Soon its users Internet they read the message Twitteran essay or a news story and they’ll wonder if it was written artificial intelligence software.

To counter this phenomenon, new tools have been created in recent months to recognize texts written by artificial intelligence machines, one of them even developed by the company itself. Open AIprogram creator ChatGPT.

The OpenAI tool uses artificial intelligence to detect differences between human-written text and machine-generated text. When the company tested the new system, it was able to pinpoint AI involvement in only 50% of the texts it analyzed, all of which came from AI software.

Experts say detecting text generated by AI machines is becoming increasingly difficult as software like ChatGPT continues to evolve, producing text that can be convincingly written by the human mind. OpenAI is currently experimenting with technology that will insert special words into ChatGPT texts, making them easier to identify. This technique is called a watermark, just like on banknotes.

The watermarking method ensures that the “keywords” in the AI ​​text are not human-detectable, so tampering is not possible. Even if someone manages to change some special words, the AI ​​text will still be recognized by the watermarking software.

Now OpenAI is experimenting with a new detection technology that will insert special words into ChatGPT texts.

University of Maryland professor and author of a study on the method, Tom Goldstein, says a watermark can be detected even from a “very short piece of text,” like part of a tweet. The OpenAI discovery tool, in contrast, requires at least 1,000 characters (150 words) to decide if a text was written by a human.

However, like any tracking tool, the watermark method isn’t perfect, Goldstein admits. Most experts also warn that it will be a long time before a fully reliable tool is found to detect texts generated by artificial intelligence software.

“The idea that there will soon be some kind of magic wand created by an AI software developer or someone else that will dispel all doubts is fantastic. By the time such a system is invented, we will be long dead,” says David Cox, head of the Watson Lab at the Massachusetts Institute of Technology in Boston.

OpenAI CEO Sam Altman said last month: “Part of the problem with AI texts is that their recognition technology will be a step ahead of new AI language models that can handle more languages, and they even add a literary flavor to the texts they create. “

Some experts believe that OpenAI and other companies in the field need to find a solution to the problem of text authentication before releasing their software to the market. OpenAI released the ChatGPT app in November, and the tracking software became available two months later. By this time, educators and researchers had already realized the need for reliable methods for detecting text written by machines.