{"slug": "you-gotta-fight-for-your-write", "title": "You Gotta Fight For Your Write", "summary": "Anthropic has announced a new text watermarking system for its Claude AI model that alters outputs in a way the company says allows it to assess the probability any block of text was generated by Claude, a technique similar to Google's SynthID but drawing criticism for its potential invasiveness. The system, detailed in an August 11, 2026 announcement, has sparked debate among users and commentators about privacy and the ethics of AI-generated content detection.", "body_md": "# You Gotta Fight For Your Write\n\nSomeone reset the 'Days Without Incident' clock at Anthropic to 0. From, I don't know, maybe 1, at most 2, [these days](https://spyglass.org/anthropic-becomes-the-villain/).\n\nLook, when I first saw [the news](https://www.theregister.com/ai-and-ml/2026/08/11/anthropic-pledges-to-embed-watermarks-to-help-discern-ai-slop-in-sop-to-eu/5285792?ref=spyglass.org) that the company would be \"watermarking\" text, I didn't think much of it. That's in part because [they didn't really explain](https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content?ref=spyglass.org) how they'd be doing it so I assumed, [like probably most people](https://daringfireball.net/linked/2026/08/11/anthropic-claude-watermarks?ref=spyglass.org), that they'd be embedding some invisible characters in text outputs. But that, of course, would only work for people who directly copy-and-paste text from Claude to elsewhere. If you re-wrote whatever the output was, or used some tool to \"wash\" the text – AI laundering? – you'd remove such \"watermarks\". Still, [many were up-in-arms](https://www.techmeme.com/260811/p2?ref=spyglass.org#a260811p2) about this because it was some level of invasive or sneaky. Others because it was going to gunk up the internet even more with superfluous characters. And some simply because it was going to make using AI to write without being known *harder to pull off*!\n\nAs it turns out, [what they're actually doing](https://www.anthropic.com/news/claude-text-watermark?ref=spyglass.org) – or planning to do – to pull this \"watermarking\" off is instead fairly insane.\n\nBasically, Anthropic has designed a system that will alter the output in a way that allows them to assess the probability that any block of text has been generated by Claude. The key is switching one possible output to another at various points in a way that they suggest would only be perceptible by them, holding the keys to how they're doing this.\n\nWhile it's true that Google has also been leveraging this same basic idea in the form of their \"[SynthID](https://deepmind.google/models/synthid/?ref=spyglass.org)\" technology for a couple of years, I suspect it hasn't been met with such outrage simply because most people don't use Gemini for such purposes – at least not to the extent they use Claude (or ChatGPT).1\n\nRegardless, Anthropic has now pulled this to the forefront and made it a debate.\n\nAgain, at first glance this may not sound so bad. Perhaps even somewhat clever. I'm reminded of how back in my reporting days, [I used to insert](https://www.yahoo.com/news/tech-writer-successfully-inceptions-tech-blogosphere-060405254.html?guccounter=1&guce_referrer=aHR0cHM6Ly93d3cuZ29vZ2xlLmNvbS8&guce_referrer_sig=AQAAAIfiYbdlMIrKBSwioMYdiCO7IPPxuytToiu7XYaeEHd53OQ2eX84W-zAIiZgVsX_4rEqKt7v2qHgjePQ8uM4KCsKTlU6K_ggbB7_afkKLKdScPgNisJU8CKNnFVvm4b4aQe5aEezenSs9e1iYWIDhdIMgD3obVrv-LvaFcadw0Un&ref=spyglass.org) some subtle tweaks to catch if someone was trying to copy a scoop without properly attributing the information source (me).2 You'd be surprised how often this worked!3\n\nTo that end, I honestly don't think I'd be as upset/annoyed if Anthropic was doing this simply as a way to catch (or, if you prefer \"identify\") people who use AI to, say, write *entire* thought pieces. Just write the damn thing you lazy sonofabitch. Or don't, but don't put your name on it like you did!\n\nOther tools and [services](https://www.pangram.com/?ref=spyglass.org) have been popping up to do this, obviously. I'm not really sure we *need* to police such things, but it is also amusing to know the type of people who do this. Then again, it also opens a can of worms in terms of if such output should be considered as a \"valid\" form of expression if it's driven by your own input (your prompt), etc.