AI has brought unprecedented opportunities and equally unprecedented risks. One of the most alarming trends is the rise of AI-generated deepfakes that impersonate celebrities, influencers, and even ordinary individuals.
In early 2024, scammers used a deepfake of Taylor Swift’s voice to trick her fans into completing fraudulent surveys, harvesting their personal information under the guise of a free cookware set giveaway. In response, Swift has filed sound mark applications for phrases such as “Hey, it’s Taylor Swift” and “Hey, it’s Taylor” in Class 41, which covers entertainment services.
If these had been made in China, how effective would the applications have been? Would they be sufficient to combat different deepfake scenarios? Are there other legal rights and enforcement tools that can be used?
These were among the key questions explored in webinars held by Vivien Chan & Co in June 2026. This article summarises the insights and examines how celebrities, brands, and content creators can leverage China's trademark and personality rights framework to combat the growing threat of deepfakes.
How effective are trademarks in combating deepfakes?
Registering a sound mark in China is possible, but it can be challenging and resource-intensive. So far, only about 5% of sound mark applications have been approved.
The CNIPA generally considers sound marks lacking in inherent distinctiveness. Simple tones, ordinary melodies, and spoken slogans are typically rejected. Applicants must demonstrate that the sound has acquired distinctiveness through extensive and long-standing use.
Li Jiaqi, who earns $1.5 million a year as a beauty influencer, filed his catchphrase “Oh my god, Buy it Buy it!” as a sound mark in Class 35 in 2020 for advertising and sales promotion services. The result? The application was refused for non-distinctiveness as the sound mark would be perceived as an advertising slogan. Although evidence of prior use was submitted, the CNIPA was not persuaded that the sound mark had acquired distinctiveness through use.
Tencent’s QQ instant messaging app notification sound, consisting of “di-di-di-di-di-di”, was finally approved in 2018 after the Chinese technology and entertainment company appealed all the way to the Beijing Higher People’s Court. Tencent submitted analyses of the sound and extensive evidence of use and public recognition, including annual reports from 2004 to 2015 to show the number of registered and active QQ accounts and a Guinness World Records certification recognising QQ as the instant messaging platform with the most concurrent users. Intel’s five-note jingle, which is commonly played at the end of promotional videos, was not approved until 2024, following an appeal to the Beijing Intellectual Property Court.
These cases show that registering a sound mark can be a lengthy and costly undertaking. Applicants must be prepared to invest time and resources in gathering evidence demonstrating extensive use and public recognition of the sound as a source identifier. In the hypothetical case of Swift, apart from evidence of her own reputation, evidence showing the use of her catchphrases in relation to the applied-for services has to be collected.
Even if registration is secured, enforcement presents another challenge. To combat deepfake false-endorsement videos, Class 35 rather than Class 41 would be the more relevant class. Assuming there were valid registrations in Class 35, Swift would be in a good position to enforce against deepfakes that reproduce the same or a confusingly similar phrase with her voice. However, it may not be effective if other words were said. Therefore, infringers could easily circumvent any enforcement by avoiding the specific words protected by the sound mark.
Accordingly, while sound mark registrations may form part of an anti-deepfake strategy, they are unlikely to be the most effective standalone solution.
Personality rights under the Civil Code
Beyond trademarks, China’s Civil Code provides a robust protection regime for personality rights. Personality rights relevant to the deepfake context include the following:
Portrait rights, which protect a person's identifiable likeness, extend to AI-generated images where the individual remains identifiable; and
Voice rights, which protect identifiable voices against unauthorised use, cloning, imitation, or AI-generated reproduction.
When determining whether portrait or voice rights have been infringed, Chinese courts apply the ‘identifiability test’ – whether the relevant public can identify or associate it with a specific individual. Exact replication is not required. It is sufficient that a deepfake image, video, or voice is capable of causing the public to recognise, identify, or associate the content with that individual. It is enough that a particular segment of the public – such as followers within a specific industry or fan community – can make that association, even if the individual is not readily recognisable to the general public.
In 2024, the Beijing Internet Court handed down China’s first landmark AI voice rights judgment, in Yin v A Certain Product Sales Company, A Certain Sound Recording Owner, and A Certain Technical Processing Party. The case involved a voice actor whose recordings, originally licensed for a specific purpose, were subsequently used to train an AI text-to-speech model without her consent. The court applied the identifiability test, holding that voice rights can extend to AI-generated voices where the synthetic output is sufficiently similar in timbre, intonation, and speaking style that the relevant public could associate it with a specific individual. The court also held that a licence to use a sound recording does not automatically include the right to use the underlying voice for AI training, as a person’s voice constitutes a protected personal identifier distinct from the copyright in the recording itself.
The identifiability test was also applied in Li v A Certain Cultural Media Co., Ltd. (2025), the first AI deepfake celebrity voice infringement case in China. An online platform used the plaintiff’s portrait together with an AI-generated voice closely resembling her own to market educational books. The Beijing Internet Court found infringements of portrait and voice rights, applying the same test to conclude that the synthetic voice was sufficiently connected to the plaintiff that viewers were likely to associate the promotional content with her and perceive her as endorsing the products. It is worth noting that the plaintiff in this case was not famous and was only known within the education industry. However, the courts still ruled that this was sufficient, showing that it is enough that the relevant public within the sector is likely to connect the content with the plaintiff.
