Artificial intelligence firm OpenAI is developing a way to prevent people taking text that AI models produce and passing it off as their own work.
The watermark-like security feature could help teachers and academics spot students who are using text generators such as OpenAI鈥檚 GPT to write essays for them, but cryptography experts say workarounds will inevitably be found.
at the University of Texas at Austin, who is spending a year working with OpenAI, said he couldn鈥檛 comment on exactly how the tool would work, but says the research would be published in the coming months.
Advertisement
AI language models like GPT interpret any basic text instructions they are given as a string of what are known as tokens, which can be letters, partial or whole words and punctuation marks. They then generate an appropriate output as another string of tokens. When creating the output, the AIs use statistical modelling to determine which token should come next to produce the most realistic results.
Aaronson has said in聽a that his watermarking project exploits the small element of randomness involved in the AI鈥檚 choice of tokens. His technique replaces that randomness with a cryptographic tool that leaves a unique signature trace hidden in patterns that are invisible to humans.
OpenAI now has a working prototype that can detect this trademark signature in even a short segment of text, wrote Aaronson. The company could use it to create a website where text can be checked to see if it was created by the firm鈥檚 AI.
鈥淏asically, whenever GPT generates some long text, we want there to be an otherwise unnoticeable secret signal in its choices of words, which you can use to prove later that, yes, this came from GPT. We want it to be much harder to take a GPT output and pass it off as if it came from a human,鈥� wrote Aaronson.
He conceded that there would be ways for students and other users to get around it, such as telling another AI model to paraphrase the output of GPT 鈥� in essence removing the telltale signs implanted by OpenAI. However, 鈥渋f you just insert or delete a few words here and there, or rearrange the order of some sentences, the watermarking signal will still be there鈥�, he wrote.
Aaronson said there are hopes that similar tools could be added to image-generating AI like DALL-E. OpenAI didn鈥檛 respond to a request for comment.
at the University of Birmingham, UK, says detection tools will either be too susceptible to small changes in the text, and therefore miss that work was created by an AI, or too cautious and therefore find false positives.
鈥淚鈥檓 a professor in a university and we do plagiarism detection all the time using various automated tools,鈥� says Ryan. 鈥淭hey鈥檙e all approximate. That doesn鈥檛 mean that they鈥檙e rubbish, but it also doesn鈥檛 mean that you can just go ahead and rely on the output of that tool. You have to treat the output with caution.鈥�
at Edinburgh Napier University, UK, says that inserting detectable signatures into images is relatively straightforward because data can be hidden in tiny fluctuations in the colours of pixels, but that doing the same in text is harder.
鈥淲ith text, you鈥檙e going to be detected quite quickly. Because all you need to do is multiple runs over the text and you can start to see patterns,鈥� he says, comparing it to how cipher crackers deciphered encrypted messages during the second world war.
