If you’re a seasoned writer or editor, you can typically spot when AI text has been given to you, but it is getting more difficult. The better the models are and the more they understand about what makes a good piece of copy, the more complex it tends to be.
It’s not just about an en-dash or even an em-dash, nor is it how many words you pack into a sentence or paragraph. Rather, there’s a flow to writing that AI doesn’t usually understand, though much like the fabled sixth finger in AI creations of people, it is getting better. And many of us are simply getting lazier, leading to more AI writing than ever.
AI can definitely help expedite processes and speed things up, but for some, it can simply bedo the work, whether it’s doing the work by making images, or doing the work of a programmer, or doing the work of a musician trained from real artists, or even doing the work of research and writing.
But what happens when AI gives off a sign that you didn’t, in fact, do what you’re saying you did?
That’s the question mark being posed by Anthropic, which has announced that text from its AI will be watermarked as part of a shift in how it works. It’s a change thanks to the company signing an AI Act in the EU on the practice of transparency in AI generated content, and one that could have a real impact on how people use AI for information.
If you use AI for research, you might simply take the information down as notes and then write your own material, or research from the results and write up something else. But if you use AI to write for you, it could mean that your text is now watermarked, and you would have no idea.
In terms of how Claude will watermark text, that’s a concept that could throw you. After all, text is simply words on a page (or screen), and probably can’t see anything specific in how it’s embedded. But the order of words and how they’re used could be how it works, with only a total rewrite likely getting around the directive, though even that might not work. Images generated from Claude, however, are likely to use metadata, and may show if they’ve been tampered with.
What this means for anyone using AI is that services may be able to work out quickly whether you have, in fact, simply copied text from artificial intelligence platforms in the near future, and may also provide those keys to schools and universities later on down the track.
The cipher and keys to work all of this out isn’t quite known, but Anthropic has said that it will offer a watermark detection system for programmers to connect to, an API that could just raise the curtains on anyone attempting to use AI in this way. In an educational institution, such as high school or university, it could just convince anyone trying this to skip the AI for their work, and go back to researching the old school way: reading the material, writing the notes, and doing it all themselves.
However, it’s a move not everyone is convinced by, and may eventually lead to tools that can strip the watermark from the text, or simply rewrite it as something else.
It’s worth noting that the addition of nearly invisible watermarking to AI text isn’t the only approach that’s been mentioned in the past week. AI music company Suno has also said new songs generated on its AI music platform will also be watermarked with a special system, one that might not be perceptible by the human ear and easily found in audio editing software.
Meanwhile, Gemini uses a variation of watermarking with its DeepMind SynthID technology, even though Google will also allow watermarks to be turned off some media, too.
It does seem as though watermarking is here to stay, and that means if you’re using it for text generation, it can be checked.
Dependent on what you might be using it for — school work, content, or simply handing a report in at work — just be aware that the end result could be one where someone might simply be able to tell you what you did to make your work, and that they may be able to find out how much you did in the end, as well.