\n\nThe key, I think, is that *many people just want to know*. They want to know if you used AI to generate what they're reading. And in some situations, like legal proceedings, people may *need* to know. To make sure they can trust that the actual legwork was done to say, prosecute a case, versus just the prompt work. Which is why the EU passed the law that Anthropic is now trying to adhere to with this! (More on this in a minute.) Should anyone be *forced* to disclose this? Again, I don't know. I think so, and certainly I'd prefer it, but this area seems pretty gray with a lot of slippery slopes. Should Anthropic be the one enforcing it? Well, like it or not, with this, they are.\n\nAnyway, the far bigger problem I have is that this isn't just about text that people are copying and pasting elsewhere. *It's about all text*. As in, the kind you're being *served up* inside of Claude as well – even if it's not meant to be put elsewhere.\n\nAnthropic is admitting that they're making that output to you *worse* in order to implement this system.\n\nAgain, that's insane. It may sound small or subtle, but it's just an awful decision and precedent to manipulate your product in this way. Even if you don't notice? **Especially if you don't notice**. This is full on subversion at scale.\n\nBut as mad as I may be about this, it's nothing compared to John Gruber, who devoted [almost 5,000 words to the matter this weekend](https://daringfireball.net/2026/08/anthropics_watermark_text_adulteration_in_claude_is_a_perversion_of_writing?ref=spyglass.org), calling it not \"subversion\" but \"perversion\". Here's a taste:\n\nOne of my fundamental problem with this is that no two synonyms carry theexactsame meaning. “He leaped at the chance” and “He jumped at the opportunity” are very similar sentences expressing the same general sentiment, but they arenotthe same. The exact words we choose when writing matter. I want any LLM I use to choose the very best, most precise words at every single decision point. An obvious constraint that I accept is time and computation. Within the constraint of executing inference quickly, and at a certain cost per token, I want the best words. This constraint matches human writing. I could surely write a better column by taking longer to write it. I write with a sense of how much care I should put into every word and punctuation choice I make. I take more time with certain paragraphs, sentences, or even individual word choices when my gut feeling says I should.\n\nIn other words, these arenecessarytrade-offs. These factors are all inmyinterest: speed, cost, quality. Ideally I would like perfect writing, at instantaneous generation speed, at zero cost. None of those things are possible. Computation is not free of charge (and cloud-based LLM inference with leading models is actually expensive). Inference is not instantaneous. And great writing, whether natural or artificial, can onlyapproachperfection.\n\nThe idea that anything other thanmyneeds should factor into the generation of textfor meis patently offensive.\n\nIncidentally, I happened to write about AI and writing [just yesterday](https://spyglass.org/ai-writing/), before I started to dive into this situation, noting the import I place on word selection. This Anthropic nonsense throws that notion right out the window. You may as well just buy a [Mad Libs](https://en.wikipedia.org/wiki/Mad_Libs?ref=spyglass.org) and run a random synonym generator.\n\nFor all [the talk](https://spyglass.org/ai-writing-inputs-outputs/) about AI slop, *this is AI slop*!\n\nAnd for a company that likes to [portray](https://spyglass.org/amodei-another-day/) and [convey](https://spyglass.org/anthropic-vs-administration/) moral [superiority](https://spyglass.org/ai-nuclear-weapons/) at every [turn](https://spyglass.org/anthropic-becomes-the-villain/), and to hold [themselves](https://spyglass.org/open-weight-ai-models/) as the [stewards of AI](https://spyglass.org/ai-overkill/), it's a bad look. Yes, it's being somewhat forced upon them by the EU, but they don't have to do this everywhere, and yet they are.\n\nThey'll say that's just easier, and I'm sure that's true. But it doesn't make it better. That makes it worse. This is a nearly $1T company. They're about to be one of the largest companies in the world. Hell, they already are. Put in the effort to not make your service shittier for every single user.