--- title: AI text watermarking and quality loss | daily.dev description: Watermarking AI text output may reduce quality, despite Anthropic's claim that it doesn't. Anthropic says 'watermarking does not impact the quality of Claude's... canonical: https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: AI text watermarking and quality loss | daily.dev og:description: Watermarking AI text output may reduce quality, despite Anthropic's claim that it doesn't. Anthropic says 'watermarking does not impact the quality of Claude's... og:url: https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g og:image: https://api.daily.dev/og/posts/BCH2plL4g.png og:image:alt: AI text watermarking and quality loss og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # AI text watermarking and quality loss **[Hacker News](https://daily.dev/sources/hn)** · 8 min read · 0 upvotes · 0 comments ## Summary Watermarking AI text output may reduce quality, despite Anthropic's claim that it doesn't. Anthropic says 'watermarking does not impact the quality of Claude's output,' and the SynthID-Text paper, which Anthropic based its design on, claims to 'preserve response quality.' But that paper's own supplement admits a 'diversity/detectability trade-off': watermarked responses to the same prompt are more similar to each other than unwatermarked ones. Consultant Natzir Turrado tested this on an open model and found 4 of 5 watermarked responses to the same prompt opened almost identically, versus 3 different openings for the unwatermarked model. This reduced diversity affects retries, creative writing, and techniques that repeat prompts to gauge model confidence. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### Does watermarking AI-generated text like SynthID-Text actually reduce the diversity of responses to the same prompt? Yes, Google's own SynthID-Text supplementary material acknowledges a 'diversity/detectability trade-off,' noting that watermarked responses to the same prompt are more similar to each other than unwatermarked ones. A test reproducing SynthID-Text on an open model found 4 of 5 watermarked responses to the same prompt opened almost identically, versus 3 distinct openings without watermarking. _Anyone weighing AI watermarking trade-offs can track this kind of analysis on daily.dev._ ### Does Anthropic's text watermarking in Claude affect output quality? Anthropic states that watermarking does not impact the quality of Claude's output, but that claim was tested only on single responses, not on how people actually use chat models through repeated prompts, retries, or multiple generations. Critics argue the real quality cost shows up in reduced diversity across multiple responses to the same or similar prompts, which single-response testing would not detect. _Developers evaluating Claude's reliability for creative or iterative tasks can follow this debate on daily.dev._ ### Why would repeating the same prompt multiple times to check for AI hallucinations be less reliable with a watermarked model? Because watermarking steers token selection using a watermark key, repeated responses to the same prompt become less independent of each other, so the apparent 'consensus' among multiple generations is less trustworthy than it would be with an unwatermarked model. This undermines the common technique of sampling a prompt several times to gauge confidence and reduce hallucinations. _Those relying on self-consistency checks to catch hallucinations can track watermarking research on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 3 discussions and 13 comments across hackernews (as of 2026-11-01). **TL;DR:** Discussion centers on whether reduced response diversity from watermarking counts as a quality loss, with the article's author defending the claim against pushback that diversity and quality are distinct concepts. **Sentiment:** 10% positive · 35% mixed · 55% skeptical **The case for** - Some argue reduced diversity could mean more consistent (not necessarily worse) answers, which some users would actually welcome. - One detailed technical rebuttal argues SynthID-style watermarking doesn't change the token distribution in expectation, so one-shot generation quality shouldn't be affected. **The pushback** - Several argue conflating diversity loss with quality loss is a flawed premise and the piece lacks benchmarks showing actual degraded quality. - One commenter who re-ran the test with hundreds of prompts found results broadly consistent with the original near-identical-response finding. - Watermarking is criticized as affecting techniques that repeatedly sample a model to infer its probability distribution or confidence. - A couple of comments mock recognizable 'AI-written' phrasing