AI startup OpenAI is testing a content moderation system based on GPT-4, which is expected to significantly reduce content moderation time.

On August 15 local time, OpenAI posted an article on its official website saying that the test of using GPT-4 for content policy development and content moderation decisions has achieved good results, achieving more consistent labels, faster policy optimization feedback loops, and reducing human moderators’ involvement. OpenAI believes that the GPT-4 content moderation tool can help companies complete about six months of work in one day.

OpenAI said that content moderation has always been a time-consuming and challenging task, requiring a high degree of meticulousness, sensitivity, deep understanding of context, and quick adaptation to new cases, and also causing an emotional burden to human moderators. Usually, changes in content policy take months to implement in the entire system due to the need for a large amount of human intervention and adjustment. In contrast, GPT-4, as a large language model (LLM), can shorten this process from months to hours.

In the specific training process, after the policy guidelines are formulated, policy experts can create a set of golden data by identifying a small number of examples and assigning labels to them. Then, GPT-4 will read the policy guidelines and assign labels to the same data set without seeing the answers, and the experts will check whether these labels are assigned correctly.

Policy experts can also ask GPT-4 to give the reasoning behind the labels, analyze any ambiguities that may exist in the policy guidelines, and make corresponding clarifications in the guidelines. These steps can be repeated until satisfied with the quality of the AI review.

In addition, to process large amounts of data on a large scale, a smaller model can be fine-tuned using GPT-4’s prediction function before formal processing.

OpenAI has been testing the technology and welcomes users to join the training. Anyone with OpenAI API access can use this method to create their AI-assisted review system.

However, OpenAI admitted that GPT-4’s labeling quality is similar to that of human moderators with mild training, but still lagging behind experienced and well-trained human moderators. In this regard, OpenAI emphasized that the review process should not be fully automated. AI still needs human leadership in the new content moderation system.

It is worth mentioning that before this, there have been several AI-driven review tools available, among which the more famous one is Google’s Perspective API, jointly developed by Google’s anti-abuse technology team and Jigsaw and used by a series of world-renowned publishers, platforms and research institutions including Reddit forum, The New York Times and The Financial Times.

However, compared with using review tools, many companies still outsource the review work completely to other manpower companies to save research costs. According to The Wall Street Journal on July 24th, in the past few years, OpenAI itself hired a large number of cheap laborers from Kenya to clean up ChatGPT’s data.

And just a week ago on August 10th, Indian media Analytics India Magazine just released an astonishing inference that OpenAI may go bankrupt by the end of 2024 because it costs OpenAI $700,000 per day just to run ChatGPT. Musk’s newly established xAI and GPU shortage problems are even worse for OpenAI which has not yet made a profit.

According to the website data analysis tool, the SimilarWeb survey shows that ChatGPT’s growth rate has been declining in the past few months. The number of users in July was 1.5 billion, down 12% month-on-month.

At the end of January this year, OpenAI launched an AI tool AI-Text-Classifier (AI text classifier), which is designed to identify text content generated by AI. However, less than half a year after its birth, this new AI tool was shut down by OpenAI in July for “low accuracy”.