OpenAI builds content moderation system based on GPT-4

OpenAI, the company behind ChatGPT, is developing an artificial intelligence-based content moderation system. Based on GPT-4, its latest and most powerful model to date. The company is testing it in-house to ensure compliance with its own content policy. traffic light.

The company explained that the method adapts to any template with which you want to regulate a specific platform. The process appears to be quite simple. First, moderation rules are introduced into GPT-4. Its performance is then tested on a small sample of the problematic content.

The human evaluators then need to analyze the successes and failures of the artificial intelligence. In cases where the system has failed, verifiers can ask GPT-4 why it made this decision. And based on that input, correct and retrain the model to make it more and more accurate.

“Shortens the content policy development process from months to hours, and you don’t need to hire a large group of moderators for this,” said Lillian Weng, director of security systems at OpenAI. Weng noted that the system can work for social media moderation and e-commerce platforms. Some of the company’s clients are already using GPT-4 for content moderation.

OpenAI notes that the model is currently You can only check texts. However, they are already working, so soon you will be able to view images and videos as well. It would be another option to stop, for example, child pornography or disinformation campaigns, two of the biggest problems for platforms like Instagram or Twitter.

OpenAI uses GPT-4 for moderation

Until recently, OpenAI used traditional content moderation methods. traffic light last year it was reported that the company had hired an outsider to scan images created by DALL-E.

At the beginning of the year Time reported that OpenAI is paying workers in Kenya to help flag offensive content. Thus prevent users from viewing it in ChatGPT. These workers were paid between $1.32 and $2 an hour. These human assessors also reported psychological trauma. They had to read and evaluate detailed descriptions of situations involving child sexual abuse, murder, suicide, torture, and incest.

A study by the University of Zurich published last March found that the use of ChatGPT artificial intelligence for content moderation was up to 20 times cheaper than hiring a person.

ChatGPT is currently moderating itself. There are still people involved in the process, both to update policies and to test edge cases. However, the new system is expected to significantly reduce the number of people involved.

OpenAI knows that this is not yet the final decision. “We can’t build a system that is 100% bulletproof from scratch… But I’m pretty sure it will be good,” Weng said.