patterns as a practical downstream effect of homogenized output. **By community** - hackernews (heated): A detailed technical disagreement unfolds between the article's author and a commenter over whether diversity loss equals quality loss, alongside side discussion mocking recognizable AI phrasing. **Hottest debate:** Whether reduced response diversity from watermarking constitutes a genuine quality loss or is a separate, unrelated property. **Open questions** - Can more complex watermarking strategies reduce the diversity loss without sacrificing detectability? - Is there a rigorous benchmark that shows watermarking consistently degrades output quality rather than just diversity? **Highlights** > This article is wrong. SynthID-style watermarking does not affect one-shot generation quality. When news about the watermarking went around a couple of months ago I too thought there must be a quality/detectability tradeoff (as there is for images), but I dug into the math more and realized that my intuition was incorrect. They key idea is this: changing the random seed for LLM generation definitionally does not change the distribution of tokens in expectation over all seeds. This can already be used to embed information in LLM streams, leading to LLM steganography work (https://arxiv.org/abs/2606.09135, but the idea dates back much further). Take your target message, whiten it via a one-time pad (stream cypher), then use it as the 'compressed stream' under an arithmetic decoding process (https://en.wikipedia.org/wiki/Arithmetic_coding) with the LLM-supplied logits as the assumed probabilitiy distribution. SynthID pulls off the same kind of trick, save that its 'tournament structure' allows application after the sampling step rather than before. The random seed under SynthID becomes context sensitive (keying on the prior N tokens as well as the watermark seed), but in expectation the next-token distribution is still the same as the unwatermarked case. If there's no change in token distribution, then there can't be a change in quality. However, the article and its sources are right about one factor: random generation diversity will suffer. Applying SynthID to an LLM stream acts like fixing the PRNG seed. Repeated generations with the same prompt and same seed will deliver the same result, up to small amounts of nondeterminism in the architecture (mostly GPU scheduling and roundoff). This matters under the same circumstances that it would matter if the provider said "we sample using the XYZ PRNG with fixed seed '42'". In my opinion, 'sample diversity' and 'sample quality' are not the same thing, and conflating the two concepts leads to poor intuition. SynthID and the like probably matter most if one is trying to repeatedly sample an LLM in order to infer the probability distribution (e.g. using a commercial LLM provider as a Jev-style classifier, looking for calibrated answers). > — [Majromax on hackernews · 1 comments](https://news.ycombinator.com/item?id=50324872) > I wrote the article. How you define quality is exactly the issue. > In my opinion, 'sample diversity' and 'sample quality' are not the same thing, and conflating the two concepts leads to poor intuition. But you're assuming everyone has the same notion of quality as you. As I point out in the piece, even the paper's writers use "quality" to show how watermarking can "lower the overall quality of a system" > SynthID and the like probably matter most if one is trying to repeatedly sample an LLM in order to infer the probability distribution Which is another impact I write about in the piece. If you evaluate the output quality of an LLM system in this way, you're impacted more by watermarking. > — [k1m on hackernews](https://news.ycombinator.com/item?id=50325468) > The linked article in Spanish contains substantially more detail. Natzir Turrado tested hundreds of prompts. For this example; he tested 30 prompts at 5 pass-pairs each; so 300 times. https://natzir.com/posicionamiento-buscadores/que-es-como-fu... > — [dannyw on hackernews · 1 comments](https://news.ycombinator.com/item?id=50322505) > It's an interesting review, but it seems to mix some things together. Gruber's claim is that we shouldn't sacrifice quality, but none of those sources show that's what's happening. Having less diversity in answers could mean you're more consistently getting a better/worse/same quality answer. There are lots of people complaining about not being able to get a real temperature=0 answer - they'd actually welcome the consistency too. The practical example of not liking the sentence is not great in my opinion. Instead of just re-rolling the answer, why not continue in that thread and say what you don't like? Or request 20 alternatives, say