Secrets of ChatGPT

OpenAI tool can detect content automatically generated by ChatGPT with an accuracy of up to 99.9%. However, they decided not to publish it.

Not long after November 2022, when OpenAI brought ChatGPT to the web platform and allowed people to use it, many schools banned students from using the tool for fear of cheating.

Students can use AI to generate articles on a variety of topics or solve assignments. To date, there has not been any tool that identifies AI-generated text.

However, according to the WSJ, an “anti-ChatGPT” application does exist.

99.9% Detection Accuracy

OpenAI has been discussing internally the ChatGPT “text stamping” tool for the past 2 years. The WSJ source said that the application was completed from at least a year ago. It can detect text generated by ChatGPT with 99.9% accuracy. The last step to launch is just “press the button”.

ChatGPT operates on an AI system that is capable of predicting which word or phrase (called a token) will appear next in a sentence. The anti-fraud tool developed by OpenAI fine-tunes the way tokens are chosen.

These changes will leave a mark called a watermark,” said the WSJ source, who has read through the document describing the anti-ChatGPT tool.

With the naked eye, these watermarks cannot be seen on the text, but OpenAI detection technology does. The tool even gives a percentage of all or part of the document written by ChatGPT.

Many schools ban ChatGPT because of concerns about cheating. Photo: iStock.

“The likelihood that the Sun will evaporate tomorrow is even higher than that the essay (AI-generated) leaves no mark,” said John Thickstun, a researcher at Stanford who was part of the team that developed the same method of stamping AI text.

However, some OpenAI employees believe that this method can be bypassed by simple techniques such as translating with Google into another language and then switching back or asking ChatGPT to add emojis to the text and then delete them manually.

Who can use the ChatGPT detection tool?

Within OpenAI, the discussion of providing AI with a text detection tool generated by ChatGPT has attracted great interest. Most think this is a difficult choice.

If there are few users, the tool will not be useful. If too many people have access, the bad guys can decipher the watermarking technique, thereby finding a way to get around.

OpenAI considers providing directly to educational institutions or 3rd party companies to help schools identify content written by AI.

Google developed the SynthID copyright stamping tool, which can detect text generated by Gemini AI. SynthID is in the testing phase and is not yet widely available.

OpenAI also has a tool to determine whether images have to be generated by each DALL-E 3 text-to-image application. The company prioritizes watermarking audio and images over text because of the more significant harm, especially during the U.S. election year.

OpenAI also developed a tool to ‘seal’ images created by DALL-E 3. Photo by Zdnet.

In January 2023, OpenAI released an algorithm that helps detect text written by several AI models, including their own. But the algorithm is only 26% successful. After only 7 months, OpenAI took it down.

There are other tools developed by several independent units and researchers to detect AI-generated text, which are used by many teachers. But sometimes they don’t detect text generated by advanced large language models and can produce fake results.

The threat to ChatGPT itself

OpenAI worries that the release of a tool capable of detecting text generated by ChatGPT will be a decision to “shoot itself in the foot”. A survey found that about 1/3 of users would abandon this chatbot if anti-fraud measures were implemented.

Specifically, OpenAI authorized a study in April 2023. The results show worldwide support for a tool that can detect ChatGPT text. On average, 3 out of 4 respondents want to have one, only one does not.

However, another study by OpenAI in the same month found that 69% of ChatGPT users thought that the fraud detection technology would lead to false accusations.

More importantly, 30% of respondents said they would use ChatGPT less if the company implemented a text stamping system that rival AI chatbots do not have.

Since then, the development team at OpenAI has debated the benefits of providing anti-fraud tools to the public.

In early June, OpenAI senior members and researchers discussed the project again. The team agreed that the watermark technology works well, but the results of the ChatGPT user survey from last year are still very noteworthy.

OpenAI is considering other, less controversial approaches. However, things still do not have a clear direction.