Um die chaotischen Zustände von AIGC zu bewältigen, ist das Einführen von AI-Kennzeichnungen nur der Anfang.

More than half a month has passed since the implementation of the new regulations on AI identification, and some changes are taking place.

On content community platforms such as Douyin, Weibo, and Xiaohongshu, pictures and videos generated by AI are clearly marked by each platform as "AI-generated content". Meanwhile, on visual content trading platforms including Visual China, AIGC content is also clearly marked.

However, according to Jiemian News' observation, some content community platforms are not accurate enough in identifying AIGC content, and some AI-generated pictures are not marked.

According to the new regulations, AIGC content needs to include either explicit or implicit identification. Explicit identification is a mark that users can clearly perceive. Currently, the common practice of various AIGC products is to add a watermark to pictures or videos. Implicit identification is a mark added to the data of the synthesized content file, which is not easily perceived by users. In actual application scenarios, many users are reluctant to add explicit identification for the sake of the aesthetics of their content works.

A technician from a large model company revealed to Jiemian News that currently, the implicit identification of AIGC cannot prevent "intentional or unintentional" damage, which is the biggest current technological bottleneck. "One possible situation is that the implicit identification is deliberately tampered with or deleted; another situation is that the implicit identification of relevant content is lost during continuous dissemination."

The person in charge of the content governance team of a certain platform said that among the false information brought by AIGC, there is a large amount of text-based content in addition to pictures and videos, and text-based AIGC cannot be identified through explicit or implicit identification.

Since the beginning of this year, false information generated by AI has frequently stirred up the public opinion field, and more and more people have begun to pay attention to the impact of AIGC on the content dissemination ecosystem. In fact, with the widespread application of AIGC, various disputes related to copyright and personality rights have also increased explosively in the past two years.

These chaotic situations are forcing the introduction of more industry regulations. Chai Jijun, CEO of Visual China, told Jiemian News that as regulators and industry participants continuously refine and optimize the problems found in practice, there will be positive changes in the governance of chaotic situations.

Implicit identification needs unified standards

Zhang Wang, vice president and general legal counsel of SenseTime, explained in detail the technical principles behind explicit and implicit identification during an interview with Jiemian News. He said that in essence, both types of identification are "post-written" information, but they are different in technology.

Specifically, the technical principle of explicit identification is relatively straightforward and belongs to post-processing technology. It only needs to mark "AI-generated/synthesized" on the corresponding pictures and videos, and the height of the text should not be less than 5% of the shortest side length of the picture; the technical principle of implicit identification is more complex and belongs to information embedding technology. Taking JPG format pictures as an example, SenseTime's AIGC service platform fills the string into the picture metadata field based on the EXIF specification, which includes multiple pieces of information such as AIGC synthesis attribute information, elements of the generation and synthesis service provider, and elements of the content dissemination service provider.

Zhang Wang said that since the "Measures" have been in effect for a short time, it will take some time for various dissemination platforms and content production platforms to reach mutual recognition of implicit identification, and there may be some challenges in the mutual recognition of implicit identification at present.

Globally, there is no unified technical standard for the implicit identification of AIGC content that is widely adopted. Currently, the existing implicit identification technology schools include those based on changing pixels/frequency domains, "birthmarks" based on the generation process, and modifying metadata. China adopts the metadata technology school.

Chai Jijun told Jiemian News that although there is no unified standard, the global trend is the same, that is, AIGC content must be identified. China's AIGC content identification standard is promoted from top to bottom by national regulations, which is bound to bring the advantage of fast implementation.

He also said that currently, the technical standards for implicit identification of content production platforms at home and abroad are not unified, which also poses some challenges to platform identification of AIGC content. "If a globally applicable watermark system is established, it will also face many challenges. For example, whether its detection algorithm is public or private is a dilemma in itself."

Automatic watermarking of AIGC pictures (Source: Generated by Doubao App)

Regarding the current situation where implicit identification information is easily lost or tampered with, Chai Jijun also said that adding a powerful and reliable watermark function to AIGC services requires real costs, which is also a relatively big challenge at present.

