Watermarking AI Text Is Fundamentally Flawed

Anthropic's watermarking of AI-generated text from its Claude models has sparked debate, but Google has been using SynthID for text since May 2024 and OpenAI announced similar provenance signals on August 2, 2026. Critics argue that watermarking is flawed due to copyright implications, false positives, and the difficulty of proving human authorship.

The recent revelation that Anthropic is watermarking text outputs from its AI models has ignited a firestorm of debate across social media. While the exact methodology is not explicitly detailed in their documentation https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content , it is presumably based on statistical word-choice biases designed to survive copying, pasting, and even subsequent AI proofreading. Anthropic themselves concede that this method "is not fully conclusive." We will explain more on this later. While the socialwebs' reaction has been swift and brutal, several critical implications are being overlooked in the mainstream discourse. Here is our perspective on why text watermarking is a fundamentally flawed paradigm. Anthropic is Not the Only One While Anthropic is currently absorbing the brunt of the negative publicity, they are hardly pioneers in this space. Google has been watermarking its generated text since at least May 2024 SynthID Paper https://www.nature.com/articles/s41586-024-08025-4 and Google Announcement of its usage in Gemini https://deepmind.google/blog/watermarking-ai-generated-text-and-video-with-synthid/ . OpenAI has similarly declared that it fully intends to add provenance signals into AI generated text as recently as August 2nd, 2026 OpenAI Help Center https://help.openai.com/en/articles/8912793-provenance-signals-content-credentials-synthid-in-openai-generated-content . The disproportionate outrage directed at Anthropic while Google has already implemented this more than 2 years ago and OpenAI have just announced as recently as 12 days ago they intend to do the same thing is a fascinating study in public relations, but it misses the larger point: text watermarking is now a widespread industry practice. We must evaluate it as a systemic shift, not on a company specific basis. The issue is much bigger than the current hate Anthropic is receiving over this issue. Copyright Implications In the United States, fully AI-generated content cannot be copyrighted, a stance reaffirmed by the US Copyright Office in recent years https://www.copyright.gov/newsnet/2025/1060.html . However, the threshold of human involvement required to make a work copyrightable remains a murky legal frontier. Watermarking introduces a dangerous variable into this already fragile equation. Consider the potential legal paradoxes: The Proofreader Dilemma: If you write a wholly original book but use AI to proofread and tighten the prose, a watermark detector might flag the entire manuscript as AI-generated. Does that strip you of your copyright? The Translator's Trap: What if you write a novel by hand, but use AI to translate it into French? The translated text's watermark could theoretically indicate the content as 100% AI generated. Does the French version lose intellectual property protections? The Coder's Canvas: Imagine spending days architecting an application making high-level design decisions, prompting, iterating, and testing. You poured human creativity into the logic, but an AI wrote the actual syntax. If the code is flagged as 100% AI-generated, is your software unprotectable, despite the massive human labor involved? Watermarking threatens to legally invalidate human-AI collaborative works by painting them with a broad, binary brush. The Danger of False Positives Both Google and Anthropic admit their text watermarks are not 100% accurate. If a completely human-authored work is falsely classified as AI-generated, the burden of proof unfairly shifts to the creator. How does one prove a negative? Must writers now record their screens and keystrokes or even have a camera pointing at them while they work to validate their authorship, or rely on vague assertions of "trust me, I wrote this"? False positives are particularly dangerous because they can be weaponized as tools for censorship or professional sabotage. If a bad actor wishes to discredit a journalist, rival, or colleague, they can simply run the target's writing through a gauntlet of different watermark detectors such as Anthropic's, Google's, OpenAI's, etc until they inevitably "detector-shop" their way to a false positive. They can then publicly shame the author, accuse them of academic or professional fraud, and let the algorithm's false authority do the damage. Each AI detector has a probability of a false positive and if they all use different methods then each time you test the same text though a different detector the probability of a false positive increases. The companies implementing watermarking text have themselves state that this method is not 100% accurate and thus watermarking has absolutely no business at all inflicting real world consequences for the people it flags. A Backdoor for Individual User Tracking There is currently no technical limitation preventing AI providers from implementing watermarking on a per-user basis. Instead of a universal Anthropic or Gemini watermark, a model could be tweaked to bias word choices in a specific, cryptographically unique pattern tied to your user account. Why is this a bad thing? It would turn AI text generation into a mass surveillance apparatus. Any text you publish on the internet such as an anonymous blog post, a whistleblower complaint, a sensitive email could be scraped, analyzed, and mathematically traced back to your specific AI account. It strips users of their right to anonymity and privacy, handing tech giants a permanent, invisible fingerprint hidden inside the very structure of our sentences. AI providers haven't shied away from split releases in the past, one for the world and one for the EU and even restrict EU's access until they comply with EU's laws. The fact that the current roll out has already been released globally and not just