AI watermarking could soon provide a way for educators to tell whether students have used AI or not.
To comply with the European Union Artificial Intelligence Act, tech companies will have to watermark content generated through their AI tools. Google is already watermarking text created through its products, and Anthropic wills start doing so with similar techniques. OpenAI is expected to develop watermarking methods for its products as well.
Sarah Fisher, a lecturer in philosophy at Cardiff University in Wales, believes this is all a step in the right direction inside and beyond the classroom, though she acknowledges the new technology won’t be perfect.
“As a society of humans, we place value on what humans produce,” she says.
What Is AI Watermarking?
Historically, watermarking has involved some type of visible marker of authenticity.
“You watermark your banknotes, and you know that they're not counterfeit, or hopefully not,” Fisher says.
AI watermarking works somewhat differently, though the end goal is the same. Claude’s watermarking system will make the chatbot follow specific statistical rules for choosing words with the same meaning. Human readers won’t notice anything, but someone with the watermark key will be able to see if the string of words Claude uses is consistent with the watermark, Fisher explains.
“So, if I'm a student, and I go to my ChatGPT, or my Claude or whatever, and I say, 'Write me a hundred words about X, Y, or Zed.' I wouldn't necessarily see a watermark in what gets produced, but if I then submitted that to an online submission system, then presumably, at some stage, we'll have some software that would be able to tell that it was generated by ChatGPT or Claude or whatever.”
Why Could AI Watermarks Be Good News For Teachers?
Fisher believes being better able to tell what is generated by AI vs. humans will be helpful both for educators and society at large, for a few different reasons.
“For some categories of content, it's really obvious,” she says. “If we're talking about especially audiovisual stuff, these deepfakes are so insidious, we really want to know whether what we're seeing before our eyes is something that really did happen.”
Identifying AI text is also important, but in different ways.
“When we get a piece of text, it's not like looking at a deepfake," Fisher says. "We already have to assess whether what we're being told is true or false. And yet, I think it still can be really, really important to know whether that bit of text came from a human or a machine.”
Fisher adds, “One reason is informational or epistemic. These Chatbots tend to make some signature errors that humans don't make and vice versa. So it's just useful to know what to look out for.”
Another factor is just the value of protecting human work. “I've heard a lot of people saying things such as they just don't want to read a story that's written by an AI," says Fisher. "Even if it was the same words as a human produced, something's missing.”
Discerning human-generated work can be particularly important in the classroom setting.
“As educators, we want to know what our students are producing," Fisher says. "We don’t really care about the end products that much. We care about them going through the process. It's not like we're filling a demand for essays in the world.”
Will AI Watermarks Slow The Tide of AI In The Classroom and Beyond?
Fisher believes the answer to this question is yes, but only to a point.
“I think it might help a bit. I don't think it's going to be the silver bullet that solves this problem, and maybe there just never is a solution to this problem,” she says. “There's always a small minority of students who will find a way to cheat. And that's unfortunate, but we should never expect to get it down to zero. But I think the watermarking might help with some of the most egregious cases, where if a student literally has generated a whole long-form piece of text and taken it straight from Claude, say, it seems like in the future the watermark would be pretty good at telling the educator, ‘Look, this is extremely likely to have been generated by Claude,’ and it's very unlikely that we'd get false positives through that.”
What Are Some Limitations To AI Watermarking
It’s already certain AI watermarks won’t catch all AI-generated content.
“There might be rogue systems which aren't doing the watermarking,” Fisher says. “We know that the big companies are trying to do it, but these tools are proliferating, and so if a student has used one that doesn't have a watermarking feature, then that may not get caught.”
Additionally, existing watermarking techniques are less good at catching short passages of AI-generated text.
“Companies are being quite open about the fact that it might be possible to edit text in a way that takes out the watermarks,” Fisher says.
Though she adds, “I don't know how worried to be about that, actually, because once a student's gone to that much trouble, maybe they've done the work by then.”
Future of AI Watermarking
Some critics of AI watermarking worry that since it won’t be perfect, it will lead society to a false sense of security and could potentially make it more likely to fall for AI-generated content, Fisher says. Another criticism is that past efforts to label traditional misinformation on social media have had mixed results, and some think the same thing will be true about labeling AI.
Fisher doesn’t dismiss these concerns, but believes on the whole, AI watermarks will be a good thing.
“I'm reasonably optimistic that this is a positive step in the right direction,” she says. “It's helpful for us all to hold humans and AI apart for now while we figure out exactly what these things are and how we're going to integrate them into all of our social practices and our institutions.”
