On August 2, Anthropic started weaving an invisible watermark into everything Claude writes. Detectors can already flag a big share of what gets posted online, with one scan of over a million posts finding more than 40 percent of LinkedIn longform reading as fully machine-generated. Schools and professors run detectors on essays. Some hiring managers run them on cover letters and resumes.
Into that, there is an obvious business sitting out there ready for the taking, and whoever builds it properly makes an absurd amount of money. A tool that writes in any specific human voice you point it at, that no detector can flag, and that carries no watermark.
The tools that exist do half the job… badly
There is already a whole industry of what people call humanizers, tools that take AI text and rework it to slip past a detector. The problem is that they don't work all that well. Some improve the writing and still get caught. Some beat one detector and fail against the next. Some produce something that scores as human and hurts your brain reading it.
And all of them are solving the smaller half of the problem. Getting past a detector is about erasing a signature, smoothing out the statistical tells that mark text as machine-made. That is just a blurring job and it makes text read like a generic human.
The hard half, the valuable half, is the opposite of blurring. It is writing like one specific person you can point at and you yourself are amazed by it. Something that feels real and authentic but it also has to be really easy to use.
Why watermarks make the prize bigger, not smaller
You would think the AI labs adding watermarks makes evasion harder and the opportunity smaller. I think it does the reverse.
A statistical detector is guessing. It looks at text and estimates a probability, and it is wrong often enough that real people get accused of being machines for writing too cleanly. That fuzziness is the humanizer's friend, because you only have to nudge the probability.
A watermark is not guessing. It is a deliberate mark placed in the text by the people who made the model, designed to survive copy-paste and some editing. It is a much stronger claim, and it turns evasion from a fuzzy statistical game into a concrete one: is the mark present or not? That sounds worse for the evader, and in the short term it is. But it also means that if someone builds a tool that reliably strips the watermark while keeping the meaning, then you have something people want. It defeats the strongest signal the industry has, cleanly, every time.
Combine the two, perfect voice in however form you want and no detectable trace, and you have not built just a writing aid, but you have built a cash printing machine that produces things that sound real and original.
The catch
The value of undetectable text depends on detection existing. The student pays to beat the detector because the school runs one. The applicant pays because the recruiter checks. The whole market is priced on the assumption that somewhere, someone is still trying to tell human from machine and mostly succeeding.
A tool that actually works, at scale, ends that. Once anyone can generate flawless text in any voice with no trace, detection stops meaning anything, and everyone knows it stops meaning anything. Schools drop the detectors because they are useless. Recruiters stop reading cover letters. The signal the product was built to defeat gets abandoned, and the reason to buy the product goes with it. You would be selling the last bottles of a poison that kills its own market.
And what it takes down on the way out is bigger than essays. Text is the last medium most people still trust by default. Seeing all the AI slop on our feeds, we already assume the photo might be fake and the voice might be cloned. Well, actually… I take that back because I see a lot of comments on social media that look like they were run in an LLM as well.
Somebody builds it and makes bank
The demand is real and enormous, the technical pieces mostly exist, and the reward for the first tool that truly nails both halves is large enough that someone, somewhere, is working on it right now. But it needs to be hard to duplicate, because if it is easily duplicated then it becomes a commodity.
Regardless, whoever connects them first and does it really well cashes out, and the rest of us lose the ability to believe a paragraph came from a human. That trade is coming whether the person who profits from it thinks it through or not.
Sources
TechCrunch: Anthropic says it will watermark text generated by its AI models - The August 2, 2026 start of embedded text watermarking on Claude output and the scope of the rollout.
Anthropic: How Claude marks AI-generated content - The embedded text watermark that persists through copy-paste and some editing, and the note that a detected mark signals processing rather than authorship.
