New changes in online comments and speech governance in the AI era

A harmful picture deleted by AI

When using Facebook or other social platforms, do you feel free to post everything online with your real name as the username? Have you ever witnessed some content deleted by the platform, and replaced by a notice saying that the content allegedly violated the platform agreement? We live in an era where the development of artificial intelligence technology and the construction of internet culture are both rapid. Judging from the current situation of the Internet environment, the popularity of mobile devices and the broad coverage of high-speed network signals enable people to access the online world anytime and anywhere. Users can use different accounts and avatars to establish different virtual identities on the platform and speak freely. These factors provide a high-quality environment for the development of artificial intelligence technology, and the same time, allow online harm and hate speech to spread to a broader range at a faster speed, harming the interests of others, inciting extreme behavior, and undermining the security of real space from the virtual world.

AI, the helper of human beings

According to the statistics from Kepios, as of January 2023, there are currently 4.76 billion social media users in the world (Kemp, 2023), so it is impossible to rely solely on manpower to censor whether the information posted by numerous users is harmful. 

(Picture from Kepios, 2023)

In 2018, at a U.S. congressional testimony, Facebook founder Mark Zuckerberg stated that AI is the ultimate answer to solving platform content governance issues when facing public doubts about the safety of his platform. Facebook has also proved that its content governance work is highly dependent on AI technology in its following practice.

Facebook’s intelligent content governance system mainly consists of two parts, DeepText and Rosetta. DeepText engine is for censoring text content, and the Rosetta system is for pictures, GIFs, and videos. DeepText has a powerful text library and can recognize text in more than 20 languages. Since released in 2016, its ability to understand text has approached that of humans through the accumulation of learning and identifying thousands of texts per second. Rosetta detects and identifies the text in the image by optical character recognition technology (OCR for short, has been widely used in daily life. The common applying example is to extract text from pictures or pdf files), realizing image analysis and management functions (Borisyuk, Gordo, & Sivakumar, 2018). Facebook specially produced more than 10,000 emoticons and fed them to Rosetta and trained it to understand the meaning expressed by human language combined with pictures. The training result was so effective that Rosetta could understand whether the emoticon package has an offensive meaning. For example, the context “Love the way you smell today” would not be considered offensive by Rosetta if it was accompanied by a picture of a flower, but it would be judged as hate speech if it was accompanied by a picture of a skunk.

Facebook’s content governance system represents the most advanced AI information identification technology in the contemporary era, proving that AI has managed the basic censor work to save manpower. Data shows that the number of successfully removed hate speech and online harm has increased significantly since the system went live. It is undeniable that in the face of the heavy daily information flow, AI, which owns a high speed of learning and working than human beings, has contributed a key force in content governance. We can also expect that after testing and updating multiple text models, it can understand the human culture in a more advanced way.

 AI is not as cunning as humans

However, human culture is profound and complex, and humans are cunning and good at anti-reconnaissance, especially after understanding the logic of AI governing content. There are two obvious disadvantages in the AI content governance system: Firstly, the lack of flexibility in literary understanding leads to misjudgment of content. Here is a vivid example to help understand this. In 2018, Facebook’s content governance system directly judged a post of an excerpt of the Declaration of Independence published by a community newspaper containing the content of “merciless Indian savages” as hate speech and automatically deleted it without considering the full text from a historical perspective, causing controversy (Wolfson, 2018). Secondly, for users with anti-reconnaissance thinking, the AI governance function fails. I have seen many users use emojis, homonyms, blank spaces, old sayings, and other ways to successfully evade detectors and publish hate speech and sensitive topics, which requires AI to have a more flexible understanding of human culture and more diverse cultural reserves.

To solve this problem from a cultural perspective, I think that on the one hand, the R&D team can combine the publishers’ personal information, for example, regional information and gender information with the analysis of the topics that different groups of people are more inclined to talk about to establish portraits of publishers of sensitive and controversial topics. The team needs to first summarize the basic characteristics of the crowd, and then refine the entries from top to bottom, to help AI review the content with focus. For example, the team can analyze which IP region users are from to post geographically discriminatory remarks more frequently, which age and gender group the users belong to are more likely to attack the LGBTQ group, and then classify users into different groups with different governance targets. The generalized entries should fully cover humanistic knowledge such as history and grammar.

On the other hand, we must admit that AI can never be separated from human assistance in terms of cultural understanding. Human flexible wisdom and AI’s large-scale information processing efficiency need to cooperate. I recommend shifting the governance focus of the human censor team to the anti-reconnaissance content area and complex ethical dilemma topics, giving humans and AI different work priorities, avoiding overlapping functions, and improving the efficiency of collecting data and governing content. Human workers should collect content that presents complex ethical dilemmas, organize them into data sets, cultivate AI’s cognition of right and wrong, and gradually establish a model of ethical principles (Wu, Huang, & Gong, 2020). Typical posting patterns that evade censorship also will be collected by human workers and handed over to the AI system for its learning.

