Call me naive all you want, but I don't think there is anything particularly dystopian about artificial intelligence itself, as long as the companies behind our frontier models are conscious about what parts of our humanity must be conserved and are making efforts to do so. Most of us have forgotten how to write, or at least how to write confidently and, more importantly, trust our own writing. Almost everywhere except in my academics and at The Dartmouth, I too have almost always relied on my trusty dear pal Claude to check my writing. “Hello, Night Owl,” it will say, as I, in the worst possible grammar and incomprehensible spelling, tell it to check if the email I am sending to my boss at 2:00 AM on a Wednesday is polite enough. Sometimes, when I am feeling daring and lethargic enough to surrender to the machines, I just ask it to write the whole email from scratch. I know that you too, dear reader, have probably, at least once, done the same. The saddest statistic of all is that now there is more suspected AI generated content on the internet than original human copy.
Anthropic, last week, announced that they intend to place a watermark on every text that is produced from Claude, and that this would survive copy-pasting. The mechanics, as far as the company has explained them, go something like this: Every new Claude model released after August 2nd embeds an invisible, machine-readable signal directly into the words it generates.
This was not done as a courtesy but rather because the European Union’s Artificial Intelligence Act now requires it of any company operating there, and Anthropic has decided to just roll the feature out everywhere rather than maintain two versions of the same product. The mark rides along inside the text itself, so it survives a copy-paste into your Notes app, your professor’s inbox and even your Canvas submission portal, for the extremely daring ones reading this. Anthropic says it may even survive some light editing.
I want to believe this saves writing. I’ve spent the last several paragraphs of a draft I deleted trying to convince myself that it does. It doesn’t. At least not on its own.
A watermark that survives light editing does not necessarily survive heavy editing, and the people running content farms and posting three “thought leadership” essays a day are likely not light editors. They are professionals at exactly the task that counteracts the intended benefits of the watermark: running text through a second pass, a second model, a translation out and back. Eventually, the signal degrades below detection. Anthropic’s own documentation concedes as much. The watermark, in practice, catches the final paper submitted whole and the press release nobody touched. It does not catch the operation built specifically to avoid being caught, because that operation was already optimizing against detection before watermarks existed. The people this feature needs to stop are the ones best equipped to route around it.
Look at what has already happened to books, because books are the closest thing we have to a preview of what a watermark can and cannot fix. Amazon’s Kindle Direct Publishing store has spent the last few years absorbing a genuine flood of AI-produced titles — thousands of them a month by some counts, clustered in genres cheap to fake at volume: romance, keto meal plans, day-trading guides and knockoff self-help. Amazon responded years ago with a cap on how many books a single account can upload per day and a disclosure requirement for AI-generated content.
Neither policy solved the problem, because the cap does nothing to stop someone running many accounts, and the disclosure field is buried in the upload flow where a shopper never sees it. An author’s name has already been used on books she never wrote, forcing Amazon to pull them after the fact rather than catch them before publication. A watermark on Claude’s output would help identify some fraction of this flood at the source, but the KDP experience is the lesson tha
Call me naive all you want, but I don't think there is anything particularly dystopian about artificial intelligence itself, as long as the companies behind our frontier models are conscious about what parts of our humanity must be conserved and are making efforts to do so. Most of us have forgotten how to write, or at least how to write confidently and, more importantly, trust our own writing. Almost everywhere except in my academics and at The Dartmouth, I too have almost always relied on my trusty dear pal Claude to check my writing. “Hello, Night Owl,” it will say, as I, in the worst possible grammar and incomprehensible spelling, tell it to check if the email I am sending to my boss at 2:00 AM on a Wednesday is polite enough. Sometimes, when I am feeling daring and lethargic enough to surrender to the machines, I just ask it to write the whole email from scratch. I know that you too, dear reader, have probably, at least once, done the same. The saddest statistic of all is that now there is more suspected AI generated content on the internet than original human copy.
Anthropic, last week, announced that they intend to place a watermark on every text that is produced from Claude, and that this would survive copy-pasting. The mechanics, as far as the company has explained them, go something like this: Every new Claude model released after August 2nd embeds an invisible, machine-readable signal directly into the words it generates.
This was not done as a courtesy but rather because the European Union’s Artificial Intelligence Act now requires it of any company operating there, and Anthropic has decided to just roll the feature out everywhere rather than maintain two versions of the same product. The mark rides along inside the text itself, so it survives a copy-paste into your Notes app, your professor’s inbox and even your Canvas submission portal, for the extremely daring ones reading this. Anthropic says it may even survive some light editing.
I want to believe this saves writing. I’ve spent the last several paragraphs of a draft I deleted trying to convince myself that it does. It doesn’t. At least not on its own.
