At OpenAI, we believe our technology can solve one of technology’s toughest problems. Content moderation at scale. OpenAI claims that GPT-4 can replace tens of thousands of human moderators while having nearly the same accuracy and more consistency. If that were true, the most harmful and mentally taxing tasks in the tech industry could be outsourced to machines.
in a blog post, OpenAI claims it is already using GPT-4 to develop and refine its own content policy, label content, and make decisions. “We want more people to act with trust, safety and moderation.” [in] Do it this way,” said Lillian Wen, head of safety systems at OpenAI. Said semaphore. “This is a very good step forward in how AI can be used to solve real-world problems in ways that benefit society.”
OpenAI has three major advantages over traditional approaches to content moderation. First, they argue that while humans interpret policies differently, machines make consistent decisions. These guidelines are as long as a book and are constantly changing. It takes a lot of training for humans to learn and adapt, but OpenAI claims that large-scale language models can implement new policies instantaneously.
Second, GPT-4 is said to help develop new policies within hours. The process of drafting, labeling, gathering feedback, and refining typically takes weeks or months. Third, OpenAI addresses the health of workers who are continuously exposed to harmful content such as child abuse and torture videos.
OpenAI could help solve problems exacerbated by proprietary technology
Nearly two decades into modern social media, and several more years into online communities, content moderation remains one of the toughest challenges for online platforms. Meta, Google and TikTok rely on an army of moderators who must scrutinize scary and often traumatic content. Most of them live in developing countries, are poorly paid, work for outsourcing companies, receive minimal mental health care, and suffer from mental health problems.
However, OpenAI itself relies heavily on clickworkers and human work. Thousands of people, many of whom live in Kenya and other African countries, annotate and label content. Sometimes I worry about emails, sometimes I get stressed at work, and bad salary.
Although OpenAI touts its approach as new and innovative, AI has been used for content moderation for years. While Mark Zuckerberg’s vision of a fully automated system is not yet fully realized, Meta uses algorithms to control the majority of harmful and illegal content. Platforms such as YouTube and TikTok utilize similar systems, so OpenAI’s technology could be attractive to small businesses that don’t have the resources to develop their own technology.
Every platform openly admits that perfect moderation of content at scale is impossible. Both humans and machines make mistakes. That percentage may be low, but it still slips through the millions of harmful posts and hides or removes harmless content.
In particular, the gray area of misleading, false and offensive content that is not necessarily illegal presents a major challenge for automated systems. Labeling such posts is difficult even for human experts, and machines often label them incorrectly. The same applies to satire, images and videos documenting crime and police brutality.
Ultimately, OpenAI could help tackle problems exacerbated by its own technology. With generative AI such as ChatGPT and the company’s image his creator DALL-E, it’s much easier to create misinformation at scale and spread it on social media. OpenAI promised to make his ChatGPT more truthful, but GPT-4 still actively produces news-related falsehoods and misinformation.