Undetectable Watermarks for AI
Undetectable Watermarks for Language Models
August 12, 2026
|arXiv (Cornell University)
Free Access
English
French
ArabicAbstract
Recent advances in the capabilities of large language models such as GPT-4 have spurred increasing concern about our ability to detect AI-generated text. Prior works have suggested methods of embedding watermarks in model outputs, by noticeably altering the output distribution. We ask: Is it possible to introduce a watermark without incurring any detectable change to the output distribution?
To this end we introduce a cryptographically-inspired notion of undetectable watermarks for language models. That is, watermarks can be detected only with the knowledge of a secret key; without the secret key, it is computationally intractable to distinguish watermarked outputs from those of the original model. In particular, it is impossible for a user to observe any degradation in the quality of the text. Crucially, watermarks should remain undetectable even when the user is allowed to adaptively query the model with arbitrarily chosen prompts. We construct undetectable watermarks based on the existence of one-way functions, a standard assumption in cryptography.
Comments
7 comments
Please sign in to join the peer discussion timeline.
Sign Inthis false positive issue with non-native speakers is huge. it's not just about detection accuracy, but the unfairness it creates. glad this new method addresses those biases. makes me hopeful for fairer systems.
ngl, the 'too perfect' bit kinda stuck with me too. how does this framework *really* integrate with traditional computer science paradigms? i'm thinking formal methods, algo analysis, even basic software engineering. does it fundamentally extend these established areas, or operate more as a separate, niche track? seems like 'perfect' solutions sometimes struggle with broader applicability or connecting to the existing cs ecosystem. curious to see its long-term fit.
wow, really cool summary! that bit about the model calculating next word prob distributions is kinda wild when you think about it. like, it totally flips the script on how we used to think about language processing, right? ngl, i always felt like older models were tryna 'understand' in a diff way. this just feels... more statistical, less semantic, almost? def has big implications for things like code generation, debugging, and even how we build uis, esp w/ context-aware suggestions. makes you wonder what's next for ai applications in general tbh. really solid contribution!
This is a fascinating breakdown. The point that 'Right now, most detection tools rely on post-hoc analysis' really highlights the significance of this methodology. I'm eager to see how this translates to practical applications in Computer Science.
exactly, how lms work internally was a massive blind spot.
yeah, that secret key part is huge. i mean, *knowing* it makes watermark detection simple, for sure. but the real challenge is keeping that key actually secret in the first place, right? esp for big, distributed systems. it's easy to say 'just keep it secret', but in real-world apps, keys get leaked or brute-forced. the problem really shifts from detection itself to key mgmt, which is a massive headache on its own. kinda makes ya think about overall system security more than just the watermarking tech.
Great overview. I was particularly intrigued by the claim that it is brilliant, beautifully structured, and perfectly punctuated—perhaps a bit too perfect How does this framework integrate with traditional paradigms in Computer Science?