A watermark that survives light editing does not necessarily survive heavy editing, and the people running content farms and posting three “thought leadership” essays a day are likely not light editors. They are professionals at exactly the task that counteracts the intended benefits of the watermark: running text through a second pass, a second model, a translation out and back. Eventually, the signal degrades below detection. Anthropic’s own documentation concedes as much. The watermark, in practice, catches the final paper submitted whole and the press release nobody touched. It does not catch the operation built specifically to avoid being caught, because that operation was already optimizing against detection before watermarks existed. The people this feature needs to stop are the ones best equipped to route around it.
Look at what has already happened to books, because books are the closest thing we have to a preview of what a watermark can and cannot fix. Amazon’s Kindle Direct Publishing store has spent the last few years absorbing a genuine flood of AI-produced titles — thousands of them a month by some counts, clustered in genres cheap to fake at volume: romance, keto meal plans, day-trading guides and knockoff self-help. Amazon responded years ago with a cap on how many books a single account can upload per day and a disclosure requirement for AI-generated content.
Neither policy solved the problem, because the cap does nothing to stop someone running many accounts, and the disclosure field is buried in the upload flow where a shopper never sees it. An author’s name has already been used on books she never wrote, forcing Amazon to pull them after the fact rather than catch them before publication. A watermark on Claude’s output would help identify some fraction of this flood at the source, but the KDP experience is the lesson that identifying the problem and a marketplace actually acting on that identification are two entirely different achievements, and so far only the first one has happened at any scale in publishing.
Suppose, generously, that some fraction of AI text still gets marked and identified. The next question is who does anything about it, and the honest answer right now is close to nobody. Detection tools for such watermarks are not yet publicly available. Even once they are, nothing obligates a platform to check for the watermark, and nothing in an ad-supported feed currently rewards authenticity over engagement. A LinkedIn post gets paid in impressions regardless of who or what wrote it. A search result ranks on relevance signals that have nothing to do with a watermark. Being labeled AI generated is not currently a cost to anyone, because no institution has built a system that treats it as one. A mark that carries no consequence is not a deterrent. It is trivia.
The watermark is not the rescue. It is one piece of infrastructure that only becomes meaningful once two things happen around it. Platforms have to build detection into the systems that decide what gets seen, the way spam filters treat a flagged sender differently rather than just knowing the message is spam. Publications and employers have to start treating an AI label the way they already treat plagiarism: as a reason to reject or investigate rather than a fact to note and ignore. Neither of these are technical problems and a company embedding a signal into its own output cannot make those decisions for the rest of the internet.
What the watermark actually did was remove the excuse. Before last week, nobody could point to AI text with certainty, so nobody had to decide what to do about it. Now, eventually, once detection ships and someone bothers to look, that excuse is gone. What happens next depends entirely on whether platforms, publications and regulators treat that as a reason to act or as a fact to shrug at.
Writing does not get saved by a fingerprint sitting quietly inside a string of tokens. It gets saved if enough institutions decide the fingerprint is worth using, and that decision has not yet been made. The watermark just made it impossible to keep avoiding.t identifying the problem and a marketplace actually acting on that identification are two entirely different achievements, and so far only the first one has happened at any scale in publishing.
Suppose, generously, that some fraction of AI text still gets marked and identified. The next question is who does anything about it, and the honest answer right now is close to nobody. Detection tools for such watermarks are not yet publicly available. Even once they are, nothing obligates a platform to check for the watermark, and nothing in an ad-supported feed currently rewards authenticity over engagement. A LinkedIn post gets paid in impressions regardless of who or what wrote it. A search result ranks on relevance signals that have nothing to do with a watermark. Being labeled AI generated is not currently a cost to anyone, because no institution has built a system that treats it as one. A mark that carries no consequence is not a deterrent. It is trivia.
The watermark is not the rescue. It is one piece of infrastructure that only becomes meaningful once two things happen around it. Platforms have to build detection into the systems that decide what gets seen, the way spam filters treat a flagged sender differently rather than just knowing the message is spam. Publications and employers have to start treating an AI label the way they already treat plagiarism: as a reason to reject or investigate rather than a fact to note and ignore. Neither of these are technical problems and a company embedding a signal into its own output cannot make those decisions for the rest of the internet.
What the watermark actually did was remove the excuse. Before last week, nobody could point to AI text with certainty, so nobody had to decide what to do about it. Now, eventually, once detection ships and someone bothers to look, that excuse is gone. What happens next depends entirely on whether platforms, publications and regulators treat that as a reason to act or as a fact to shrug at.
Writing does not get saved by a fingerprint sitting quietly inside a string of tokens. It gets saved if enough institutions decide the fingerprint is worth using, and that decision has not yet been made. The watermark just made it impossible to keep avoiding.
Opinion articles represent the views of their author(s), which are not necessarily those of The Dartmouth.
