OpenAI Working on Proprietary Watermark for Its AI-Generated Text
January 24, 2023
Even before OpenAI made its text generator GPT-3 available to the public, folks were concerned the tool was too good at mimicking the human-written word. For example, what is to keep students from handing their assignments off to an algorithm? (Nothing, as it turns out.) How would one know? Now OpenAI has come up with a solution—of sorts. Analytics India Magazine reports, “Generated by Human or AI: OpenAI to Watermark its Content.” Writer Pritam Bordoloi describes how the watermark would work:
“We want it to be much harder to take a GPT output and pass it off as if it came from a human,’ [OpenAI’s Scott Aaronson] revealed while presenting a lecture at the University of Texas at Austin. ‘For GPT, every input and output is a string of tokens, which could be words but also punctuation marks, parts of words, or more—there are about 100,000 tokens in total. At its core, GPT is constantly generating a probability distribution over the next token to generate, conditional on the string of previous tokens,’ he said in a blog post documenting his lecture. So, whenever an AI is generating text, the tool that Aaronson is working on would embed an ‘unnoticeable secret signal’ which would indicate the origin of the text. ‘We actually have a working prototype of the watermarking scheme, built by OpenAI engineer Hendrik Kirchner.’ While you and I might still be scratching our heads about whether the content is written by an AI or a human, OpenAI—who will have access to a cryptographic key—would be able to uncover a watermark, Aaronson revealed.”
Great! OpenAI will be able to tell the difference. But … how does that help the rest of us? If the company just gifted the watermarking key to the public, bad actors would find a way around it. Besides, as Bordoloi notes, that would also nix OpenAI’s chance to make a profit off it. Maybe it will sell it as a service to certain qualified users? That would be an impressive example of creating a problem and selling the solution—a classic business model. Was this part of the firm’s plan all along? Plus, the killer question, “Will it work?”
Cynthia Murrell, January 24, 2023