsify his voice.
Voice, like portraits and names, can become a symbol of a person, allowing the public to associate with the sound and fix it to the owner of the voice. However, it does not mean that this sound cannot be produced or used by others. The German scholar Lei Bingde once said: "In the copyright law, the act of imitating artistic performances is free. In fact, the common "live-action parody" imitator does not infringe on the rights of the performer being imitated. The plastic surgeon does not infringe on the portrait rights of the person being imitated. In the same way, if a real person imitates the voice and appearance of another person, it will not infringe on the rights of the person being imitated, so artificial intelligence, as a machine that simulates the voice model of others to "sing", obviously does not infringe on the rights of the performer. Using AI technology to simulate other people's voices to make songs, this production simulation will not cause ugliness or defacement of the singer's voice image. In the case that the publisher declares that the song was completed by AI at the beginning of the release, then the publisher does not constitute confusion about the sound model, and it is not to say that it is forging other people's voices to deceive the public. The stable connection between the sound model and the owner is the effort behind the natural person and the value of human labor. By using another person's voice model to make a song, the producer uses the commercial value behind the specific voice model to arouse the public's interest in appreciating singing with this voice, and this behavior will damage the property value of the voice, so it constitutes an infringement of the personality rights of his voice.
Conclusion: The value orientation of the copyright law to the songs generated by AIGC
AIGC generated songs have sprung up, which has had a great impact on the traditional music industry. On March 16, 2023, the Recording Industry Association of America (RIAA), together with more than 30 social groups, jointly launched the "Human Art Movement", calling for ensuring that the development and use of AI is always committed to promoting the development of human culture and art, rather than replacing or eroding human creation. The giant wheel of history is rolling forward, the "AI music era" is in the ascendant, and technological innovation and industrial development coexist with opportunities and crises for creators. But under the law of survival of the fittest, we must accept everything that artificial intelligence gives us. On May 23, 2023, the mainland promulgated the Interim Measures for the Management of Generative AI Services to regulate the development of AI, but this also indicates that the law affirms the value of AI. The emergence of AIGC-generated songs has greatly promoted music creation and commercial development, but it has also brought challenges to existing legal norms. There is no doubt that such songs can promote creative and innovative forms to a greater extent, so the copyright law should play a role in escorting and promoting development. Scientifically solve the problem of imbalance of interests in the content industry that may be brought about by "AI singer covers", and build a reasonable return mechanism for intellectual labor in the field of copyright. In the face of the existing infringement risks and problems, it is necessary to better implement the regulatory system to prevent the wanton production and arbitrary dissemination of AICG songs, and maintain compliant and effective production behaviors. At the same time, through a variety of cooperation methods, we can obtain the authorization of the original song and the imitated performer, reduce the risk of infringement, establish a good creation platform, and let the songs generated by AIGC germinate more creative vitality.