No matter from which aspect the content governance system is refined, the AI review process should ensure that it runs through the entire text and that the keyword entries are analyzed in combination with the whole context, which greatly increases the accuracy requirements of AI work. In addition to improving the AI censorship function, I think the platform can also increase the cost that users should bear for publishing harmful information on the Internet after managing to reach a consensus among government departments, legislatures, and the public. For example, with the permission of law enforcement agencies, platforms can use the data collected by the algorithm to target the identity information of the users who post hate speech and harmful comments. According to its degree of harm to determine the degree of punishment. This can reduce the network concealment of users who often engage in inappropriate network activities and regulate users’ online behavior.

AI, beats magic with magic

Although AI technology has made great contributions to facilitating the governance of Internet culture, it is also disrupting the order of Internet culture. With the rapid development of AI technology, the threshold for people to use AI has become lower. Just input commands, and within a few seconds, AI can generate content that is difficult to distinguish with the naked eye. There are more and more cases of misusing AI technologies such as AI face-changing, and AI forging screenshots to create online harm and undermine cyberspace security. That is the reason why many people have begun to show resistance and fear to the application of AI technology.

An entry was featured on the trending search of a Chinese popular social platform Weibo in March this year: #A female’s taking-subway photo undressed by AI was being spread#. The cause of the incident was that in July last year, a girl uploaded a photo of herself wearing normal clothes and taking the subway on the social platform Xiaohongshu. However, after a few months, the girl in the photo was photoshopped naked by someone using AI technology. The “Guangzhou subway beauty exposure incident” became a hot topic on the Internet for a while, and the innocent girl was inexplicably subjected to a wave of online attacks. In response to this incident, many netizens not only expressed anger at the behavior of the rumormonger but also expressed their uneasiness and fear with the rapid development of AI technology. Therefore, to deal with the resistance and fear that many people have begun to show about the application of AI technology, it is urgent to govern the emerging content created by AI programs.

On March 22, 2023, the Future of Life Institute issued an appeal letter titled Pause Giant AI Experiments: An Open Letter to the whole society, in which it advocated that all artificial intelligence laboratories suspend artificial intelligence system projects that are more powerful than ChatGPT-4 for at least 6 months so that all human beings can deal with the impact of AI technology on human civilization and the risks it poses to the public, improve the governance system, and reasonably control AI misuse. This letter has received the support of more than a thousand AI experts and related practitioners, including SpaceX founder Elon Musk and OpenAI founder Sam Altman. The AI algorithm surpasses the human brain’s memory and learning capabilities, making AI technology update and iterate at an inestimable speed. In contrast, the establishment speed of control facilities, policies, and systems cannot keep up with the development of AI. This even makes the creators of AI uneasy. 

Although Stephen Hawking predicted in his speech Guiding AI to benefit humanity and the environment in 2017 that if the development direction of AI is not controlled in time, AI will destroy human society, I think the development of AI will never lead to the decline of human civilization. I agree with Kate Crawford’s understanding of artificial intelligence in the book Atlas of AI. He wrote that AI systems are ultimately designed to serve existing dominant interests (Crawford, 2021). AI is a product of the material world, which is influenced and determined by various natural existences, and created by humans with a wide range of data and rules. AI is neither rational nor autonomous, so it is always humans who endow AI with rules to identify things.

Some companies have begun to exert human initiative and are committed to developing AI software and algorithm programs for the governance of AI-generated content. The AI war has kicked off. For example, Baidu, one of the three biggest internet companies in China, launched an AI deep fake detection system in 2019 to specifically detect content generated by AI technology. Its face deep fake detection function, which has already been launched, can distinguish whether the picture is synthesized by AI technology with a correct rate of up to 99%, and another function of human voice forgery detection is still in the development process. In the process of curbing the abuse of AI, it is not just the AI programmers who are responsible for this battle, but also the platforms. The platforms need to resolutely crack down on AI software that is designed with evil intentions, and curb chaos at the source.

(Sample of Baidu face deep detection function, Anquan Baidu, 2023)

The development of AI technology not only brings infinite opportunities to Internet cultural governance but also brings challenges. Enterprises and platforms must assume social responsibilities, pay attention to the threat of hate speech to social security, and actively promote the development of more advanced AI governance projects, so that hate speech has nowhere to hide. However, ensuring a positive and orderly Internet environment requires not only the constraints of hardware conditions but also the conscious establishment of ethical guidelines by users. As users, we should restrain our behavior to be civilized netizens, and actively report hate speech and harmful content. The government, platforms, and individuals should work together to achieve harmony in the Internet environment.


1. Kemp, S. (2023). Digital 2023: global overview report. Kepios.

2. Borisyuk, F., Gordo, A., & Sivakumar, V. (2018). Rosetta: Large scale system for text detection and recognition in images. Paper presented at the Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining.

3. Wolfson, S. (2018). Facebook Labels Declaration of Independence as ‘Hate Speech’. The Guardian, 5

4. Wu, W., Huang, T., & Gong, K. (2020). Ethical principles and governance technology development of AI in China. Engineering, 6(3), 302-309. 

5. Crawford, K. (2021). The atlas of AI: Power, politics, and the planetary costs of artificial intelligence: Yale University Press.

6. Hawking, S. W. (2017). Guiding AI to benefit humanity and the environmentIntelligent robot, (2), 21-23.

7. Future of Life. (2023). Pause Giant AI Experiments: An Open Letter.

8. Anquan Baidu. (2023). [Sample of Baidu face deep detection function] [Picture].

Be the first to comment

Leave a Reply