in the EU indicates that the AI providers themselves see benefit to them implementing this. One must ask themselves the reason that they see benefit adding watermarks to text. Degraded AI Output Quality Basic statistical logic dictates that forcing a model to adhere to a watermark degrades its performance. By artificially biasing word choices to create a statistically relevant signature, you limit the model's natural vocabulary, expressive range and the amount of options it can choose from. This all must have a negative impact on the quality of the model's output. It is our opinion that creativity and quality of answers never increased when the specific way you have to answer gets restricted. This inevitably forces a trade-off: model quality versus watermark detectability. It restricts nuance. Furthermore, if a model is heavily constrained by watermarking algorithms, will it ignore complex stylistic prompts e.g., "write this in the style of an 18th-century pirate" because adopting that persona would require abandoning its mandated statistical word biases? Some research is already supporting the idea that watermarking impacts model quality OpenReview https://openreview.net/forum?id=097IlaUfQY , ACL Anthology https://aclanthology.org/2024.findings-naacl.223.pdf , OpenReview https://openreview.net/pdf?id=PuhF0hyDq1 , arXiv https://arxiv.org/pdf/2403.19548 . The Illusion of Academic Integrity Educators are understandably anxious about students using AI to cheat on essays and exams. Proponents argue watermarking is the solution. We strongly disagree: because the method is probabilistic rather than absolute, it has no place in a punitive academic setting. A single false expulsion undoes any theoretical good this technology provides. Furthermore, it creates a Kafkaesque nightmare for students, characterized by a massive institutional power imbalance. A university runs an essay through an opaque, proprietary API; the computer says "AI generated." The student has no recourse, no way to audit the black-box algorithm, and no way to defend themselves other than pleading with the institution to "run it again." It replaces educational trust with algorithmic dogma. Combatting Disinformation: A Moot Point The argument that watermarking combats spam, astroturfing, and phishing ignores a fundamental reality: using AI as a collaborative writing partner is rapidly becoming the global default. Many professionals and students already use AI for brainstorming, translating, and drafting; within the next few years, writing without some form of AI assistance could very well be the exception, not the rule. If nearly all published text features some level of AI collaboration, detecting an AI watermark becomes entirely moot. It is reminiscent of the historical panic over spell-checkers and calculators "dumbing down" society an argument long since discredited. The real battle against disinformation is verifying whether a real human or an automated bot farm is publishing the text not agonizing over whether the text itself was partially generated by AI. The "Model Collapse" Fallacy Some argue watermarking is necessary so AI companies can filter out synthetic text when training future models, preventing "model collapse." However, this approach lacks nuance. A piece of text generated through deep human-AI collaboration is novel, high-quality data that should be used for training. By blindly filtering out anything with a watermark, AI companies will starve their next-generation models of valuable human-directed content. Instead AI training companies should evaluate works for novelty and actual usefulness of the text rather than relying on a blunt watermark to see if AI had touched the text at any time. An Arms Race is Afoot and Honest People Will Lose Out Here is the core failure mode that makes this entire exercise amount to security theater: watermark removal is a solvable adversarial problem, and the people most motivated to solve it are exactly the people the system is trying to catch. A statistical text watermark works by nudging word choice probabilities. Therefore, it can be tested against, just like any other classifier. Anyone with a motive to strip a watermark such as a plagiarism mill, a disinformation farm, or a spammer can run their output against public detectors, tweak the wording, paraphrase it through an un-watermarked open-source LLM, and iterate until the text registers as "clean." This is not speculative; it is the exact cat-and-mouse dynamic we have seen with plagiarism checkers and email spam filters. A lucrative cottage industry of "AI humanizer" tools already exists for this single purpose. As a result, the people these detectors actually catch are not the bad actors. They catch the honest users: the student who used AI just to fix their grammar, the author who asked an AI for feedback and then rewrote the passage, or someone who leaned on AI for translation due to language difficulties. These users don't run their work through five different evasion tools because they aren't trying to hide anything. A detection system whose false-negative population is "anyone who actively tried to evade it" and whose false-positive population is "anyone who didn't bother to hide" is structurally broken. It screens out honesty while giving dishonesty a free pass. Conclusion Watermarking AI text is a crude, superficial patch applied to a complex sociological shift. In its current form, it threatens to degrade the quality of AI outputs, muddy the waters of copyright law, create massive vulnerabilities for user tracking, and substantially hurt people whose work is falsely flagged as AI-generated. Worst of all, it acts as a trap for honest, transparent users while failing entirely to stop the malicious actors it was ostensibly designed to thwart. Worse still, it's accelerating something already underway: society is entering a new witch-hunt era around AI usage, where disliking someone's work increasingly means reaching for "It's ~~a witch~~ AI-generated" instead of engaging with the writing on its merits. We must move past primitive tools like watermarking before they cement that mentality further.