which go in the right direction and request 20 more? Retrying from scratch with one sample is a waste of time and doesn't indicate anything about quality. Basically, it's all theoretical - show us the actual benchmarks/samples where the quality is consistently degraded. > — [viraptor on hackernews · 1 comments](https://news.ycombinator.com/item?id=50321657) > > Natzir Turrado, a technical consultant, actually tested this (article in Spanish) by reproducing SynthID-Text on an open model with the paper's settings and a watermark key. He asked the watermarked model five times to describe a memorable trip to a coastal town. Four of the five answers he got back opened in a near-identical way: “My heart still sings/dances when I think back to my trip to the quaint coastal town of…”. For the unwatermarked model, three opened in different ways. A test of 2/5 vs 4/5 seems to be an obviously invalid experiment. > — [ethanj8011 on hackernews · 1 comments](https://news.ycombinator.com/item?id=50321944) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=50320599) · 43 points · 13 comments - [hackernews](https://news.ycombinator.com/item?id=49426464) · 3 points · 0 comments - [hackernews](https://news.ycombinator.com/item?id=49436526) · 2 points · 0 comments ## Similar posts on daily.dev - [Readers can't identify watermarked AI text](https://daily.dev/posts/readers-can-t-identify-watermarked-ai-text-me1g7n4eb) · sean goedecke · 0 upvotes · 0 comments - [Text AI watermarks will always be trivial to remove](https://daily.dev/posts/text-ai-watermarks-will-always-be-trivial-to-remove-borbz7x9u) · sean goedecke · 6 upvotes · 4 comments --- Tags: [#llm](https://daily.dev/tags/llm), [#claude](https://daily.dev/tags/claude), [#ai-safety](https://daily.dev/tags/ai-safety) [View this post on daily.dev](https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"AI text watermarking and quality loss","url":"https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g"},"datePublished":"2026-11-01T04:36:22.535Z","dateModified":"2026-11-01T14:37:12.239Z","description":"Watermarking AI text output may reduce quality, despite Anthropic's claim that it doesn't. Anthropic says 'watermarking does not impact the quality of Claude's...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d4b52f9efed4b39a14492ab603348fae?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/d4b52f9efed4b39a14492ab603348fae?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"llm,claude,ai-safety","timeRequired":"PT8M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"AI text watermarking and quality loss"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/ai-text-watermarking-and-quality-loss-bch2pll4g#faq","mainEntity":[{"@type":"Question","name":"Does watermarking AI-generated text like SynthID-Text actually reduce the diversity of responses to the same prompt?","acceptedAnswer":{"@type":"Answer","text":"Yes, Google's own SynthID-Text supplementary material acknowledges a 'diversity/detectability trade-off,' noting that watermarked responses to the same prompt are more similar to each other than unwatermarked ones. A test reproducing SynthID-Text on an open model found 4 of 5 watermarked responses to the same prompt opened almost identically, versus 3 distinct openings without watermarking. Anyone weighing AI watermarking trade-offs can track this kind of analysis on daily.dev."}},{"@type":"Question","name":"Does Anthropic's text watermarking in Claude affect output quality?","acceptedAnswer":{"@type":"Answer","text":"Anthropic states that watermarking does not impact the quality of Claude's output, but that claim was tested only on single responses, not on how people actually use chat models through repeated prompts, retries, or multiple generations. Critics argue the real quality cost shows up in reduced diversity across multiple responses to the same or similar prompts, which single-response testing would not detect. Developers evaluating Claude's reliability for creative or iterative tasks can follow this debate on daily.dev."}},{"@type":"Question","name":"Why would repeating the same prompt multiple times to check for AI hallucinations be less reliable with a watermarked model?","acceptedAnswer":{"@type":"Answer","text":"Because watermarking steers token selection using a watermark key, repeated responses to the same prompt become less independent of each other, so the apparent 'consensus' among multiple generations is less trustworthy than it would be with an unwatermarked model. This undermines the common technique of sampling a prompt several times to gauge confidence and reduce hallucinations. Those relying on self-consistency checks to catch hallucinations can track watermarking research on daily.dev."}}]} ```