Why AI text watermarks — like Claude’s — are a great tool against slop

Anthropic, the company that created the AI tool Claude, has announced it will include invisible watermarks in Claude-generated content, including text, graphics, and code. As a writer and editor, I’ll concentrate on what this means about AI-generated text content.

These are some key features of Claude’s watermarking:

• Watermarking can’t be turned off. It’s not optional. It will be included in all content generated by Claude’s new models as they are released, and may be retrofitted to old models as well. Content created before this policy was implemented won’t include a watermark.
• Anthropic is complying with EU regulations requiring that AI output be easily identifiable. As a result, you can expect all other major AI suppliers to deliver a similar feature soon. If, for example, you switched from Claude to ChatGPT to avoid watermarking, you’ll find soon watermarking present in your new tool.
• It persists through cutting and pasting text. It’s likely that this means that the watermark is based on word choices, rather than, say, using different but similar-looking characters for spaces or punctuation. This makes it technically difficult to strip out.
• It’s invisible to readers. In any given sentence, there are a number of possible word choices that still make sense for a given Large Language Model. The word choices that encode the watermarking won’t affect the readability of the text. (Or as Claude might write, “It’s not only a security feature, it’s unobtrusive, invisible, and benign.”)
• AI-detection tools will soon have access to it. Anthropic will create a simple system that other tools — like the AI detection product Pangram — can use to detect AI-generated content.

Why this is great

Tools like Pangram are making the best possible guesses about whether a passage is AI-generated. They make mistakes. (I’ve experienced this myself, as Pangram marked text I’d worked on as AI, even after it was heavily edited.)

Anthropic’s watermarking, while not perfect, is far better. While it won’t be able to catch a snippet of a few sentences or two, I expect it to detect AI dependably in larger passages. When the other models follow suit, AI detection should be on far more solid ground.

One of the best features of this AI watermarking is that it only marks literal text output. If you use AI for research, for brainstorming, as a thesaurus, to recommend the best title for your article, or to analyze your writing tics, watermarking won’t flag your AI use. But if you use AI to generate text and include it in an article or book, the watermarking will make it far easier to definitively detect.

Right now AI detection is equivocal and subject to argument. This makes life difficult for editorial professionals who are dealing with potentially AI-generated source material, or who are accused of using AI to write things instead of their own brains. If AI detection is dependable, it will eliminate a lot of those problems.

I expect these potential applications of dependable AI detection to emerge:

• Teachers will be able to require students to write assignments themselves, and can flag AI-generated homework responses with so much worry about making false accusations. This will get the writing curriculum back on track and help better develop students’ writing skills.
• Publishers will be able to definitively test and reject AI-generated author content. They won’t have to insist that authors document how they wrote things themselves, which creates a tedious and time-consuming amount of overhead.
• Writers for hire will be able to prove they did the work they’re being paid for.
• The Copyright Office will be able to test work submitted for potential copyright registration and reject it if it is mostly AI-generated. This is consistent with the Copyright Office’s policy that AI-generated content is not eligible for copyright.
• Online bookstores like Amazon will be able to detect and reject AI slop books, improving the quality of their content selection and the experience for buyers. Far fewer people will mistakenly buy crappy AI-generated knockoffs.
• Workers will be able to tell if their colleagues’ work is original writing or AI-generated.
• I expect a system to emerge soon in which content can be dependably certified as human-authored. (VerifyMyWriting offers one such certification — it will likely become much more precise once watermarking is universal.) Legitimate news and content sites will likely post badges certifying their content, reassuring readers that it’s worth their time to read it.

Who’s against this?

If you object to this watermarking, you are effectively saying, “I think people should be able to create AI-generated content and share it without identifying it.” And no one is fooled. You really mean that you should be able to do this.

No one is stopping you from using AI for whatever task you want — ghostwriting text for hire, writing memos, completing assignments, or even authoring a book. But you won’t be able to conceal how you did it. If you’re embarrassed to have people know the tools you use, why? A real writer never needs to hide their writing process.

Because the AI watermarking is based on word choice, the only way to undo it is to rewrite the text with different word choices. And Anthropic won’t make it easy to identify which words incorporate the watermark, so you’ll have to rewrite everything. If you’ve gone to that trouble, you haven’t saved time; you’d probably be better off just writing the text the usual unautomated way and asking an editor to improve it.

If you use AI to generate masses of content for, for example, promoting a product to people, you either should be willing to admit your use of AI, or you should stop and get a real copywriter.

Anthropic is not violating your privacy. It never promised you that your use of its tools would be kept secret.

No legitimate writer should have a problem with this.

I’m sure many of you think I’m wrong. I look forward to reading your well-argued, logical justification in the comments.