Systems to detect when AI has been used to write an assignment, for example, are already being used at schools and universities.But there's a change underway at some of the biggest artificial intelligence firms that could make detectors much harder to fool.Watermarks are coming to text generated by AI chatbots. But you can't see it and you probably won't be able to tell if it's been applied.How does it work? Program:More from ABC News Top StoriesTranscriptDavid Coady: Have you ever read an article, an email or a post on social media and thought, I reckon this is just AI? It could soon be clearer which words are the product of a human's mind and which are the output of an artificial intelligence tool. Anthropic, which is behind Claude, one of the most popular AI chatbots, has revealed it will start weaving a secret watermark into all of the text its models generate. Why is it doing this? It's because of the Europeans. Article 50 of the EU AI Act came into effect at the start of August. It requires AI systems to mark their output so that a machine can detect it as artificially generated. So Anthropic is moving first and the feature won't be limited to Europe. It's different to how some AI tools watermark images or documents they create. There, some metadata can be put into the actual file. But with text that can be copied and pasted, it's the words themselves that are being tinkered with. There are no hidden characters. It's a pretty amazing system to do with patterns, synonyms and how AI picks the next word it spits out. To explain, here's AI expert Shaanan Cohney, an assistant professor at Tufts University in Massachusetts.Shaanan Cohney, Assistant Professor of Computer Science at Tufts University: So when you put in a question into one of these models, it uses some randomness along with what it knows to try and figure out what words to put out next. And what you can do is you can put a subtle bias in the randomness that has to do with a secret code. So for example, the sky is overcast today. The sky is grey today. Both of those seem quite similar. But if you know the secret code that's been locked up, you can guess which of those is going to be more probable, given that it comes from Claude and not from somewhere else.David Coady: But don't we want chatbots to use the best word in all circumstances, not change up words so that they're detectable as AI?Shaanan Cohney, Assistant Professor of Computer Science at Tufts University: This is actually super interesting. The authors of the research paper on whom this technology is based, they took 20 million responses from Google's chatbot, and they actually found that there was no meaningful difference in the ratings that people gave to them. Practically day to day, it seems like the sort of thing that most people are unlikely to notice.David Coady: The watermark won't reveal anything about you or what you asked, Claude. It'll just reveal, once Anthropic releases the detection tool, that this text came from Claude. It won't work on very short bits of text and won't apply to proofreading if the words you put into the chatbot are not substantially changed. There are AI detectors out there already, but Shaanan Cohney expects this method will be better. So, for anyone using AI to generate huge chunks of text for assignments, job applications or reports at work, what now? How do you hide it?Shaanan Cohney, Assistant Professor of Computer Science at Tufts University: So that presupposes that we should be hiding it. For example, with cover letters, the more people use AI, the less valuable they are for everyone else. Because if everyone is just sending in slop, no one's going to read it at all and it won't be useful for the hiring process. So I think society's got a lot of hard questions to ask itself and the technologies that we need to develop in order to protect ourselves and to build the kind of society we want to see.David Coady: At first, the feature will be part of new Claude models launched by Anthropic. Other major AI companies have pledged to comply with the EU law, so it's becoming clear that soon it'll become hard to hide if you've got a big helping hand from AI.