In recent years, big tech companies have shown significant advances in areas like language processing, data prediction, and personalization. However, they continue to face difficulties in effectively detecting and removing harmful content on their platforms. This challenge becomes more relevant when we remember how conspiracy theories about elections and vaccines have caused real damage in the United States. So, the question arises: why haven’t tech companies gotten better at moderating content? Can they be forced to do it? And how can advances in artificial intelligence improve our ability to combat misinformation?
This topic has been recently addressed by the Tate Ryan-Mosleyarchive at technologyreview.com, an article that is worth reading and that I am commenting on here.
The complexity of language and the limitations of artificial intelligence
When companies like Meta, Twitter, Google or TikTok have been challenged for spreading hate and disinformation, they often argue that the inherent complexity of language makes it difficult to interpret hate speech across languages ​​and on a large scale. Executives of these companies argue that they should not be responsible for solving all the world’s political problems.
Companies currently use a combination of technology and human moderators to combat harmful content. For example, on Facebook, artificial intelligence detects 97% of the content removed from the platform. However, artificial intelligence is not efficient at interpreting nuance and context, which is also a challenge for human moderators, who are not always good at interpreting these aspects.
Cultural and linguistic complexity also present challenges, as most automated content moderation systems were trained primarily on English data and do not work well with other languages.
Economic interests or lack of regulation?
hany farid, a professor at the School of Information at the University of California, Berkeley, offers a more obvious explanation. According to him, content moderation has not advanced enough due to the lack of economic interest from technology companies. He says it all comes down to greed and we need to stop pretending it’s about something other than money. On the other hand, the lack of federal regulation in the United States makes it difficult for platforms to be held financially responsible for online abuse.
Content moderation seems to be a never-ending war between tech companies and malicious actors. Companies set rules to monitor content, but malicious actors find ways around them, such as using emojis or deliberate misspellings to avoid detection. Companies then try to close these loopholes, but the perpetrators find new ways around the rules, and when blocked they scream CENSORSHIP out loud.
The emergence of generative language models
The advent of generative artificial intelligence and large language models like ChatGPT pose new challenges for content moderation. These models have significantly improved their ability to interpret language, but they also present risks. On the one hand, malicious actors could use generative AI to carry out disinformation campaigns on a larger scale and at greater speed. This represents a concerning prospect, especially given the lack of effective methods to identify and tag AI-generated content.
On the other hand, the new large language models have a better ability to interpret text compared to previous AI systems. In theory, they could be used to improve automated content moderation. However, this would require companies to invest in adapting these models for that specific purpose. Although some companies, such as Microsoft, have begun to investigate in this area, significant progress has not yet been observed.
New tools for a new technology
To address the problem of AI-generated content, it is necessary to develop effective tools that can detect and label such content. Techniques such as digital marking (digital watermarking), which inserts a code to indicate that the content was generated by artificial intelligence, have been proposed as a possible solution. Cryptographic signatures that record information about the origin of the content have also been mentioned. However, these techniques still face challenges in terms of implementation and accuracy.
The new EU draft law on AI requires companies to inform users when content is generated by machines. It is likely that more will be heard about emerging tools for detecting AI-generated content in the coming months, as the demand for transparency in this area increases.
What is clear is that content moderation remains a complex and ever-evolving challenge for tech companies. While there are advances in artificial intelligence and language models, interpreting contexts and nuances remains an obstacle. It’s critical that tech companies invest in robust tools and approaches to combating harmful content, but that costs money, and you have to make investors smile first.