Most recently, in Dilraba Dilmurat v Hangzhou Qianyan Wanyu Culture Communication Co., Ltd. and Fuyang Zhishang Culture Media Co., Ltd. (2026), a landmark AI face-swapping case, the Beijing Internet Court considered whether an AI-generated image must be an exact replica before portrait rights can be infringed. The court held that exact replication is not required. Rather, the relevant question is whether the general public, or a relevant segment of the public, can recognise the individual from the AI-generated image. In reaching its decision, the court relied on evidence of the celebrity’s public recognition and online discussions in which viewers identified the face-swapped character as the plaintiff. This case confirms that the identifiability test focuses on public recognition and association, rather than technical similarity alone, and that personality rights may be invoked even where the AI-generated content differs in certain respects from the person’s actual appearance.
Another interesting case is Liao Mou v A Certain Science & Technology Culture Co., Ltd. (2024), in which the Beijing Internet Court distinguished between portrait rights and personal information rights in the context of AI face-swapping. The court held that there was no infringement of portrait rights because the plaintiffs were no longer identifiable after their faces had been replaced by users’ faces. However, the defendant’s use of the plaintiffs’ facial features to create AI face-swap templates constituted an infringement of their personal information rights, which are also protected personality rights, as facial features are biometric personal information protected by law. The case highlights that even where the final AI-generated output is not identifiable as a particular person, the unauthorised collection or processing of biometric data may give rise to liability.
These cases demonstrate that the various personality rights, portrait rights, and personal information rights are broadly protected in China and therefore can be relied upon in the protection against deepfake content.
Deepfake regulation and enforcement
China has developed a multi-layered framework for combating deepfakes, combining private enforcement with increasing regulatory oversight. Rights holders can seek the removal of infringing content through the complaint mechanisms of major e-commerce and social media platforms, while more serious or repeat infringements may be escalated to government authorities – including the Cyberspace Administration of China (CAC), the Administrations for Market Regulation (AMRs), and the Public Security Bureaus (PSBs) – and Chinese courts.
The Deep Synthesis Provisions, effective from 2023, require service providers to:
Obtain appropriate consent when manipulating another person’s voice or facial features;
Implement content governance measures; and
Label certain AI-generated content.
The AI-Generated Content Labelling Measures, effective from 2025, further mandate visible AI disclosures and embedded metadata for a broad range of AI-generated text, audio, images, videos, and virtual content. The CAC has actively enforced these rules through nationwide campaigns targeting voice cloning, face swapping, misinformation, and unlabelled AI-generated content, and established a complaint channel for internet fraud and misuse of AI technologies.
Deepfakes may also trigger further liability. Where AI-generated content creates a false impression of celebrity endorsement, it may constitute false advertising under the Advertising Law and Anti-Unfair Competition Law, exposing advertisers to administrative penalties. In 2025, an AMR ruled in Beijing’s first AI false-advertising enforcement case that using a deepfake of a well-known China Central Television host in product advertisements created a false impression of endorsement and constituted false advertising.
In more serious cases involving fraud, online scams, misuse of personal information, or deepfake pornography, complaints may be filed with a PSB, potentially leading to criminal prosecution. In a 2023 criminal case, an individual used AI face-swapping technology to create and distribute pornographic videos featuring celebrities and ordinary individuals. The offender was sentenced to seven years and three months’ imprisonment and fined RMB 60,000, demonstrating that serious misuse of deepfake technology may attract criminal liability.
Together, these enforcement mechanisms provide rights holders with a broad range of tools to address deepfake impersonation in China.
Practical tips for celebrities and influencers
A comprehensive protection regime would include intellectual property filings, contractual protections, and proactive enforcement.
Early registration – register stage names, Chinese translations, and fandom names as trademarks, covering relevant classes such as downloadable music, entertainment, and sales promotion services. Stylised logos may also be recorded as copyrights to provide prima facie evidence of ownership and existence.
Collection and preservation of reputation evidence – reputation is critical to the enforcement of personality rights. Celebrities and influencers should systematically collect and preserve evidence of their public profile. Vivien Chan & Co regularly assists with reputation evidence collection, including conducting searches of the National Library of China for references in print and online publications, arranging notarised visits to promotional events, periodically timestamping social media accounts and music or video streaming platforms, and conducting market surveys.
Robust licensing arrangements – put all licences in writing and include essential terms, such as scope of use, express prohibitions on AI training and cloning, alteration rights, sublicensing rights, pre-approval mechanisms, and termination provisions. If the licence term is unclear, under Chinese laws, either party may terminate on reasonable notice; even a fixed-term licence may be terminated for justified reasons such as a change in personal image or licensee misconduct.
Rapid enforcement – given the rapid spread of information online, timely enforcement is critical to mitigating damage and preventing further dissemination of infringing content. Vivien Chan & Co’s in-house investigators can carry out immediate timestamped captures of online infringements. To expedite complaints to e-commerce and social media platforms, the firm can prepare the required power of attorney, identity proof, and reputation evidence package in advance, allowing prompt submission when infringements are detected.
Protection in the AI age
The Taylor Swift deepfake scandal is a wake-up call. As AI technology becomes more accessible, the risk of impersonation grows exponentially. China offers a rich toolkit of trademarks, personality rights, deepfake regulations, and enforcement mechanisms, but effective protection requires vigilance and a well-planned strategy.
Celebrities, influencers, and brands should seek counsel to navigate this evolving landscape through strategic intellectual property filings, active monitoring, investigations, evidence preservation, and enforcement planning, enabling them to effectively protect their identity, reputation, and brand value in the age of AI.