\n\nAgain, it may seem like a small thing. But small things matter. And they tend to snowball...\n\nHonestly, what Anthropic should do here is tell the EU, just as their soon-to-be Big Tech peers, Apple, Google, and Meta do, that their rules are stupid. As with [other such EU rules](https://spyglass.org/tag/eu/), they appear at first glance to perhaps be coming from the right place, but they're usually just [political grandstanding](https://spyglass.org/red-red-whine/). Worse, [they seemingly don't think](https://spyglass.org/the-eu-clown-town/) through the [very obvious second-order effects](https://spyglass.org/eu-no-tech-zone/) of [such rules](https://spyglass.org/europe-cookie-banners/).\n\nIf the EU mandates that the big AI tools have to implement these \"safeguards\", the people most likely to take advantage of such things are just going to turn to other tools. And what happens when running models locally on machines becomes far more commonplace? There's just no way this actually works. Again, all it will do is make what were the most-used AI tools worse. And that, in turn, may make the most-used AI tools less used.\n\nOf course, that often does seem like the EU's basic mandate. So...\n\nI digress. This is all just so nutty. Not that it's happening – again, Google already has been doing this and the EU is mandating it – but that it needs to be explained to these companies why it's such a batshit idea. Again, they're making their products worse – *on purpose*. [Writing with AI was already contentious](https://spyglass.org/ai-writing-inputs-outputs/). Now it's nuclear.\n\nWhy does this need to be spelled out? It's almost as if they have [a reading-the-room problem](https://spyglass.org/i-am-the-great-glassholio/) here. And [perhaps](https://spyglass.org/ai-steve-jobs/) that is [impacting](https://spyglass.org/openai-tbpn-soft-power/) the overall [perception](https://spyglass.org/ai-perception-problems/) of AI. [You think](https://spyglass.org/ai-luddites/)?\n\nI'll let Gruber take us home:\n\nThis isn’t just about text one might generate with the intention of passing it off as their own natural work. This isn’t even about LLM proofreading of work written by hand. Anthropic is saying thatallnew Claude models are going to adulterate every single bit of text longer than 200 tokens (~150 words) they generate, including everything it presents to its users to read. So even in a private conversation between a user and Claude, which will never be read by anyone other than the user, Claude will begin making word choices in the name of marking its output in statistically predictable ways rather than maximizing clarity and precision.\n\nEven today’s so-called frontier models are already decidedly[lacking in lucidity]. Claude, ChatGPT, Grok, et al. are “better writers” than most humans and produce better prose than the median human. But: no shit. Most people are terrible writers. The “average person” is pretty stupid and[half of all people are stupider than that]. And there are many smart, interesting people who are miserable writers. So as impressive as LLMs are, the bar is low. The best writing I see come out of these models is worse than anything I would choose to read for pleasure. And now Anthropic is saying they’re going to make it worse, on purpose, for purposes that do not benefit me in any way? Even if only slightly worse?\n\nGet fucked.\n\n**Previously, on**\n\n**Spyglass**\n\n**...** 1 It was also framed more around images, video, and audio \"watermarking\" and less around the text element.\n\n2 And [what outrage this caused back then in the blogosphere](https://massivegreatness.com/i-cant-not-be-trusted-fuckers?ref=spyglass.org). Which feels so quaint now.\n\n3 Some companies employ similar tactics for internal memos too. Best not leak!", "url": "https://wpnews.pro/news/you-gotta-fight-for-your-write", "canonical_source": "https://spyglass.org/anthropic-ai-watermarks/", "published_at": "2026-08-17 15:38:10+00:00", "updated_at": "2026-08-17 15:42:33.834099+00:00", "lang": "en", "topics": ["generative-ai", "ai-ethics", "ai-policy"], "entities": ["Anthropic", "Claude", "Google", "SynthID"], "alternates": {"html": "https://wpnews.pro/news/you-gotta-fight-for-your-write", "markdown": "https://wpnews.pro/news/you-gotta-fight-for-your-write.md", "text": "https://wpnews.pro/news/you-gotta-fight-for-your-write.txt", "jsonld": "https://wpnews.pro/news/you-gotta-fight-for-your-write.jsonld"}}