Behind the “AI Stefanie Sun” screen-sweeping incident, how does AIGC face the pain of infringement?

Introduction: The initial application of AIGC technology has been popularized on the Internet recently, but AI voice-changing technology has been warned of infringement. The author talks about the copyright issues that need to be paid attention to in the application of AI technology, and points out that only by paying attention to these issues can the healthy development of AIGC be guaranteed.

The unpopular singer Stefanie Sun is “retiring”, but AI Stefanie Sun has quietly become the traffic code for the entire network.

"In one night, the number of views reached more than 100,000. After posting less than ten videos in a row, the number of fans exceeded 1,000." Xiao Su, who has been a UP host on Bilibili for less than a month, is one of the people behind the AI ​​Stefanie Sun. Obviously, she is very satisfied with her content data. "It's great for a new account to achieve this level."

"In four hours, the number of views on all platforms exceeded 500,000."

Qiqi also creates content for AI singers on platforms such as Bilibili and Video Account. The growth of content data continues to stimulate her creative enthusiasm.

The technological changes triggered by ChatGPT spread rapidly, directly impacting the field of content creation, and even turning AIGC into an important part of the content platform. The AI ​​Stefanie Sun and AI Jay Chou that have emerged on Bilibili and Douyin are showing a trend of reviving the Chinese music scene. It is undeniable that the replacement of technical tools has once again lowered the threshold for content creation, and this more participatory and even somewhat curious content has quickly spread across the entire network.

However, behind the national craze, AI-generated content has always had major inherent problems - copyright disputes and content security.

Li Yan, founder of Kuaiyin, used AI Stefanie Sun as an example to explain, "You will see AI Stefanie Sun singing Jay Chou's songs on content platforms, but this kind of content not only infringes on Stefanie Sun's voice, but also involves copyright infringement of the songs. This problem is still very serious in the music industry."

On May 9, Douyin issued an initiative against AI-generated content, prohibiting the use of generative AI technology to create and publish infringing content, including but not limited to portrait rights, intellectual property rights, etc., and will impose strict penalties once discovered; in addition, it also requires publishers to clearly mark AI-generated content and add a unified official "watermark".

The wildly growing AIGC is about to face its first baptism of industry self-discipline.

01 AI Stefanie Sun became popular quietly

"In the past one or two months, it has been very easy to become popular by making this kind of content. AI Stefanie Sun's video has even been popular on Bilibili."

In addition to Bilibili, the trend of AI singers is also growing on short video platforms such as Douyin and Kuaishou.

Xiao Su has summarized a set of logic for AI singers to become hits. "Generally, it is a powerful combination of superstars, such as AI Stefanie Sun's cover of Jay Chou and David Tao's classic songs." The AI ​​Stefanie Sun and Jay Chou's duet version of "Peninsula Iron Box" has been played more than 800,000 times on Bilibili. In addition, Xiao Su said that using the singing style of superstars to cover popular Douyin hits is also likely to go viral because of the novelty and contrast. AI Jay Chou's cover of "Wumeizijiang" is one example.

300 online viewing, image source: screenshot from B station

AI cover singing is not a new concept, but the change of AIGC technology has really added fuel to the explosion of AI singers. Guo Jing, CEO of Time Domain Technology, explained to the host the creation process of AI singers, "AI singers mainly involve two technologies: singing voice synthesis (SVS) and singing voice conversion (SVC). Singing voice synthesis is to let AI singers sing the music score input by users directly, and singing voice conversion is to input the voice of an "original singer" and convert the original singer into the voice of the target singer, which is similar to the voice version of AI face-changing. At present, open source models for singing voice synthesis and conversion are not difficult to find. Xiao Su uses the overseas project "so-vits-svc".

Among all AI singers, AI Stefanie Sun stands out because of her unique voice and the technical logic. Li Yan revealed that Stefanie Sun's voice was the first to be open sourced. "The open sourcer first used Stefanie Sun's voice to demonstrate his own solution, which naturally attracted a group of users who copied it." He said that the discussion about AI singers began as early as March 11, and someone open-sourced Stefanie Sun's voice in early April.

However, despite the popularity of AI Stefanie Sun's videos, Xiao Su does not feel safe. "Every video description needs to clearly indicate that the video content is generated by AI tools, is not for commercial use, and will be deleted immediately if it infringes." In addition to the creators' own active annotations, some platforms will also give warnings under the video to remind users to identify AI content. The reason behind this is naturally due to copyright and security considerations. "Copyright supervision requirements are getting stricter and stricter. I am very worried that one day the platform will remove all my content."

Disclaimer in a video introduction Image source: Screenshot from Station B

At the same time, how to maintain the popularity of AI Stefanie Sun is also a problem faced by creators. In Guo Jing's view, the popularity of AI Stefanie Sun is more about consuming the concept of AI. "For most users, the fact that AI can make Stefanie Sun cover Jay Chou's songs is more important than whether the songs themselves are good. In the AI ​​world, achieving this kind of 'once in a lifetime' is very attractive to users, and the result of its popularity is naturally expected."

At the same time, Guo Jing also affirmed that from a long-term perspective, AI Stefanie Sun has verified the significance of AI technology as a productivity tool. "Voice synthesis plus sound conversion can allow more voices to sing more songs, thereby greatly expanding the diversity of the content space."

In addition to copyright concerns at the creator level, the platform naturally needs to bear corresponding responsibilities for AI content. "Neither record companies nor singers will target ordinary users, because users' behavior is highly entertaining, and these users are even fans of singers. However, singers can hold content platforms accountable." A music product manager explained why Douyin vigorously regulates AIGC content.

