Photo Courtesy of Olivia Gehm.
Coupled with preexisting AI detection software, watermarking has the potential to help institutions crack down on generative AI use.
By Olivia Gehm
TAMPA, Fla. — As generative artificial intelligence continues to rapidly expand, legislators around the world are drafting laws to regulate it. The European Union (EU), for example, passed regulatory legislation surrounding generative AI, effective Aug. 2.
According to the EU Artificial Intelligence website, generative AI is required to “mark outputs in a machine-readable format and ensure they are detectable as artificially generated or manipulated.”
In other words, AI outputs will have to include an indication that the text is machine-generated, detectable by software. The watermark is invisible to the user, and doesn’t change the quality or style of the output — but it’s there.
Claude, a popular AI model by Anthropic, has been among the first to comply, and the company has posted an announcement explaining the watermark.
“Large language models like Claude work by generating one word at a time,” Anthropic said. “Each time the model decides on the next word, it chooses among a list of potential candidates, ultimately selecting the most sensible or likely based on the preceding text.”
Anthropic then explained that the function of a watermark is to leave a pattern in Claude’s responses, which will be detectable to anyone with the “key.”
“The words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key,” said Anthropic. “If it is, one can assign a probability that the text was generated by Claude.”
While AI models such as Claude will be watermarking outputs for users regardless of location, it is unclear whether Anthropic will provide the “key” to anyone other than those mandated by EU law. However, if access is expanded to, say, educational institutions in the U.S., professors could have an additional tool to detect plagiarism.
While AI model watermarking is relatively new, AI detection software has been around for a few years. Pangram has received mixed reviews, but the software is fairly accurate.
Steven Mollman, a professor in the department of English and Writing at The University of Tampa, said that many studies have given Pangram a very low false-positive rate, meaning rarely does the software flag human-written text as being AI-generated.
“I think the flip side of that, though, is the false-negative rate,” he said. “Like, how often is something that’s AI getting through?”
Mollman also said that he believes many detection software programs, namely Turnitin, focus heavily on lowering their false-positive rate, often at the expense of their false-negative rate.
“They don’t want you filling out an academic integrity case against someone who didn’t do anything wrong,” he said.
Although false-positive results are rare, many professors are still wary of using test results as anything other than a confirmation of plagiarism that they already suspected.
“I would never, for example, solely prosecute an academic integrity case based on a Pangram result,” Mollman said.
That’s where watermarking comes in. While its novelty makes it unclear how helpful it will be to higher education on its own, it could be important when used in tandem with current detection software.
If higher education can properly combine the use of watermarking and third-party detection software, students might be deterred from submitting AI-generated work. At the very least, institutions can ensure that less AI-generated material will slip through the cracks.
While there aren’t yet any obvious tangible downsides to watermarking, some have expressed concern that AI models such as Claude will have worse outputs once watermarking becomes standard.
In the same announcement, Anthropic denied any claims that Claude’s outputs will suffer from watermarking, and said that the technology is not detectable by the user.
“Watermarking does not impact the quality of Claude’s output,” the company said. “To a reader, a watermarked response is indistinguishable from an unwatermarked one.”
Regardless of the efficacy and quality of watermarked AI models, they are of no use to higher education if institutions do not have access to the key. Anthropic said they plan on expanding access to that key over time, but in the event that schools in the U.S. are unable to use the technology, there’s always the alternative of going completely analog.
Some professors have already made this switch to tech-free classrooms, such as Victorio Reyes Asili, professor of English and Writing at The University of Tampa. Reyes said that while he was reluctant to go analog for his academic writing classes, he has noticed that students seem much more engaged and invested in learning.
“It has now become clear that paper, pens, and books are more effective at cultivating reading literacy and promoting students’ cognitive development than digital learning tools,” he said.
While it is unclear what academic integrity will look like as AI continues to improve, schools have several tools at their disposal, such as analog classes, AI detection software, and possibly watermarking in the not-too-distant future.