Can a Hidden Watermark Help Detect Text Written By AI? Maybe

But everyone must be willing to use it

  • A watermarking tool can identify text written by AI with near certainty.
  • It would require significant effort to circumvent. 
  • All AI-text tools would be required to add it to their text-generating models.
Closeup on a page in a ribbon-based typewriter with text that reads "This is my story..."

Constantine Johnny / Getty Images

An old technique might help combat the biggest problem of AI-created text—fraud.

Text written by AI (artificial intelligence) can be useful, but it is also ripe for exploitation by bad actors. Fake, falsely-attributed political texts, and cheating in school, for example. And detecting computer-written text is hard. Even OpenAI, the company that built Dall-E and ChatGPT, can only identify AI-written texts 26% of the time, which is where watermarking comes in

"In my opinion, the biggest danger of AI-generated text is the potential for it to be misused or exploited by unscrupulous actors. Watermarking can help to reduce the risk of this happening, as it can make it easier to trace the source of any text and take appropriate action if necessary," Ataur Rahman, CEO of AI assistant company GetGenie, told Lifewire via email.

High Watermark

Unlike image watermarks, which are designed to be obvious and to shame anyone who tries to use the image without permission, AI text watermarks use hidden patterns encoded into the text itself. These are not noticed by human readers but can be detected by software. 

It works using an algorithm to change the choices made by the AI. Bots like ChatGPT work by creating texts one word at a time. The watermarker maintains two-word lists, a greenlist and a redlist. Whenever the chatbot chooses its next word, it will be more likely to pick one from the greenlist. For instance, if the next word is a kind of dog, it would choose a greenlisted breed instead of a redlisted breed.

Colorful plastic letters agains a pink and blue background.

Towfiqu barbhuiya / Unsplash

To detect whether a text was AI-generated, you just run it through software that counts the number of green vs. red words. The technique is more effective on longer texts and could, in theory, be built into a web browser extension to automatically highlight AI-generated text. 

"While I like the idea of watermarking AI-produced text, I believe that it could be hard to enforce. If this concept were to truly be effective, the watermark must work across everything from websites to word documents," Ashton Frye, founder of AI text tool company TypeDroid, told Lifewire via email.

Is This Text Written By AI?

The tool, according to the creators, is extremely effective, but that doesn't make it infallible. If you have access to the word lists, you could change them programmatically, for example. And if you want to tweak it manually to avoid detection, as one might do if using AI to cheat on college assignments, then you can. Although that would probably require as much work as just writing the essay in the first place. Speaking to MIT Technology Review's Melissa Heikkilä, Tom Goldstein, one of the co-researchers involved with the project, said that you would have to change around half the words in order to remove the watermark.

The real challenge is that, to be effective, all AI text generators would have to implement this tool. And even if all of them agreed to add it to their models, it would always be possible to create another AI text generator that did not comply.

This is especially important in terms of copyright infringement, where users might want to avoid detection.

Someone capturing the text in a book by taking a picture with a smartphone.

patpichaya / Getty Images

"While the written and visual content is considered 'generative,' the reality is it is entirely 'derivative' of existing material. And when it comes to AI-created text, the most valuable material is data," TJ Leonard, CEO at stock image, audio, and text company Storyblocks told Lifewire via email. 

The AI revolution is already underway, and it seems that it may already to too late to reign in its worst aspects. Future implementations may be better regulated, but there are already open-source image-generation models out in the wild, free (and easy) to use, and already incorporated into phone apps.

And for the user, things probably won't get any less confusing. We won't be able to tell if it's text written by AI or just blandly written. In the meantime, we will start to see AI text tools in the software we already use, offering to write reports and automate other tedious work tasks.

Like any tech, it's a tool that can be used for good or evil. But so far, in the case of AI, that balance is still heavily skewed towards evil.

Was this page helpful?