OpenAI is watermarking ChatGPT outputs to fight cheating

ChatGPT has a well-earned reputation among tech observers as a confident bullshitter. It sounds like an undergraduate who skipped every single lecture, yet still manages to write a remarkably convincing exam essay full of plausible nonsense.

While students rushed to use the tool to crank out passing coursework, university lecturers and school administrators started sounding the alarm. Because the generated text is brand new every time, existing plagiarism checkers are completely blind to it.

Now, OpenAI is rushing to watermark its own machine-generated text to stop an impending wave of what investor Paul Graham calls "AIgiarism."

Statistical breadcrumbs in your prose

Speaking at the University of Texas, OpenAI guest researcher Scott Aaronson revealed that the company already has a working prototype of a text watermarking scheme.

The system works by subtly tweaking the model's word choices behind the scenes. To a human reader, the phrasing feels entirely natural and fluid, but the underlying distribution of words leaves a distinct mathematical footprint.

Aaronson noted that just a few hundred words are typically enough to generate a statistically reliable signal that the piece came directly from GPT.

OpenAI wants to use this safeguard for more than just classroom integrity. Aaronson pointed out that it could help curb automated political propaganda campaigns and prevent bad actors from mimicking a real person's writing style to frame them.

The cleanup before the flood

Education is not the only sector feeling the pinch. Developer Q&A giant Stack Overflow recently banned ChatGPT-generated answers entirely.

The site's moderation team explained that while ChatGPT generates answers with a high failure rate, they look convincing enough at a glance to fool people. The sheer volume of cheap, bad code answers threatened to drown out genuine human help.

Paul Graham argued that the rules against AIgiarism should mimic standard anti-plagiarism guidelines, since taking credit for words you did not actually write is still deception.

Why it matters

OpenAI is trying to get ahead of the regulatory sledgehammer before schools and governments decide to ban its tools outright.

For everyday users and students, the message is clear: using ChatGPT to ghostwrite your papers is already living on borrowed time. Once these watermarking detectors roll out, backwards-looking detection could easily catch up with unedited submissions.

For the broader industry, OpenAI is effectively setting the standard for responsible AI deployment. If watermarking works reliably, expect schools and corporate platforms to demand the same traceability from every generative text engine on the market.

We think the concept is clever, but it has obvious blind spots. Statistical watermarks are notoriously susceptible to tampering. A user can easily disrupt the word probability pattern by running the output through a paraphraser, swapping synonyms, or simply rewriting a few sentences by hand. It marks the start of an endless cat-and-mouse game between model creators and determined cheaters.