Although there are many challenges in identifying implicit identification, currently, all platforms are trying their best to implement the regulations in the links they can control to avoid chaotic situations. For example, content dissemination platforms such as Weibo and Douyin are improving their ability to identify AIGC content through technical means and minimizing the spread of AIGC false content through operational means.

Chai Jijun revealed that Visual China does not collect AIGC works on news reporting topics. For design and creative topics, Visual China accepts AIGC works and regulates those involving portrait rights and property rights. At the same time, mainly soliciting contributions with clear topics arranged by editorial and operation personnel can avoid problems such as non-compliance of AIGC content to the greatest extent.

The demand for AIGC commercial use is increasing, and compliance issues cannot be ignored

AIGC has greatly lowered the threshold for content production. Not only are more and more ordinary users starting to use AIGC for content sharing and entertainment, but also some brands and e-commerce merchants are starting to use AIGC for advertising and marketing. However, disputes have also arisen.

In August last year, when Xiaomi was preheating its new folding-screen mobile phone, it released a fashionable poster of a female model on its official Weibo. However, the female model's fingers lacked joints and her collarbone was serrated, which was discovered by sharp-eyed netizens to be generated by Midjourney, triggering a lot of controversy. There were also comments such as "plastic people" and "uncanny valley" in the comment section of this Weibo post.

In September last year, Hangzhou Jellyfish Intelligence Technology Co., Ltd., the AI operation entity of TouchPal, was sued by Shanghai Xinchuanghua Culture Development Co., Ltd. for infringement of its information network dissemination rights because it allowed users to upload pictures and generate works similar to the image of Ultraman. The latter has the exclusive authorization of the relevant rights from the copyright owner of the Ultraman series of works. The court finally ruled that although users uploaded the materials, the platform failed to take effective measures to prevent infringement, which constituted contributory infringement, and ordered the defendant to stop the infringement and compensate the plaintiff 30,000 yuan.

Similar disputes have occurred repeatedly both at home and abroad. Chai Jijun told Jiemian News that AI is the core driving force of the content industry in the next decade, and practitioners in the industrial chain will invest fully. However, practitioners also need to realize that issues such as copyright, ethics, and false information associated with AI technology must be avoided as much as possible to prevent potential compliance risks during commercial use.

Jiemian News has counted the disputes and lawsuits caused by the commercial use of AIGC at home and abroad in the past two years, which mainly involve multiple aspects such as data compliance and information disclosure. Among them, there have been disputes between domestic and foreign artificial intelligence companies over the unauthorized use of training data by large models from content platforms. Some large model companies have also used photographers' original works or well-known original anime images during the training process, resulting in the appearance of original work content in the generated pictures.

In terms of information disclosure, one type of dispute comes from the use of AIGC in advertising and marketing materials without the brand side marking or truthfully informing users, which triggers a trust crisis; another type of situation is caused by the unclear data authorization chain. Especially when commercial AIGC content involves a person's portrait, the brand side must obtain clear and sufficient portrait rights authorization from that person.

You Yunting, a senior partner and intellectual property lawyer at Shanghai Dabang Law Firm, told Jiemian News that currently, the relevant legal regulations on copyright disputes caused by AIGC training data in China are not clear, and there are no precedents in judicial practice. Regarding disputes over personal rights, the "Civil Code" has relatively clear regulations on the reasonable use of portraits and voices. Regarding copyright infringement disputes related to comics, the "Copyright Law" has relatively clear regulations.

Chai Jijun said that Visual China is cautious about AI visual content involving real human portraits. In the scenario of the content trading platform, the review mechanism clearly prohibits signed contributors from uploading AI visual content with recognizable real human portraits; in the scenario of content customization, Visual China will use data with portrait rights authorization to generate AI visual content to help customers avoid infringement risks.

He also advised enterprises and brand owners that when using AI tools to generate and use visual content, they should choose compliant platforms that meet traceability requirements to ensure that the content source is clearly authorized. At the same time, enterprises should pay attention to distinguishing usage scenarios and avoid using AI-generated content that may cause misunderstandings or discomfort to consumers.

This article is from "Jiemian News", author: Xiao Fang, published by 36Kr with authorization.