ChatGPT Says It Has A Fix To Catch Plagiarizers Claiming Its Work As Their Own

It’s only the dawn of artificial-intelligent generators, and already, it has caused an overturn in the way people work. Besides the instant gratification that comes with art-creating tools like DALL-E and Lensa, such technology has spurned a revolutionary way for students to cheat.

Almost as soon as ChatGPT, a new text-generating program, became open-source, academics have reported incidents where students were suspected of utilizing ChatGPT and passing off AI-written essays as their own.

ChatGPT, released to the public in December, can process textual questions and requests and churn out wholly “original” works in coherent English within seconds. Those unaware of such tools may not notice the difference and assume that a human author had composed the results themselves.

Granted, the program isn’t perfect and can oftentimes misunderstand the brief. One student who was caught plagiarizing with ChatGPT, for example, had been tasked to write about a philosophical theory. Instead, their work centered around the philosopher this concept was inspired by.

ChatGPT is also bad at giving citations—an understandable flaw, as, just like software by the likes of DALL-E and Midjourney, ChatGPT’s texts are cumulative products of millions of human-made works, making it hard to detect a true source.

ChatGPT’s creators are now working out a way to “fingerprint” watermark texts to avoid “AIgiarism,” a term coined by American venture capitalist Paul Graham, whose wife Jessica Livingston is a backer of ChatGPT creator OpenAI.

During a lecture at the University of Texas, OpenAI guest researcher Scott Aaronson said the organization is pushing safeguards to “statistically watermark” ChatGPT-generated texts. He’s published the full transcript of the lecture on his blog.

ChatGPT’s watermarks will be made up of cryptographic signals that read as natural text to the unobserved eye. In actual fact, the language is distributed in a specific way that’s in line with a code decipherable by anyone with the key to unlock it.

The choice of words is “statistically predictable” and comes as part of a sequence of the chunks of text before. Though, it will be undetectable by the casual reader because it will look as random as all other parts of the text.

The watermark looks completely natural to those reading the text because the choice of words is mimicking the randomness of all the other words.

ChatGPT’s fingerprinting system, currently a “working prototype,” will otherwise be identifiable by a machine as it generates “a few hundred [words]” to give researchers “a reasonable signal that, yes, this text came from GPT,” Aaronson explains.

The technology won’t just be helpful for academic plagiarism detectors like Turnitinm, but it will also help weed out disinformation by malicious parties.

“We want it to be much harder to take a GPT output and pass it off as if it came from a human,” says Aaronson. “This could be helpful for preventing academic plagiarism, obviously, but also, for example, mass generation of propaganda—you know, spamming every blog with seemingly on-topic comments supporting Russia’s invasion of Ukraine without even a building full of trolls in Moscow. Or impersonating someone’s writing style in order to incriminate them.”

[via The Guardian, Bloomberg, Scott Aaronson, Search Engine Journal, cover photo 69519886 © Maxuser2 | Dreamstime.com]