According to this logic, other content platforms will inevitably accelerate to keep up with the pace of regulation, and the wild growth stage of AIGC content will finally come to an end.

02 Watermarking and labeling, AI-generated content needs to be warned

In this wave of artificial intelligence, the source of humanity's sense of crisis is the rapid evolution of AI, which even has the ability to make the fake look real.

The AI ​​painting that was criticized for incomplete hand details two months ago has been upgraded in a short period of time. It can not only compete with Taobao models for jobs, but can even disguise itself as a beauty blogger. Otakus are unable to distinguish the true face of AI and instead fall for it.

In the process of AI evolution, the concept of "digital life" has also begun to take shape. First, a tech ape used AI to revive a deceased relative, which became a hot topic. Then, there was a cross-time conversation between AI Musk and AI Jobs, and a poetry debate between AI Li Bai and AI Lin Daiyu.

Image source: Video account AI.Talk

These rather confusing contents are both innovative in content form and fresh in technological change, and naturally become the "traffic code" of the entire platform.

Kuaiyin has counted the user recognition of AIGC content. Li Yan introduced that Kuaiyin mixes the AI ​​version of music content with the original version and distributes them to users. According to the data from the user side, 30-40% of the content users have better indicators than the original version.

"With the support of current technology, the quality of AI-generated content is already high enough. At the same time, AI-generated content is still rare and will bring a sense of freshness to users. Therefore, users will have high feedback on AI-generated content."

However, in the context of the explosive development of Internet communication, AI-generated content that is difficult to distinguish between true and false has made the information environment more complicated. As the technology sinks into the context of the whole people, uncontrollable factors have also increased. It can be said that the higher the popularity of AI-generated content, the greater the hidden concerns it will cause, and the impact on the boundaries of privacy and content security is unprecedented. Therefore, the supervision of content platforms has become particularly important.

At present, the solution provided by Douyin is to add watermarks and unify the labeling. In terms of content form, the text content needs to be clearly marked with the words "AI generated", and the video content also needs to add a watermark of a specified proportion on the canvas to highlight the AI ​​logo. During the push process, the platform will also add a notification label "This content is generated by AI" in the information flow, video details page and other parts.

Image source: Douyin Rules Center Official Account

The explicit watermark has become the first "gatekeeper" in the dissemination chain of AI-generated content, reminding users in the form of an identification to treat the content with caution, which can avoid the chaos caused by AIGC in the content field to a certain extent.

Regarding AI logos, Li Yan added an observation that after adding words such as "AI creation" or "AI generation" to titles or texts, user data has declined significantly. For content creators, how to deal with the impact of such watermarks on content has become an issue that must be considered.

For the platform, before AIGC content supervision has formed clear provisions, this identifying "watermark" has put an end to the barbaric development of AIGC content creation and formed a relatively safe content creation environment within the platform. At the same time, this can also be regarded as a disguised disclaimer for the platform to escape from the risks of AI-generated content.

03 AIGC is accelerating on the right track

TikTok fired the first shot in AIGC's content supervision.

On May 9, short video content platform Douyin issued an initiative for AI-generated content. In addition to requiring publishers to clearly mark AI-generated content and add a unified official "watermark", Douyin also strengthened the restrictions on virtual people, requiring virtual people to register on the platform and obtain a unique "ID card". At the same time, users of virtual human technology also need real-name authentication.

As an industry that has already grown to a considerable scale, Douyin's regulatory actions on virtual people are timely. The outbreak of virtual people can be traced back to 2020. The development of technologies such as AI, cloud computing, and 5G has unlocked the value of virtual people's content, marketing, and even industrial development.

However, as a new thing, a series of problems inevitably arise in the process of virtual humans moving towards industrial development. In addition, after generative AI technology reaches the public level, the governance of virtual humans has become more complicated.

The first is the general personality rights of virtual people, such as portraits and voices. With the support of AI painting tools, the technical threshold has been lowered again, and it is not difficult to design a set of virtual images exclusively for ordinary users. However, the convenience of technology has blurred the boundaries of rights. AI technology has become a tool, and the personalized content of users available on the Internet has become the material. Eventually, there have been chaos in the abuse of other people's images and data security, which have become the companion problems of technological openness.

At the same time, as the era of everyone having a virtual human arrives, the entry threshold for content creators will be lowered, which will easily lead to the uneven quality of content such as virtual human live broadcasts and videos.

"In general, the control of Douyin will mainly affect some small companies and ordinary users operating in the gray area," virtual human operator Bo Ming told Douzhu. "In order to save costs, these small companies and ordinary users will take shortcuts in technology and supervision, and even try to stand out in content. This is a common way of playing."

In Bo Ming's view, "real-name authentication will not have much impact on virtual people who have already formed a mature operating model. On the contrary, it will effectively clean up the development environment of virtual people, eliminate the malignant factors affecting the development of the industry, and allow virtual people to develop healthily."

The real-name authentication initiative is actually leveraging the power of the platform to screen the producers of virtual humans, to govern the virtual human industry with increasingly lower barriers to entry, and to form a virtuous cycle in the content ecology.

From the initial popularity caused by technological changes to the curiosity of the public to try it out, and then to its final landing in a new form of content with healthy development, AIGC is bound to go through a period of development pain and be constrained by the power of supervision before it can finally move towards a formal development path.

Author: Bukong

Source: WeChat public account "Dou Jiao TMT"