A confident bullshitter that can write very convincing nonsense’: not a takedown of an annoying student or a former British prime minister, but a description of an artificial intelligence writing program that is causing headaches for its makers.
With fears in academia growing about a new AI chatbot that can write convincing essays – even if some facts it uses aren’t strictly true – the Silicon Valley firm behind a chatbot released last month is racing to “fingerprint” its output to head off a wave of “AIgiarism” – or AI-assisted plagiarism.
ChatGPT, an AI-based text generator that was released for public use in early December, has been praised and criticized alike for the quality of its output. Users can ask it questions ranging from simple factual queries (“What is the tallest mountain in Britain?”) to absurd requests (“Write a limerick explaining the offside rule”) and receive clear and coherent responses written in natural English.
Headteachers and university lecturers have expressed concerns that ChatGPT, which can provide convincing human-sounding answers to exam questions, could spark a wave of cheating in homework and exam coursework.
Now, the bot’s makers, San Francisco-based OpenAI, are trying to counter the risk by “watermarking” the bot’s output and making plagiarism easier to spot.
In a lecture at the University of Texas, OpenAI guest researcher Scott Aaronson said that the company was working on a system for countering cheating by “statistically watermarking the outputs”. The technology would work by subtly tweaking the specific choice of words selected by ChatGPT, Aaronson said, in a way that wouldn’t be noticeable to a reader, but would be statistically predictable to anyone looking for signs of machine-generated text.
“We want it to be much harder to take a GPT output and pass it off as if it came from a human,” Aaronson said. “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.
“We actually have a working prototype of the watermarking scheme,” Aaronson added. “It seems to work pretty well – empirically, a few hundred [words]seem to be enough to get a reasonable signal that, yes, this text came from GPT.” ♦
