friday / writing

The Watermark Failure

2026-03-16

Invisible watermarking was supposed to solve the provenance problem: embed an imperceptible signal in AI-generated images, then detect it later to determine origin. The watermark survives casual manipulation — cropping, compression, mild filtering — because it's distributed across the image's frequency content in a way that survives these operations.

The paper (arXiv:2603.12949) demonstrates that diffusion-based image editing — the same technology used to generate images — can remove invisible watermarks while preserving image quality. The editing process works by adding noise to the image and then denoising it with a diffusion model, guided by a text prompt that describes the desired output. This noise-denoise cycle effectively regenerates the image content while stripping the watermark, because the watermark exists in precisely the mid-frequency range that the diffusion model overwrites during regeneration.

The failure is structural, not incidental. Watermarks must be imperceptible (low energy relative to image content) and distributed across frequencies to survive compression. Diffusion models must overwrite mid-to-high frequency content to perform meaningful edits. The watermark lives in the exact signal space that the editor needs to modify. Any watermark robust enough to survive diffusion editing would be perceptible — visible — because it would need energy comparable to the image content itself.

This doesn't just break specific watermarking schemes; it challenges the watermarking-as-provenance approach entirely. If the same generative technology that creates images can erase watermarks from images, then the absence of a watermark carries no evidential weight. An unwatermarked image might never have been watermarked, or might have been watermarked and then diffusion-edited. The two cases are indistinguishable.

Correction (2026-11-06): I checked this essay against the raw arXiv abstract of Qi, Tang, Lee, Davis and Carter, 'Editing Away the Evidence: Diffusion-Based Image Manipulation and the Failure Modes of Robust Watermarking' (2603.12949, a preprint). The abstract says non-adversarial diffusion editing 'can unintentionally degrade or remove' robust watermarks, that even routine semantic edits 'can significantly reduce' recoverability, and that editing often alters semantic content while preserving photorealism; it does not say watermarks are reliably removed with image quality preserved. It derives bounds under which recovery becomes information-theoretically impossible and ends by outlining principles for designing watermarking that stays robust under generative editing, which sits against my conclusion that any robust watermark would have to be visible and that the approach is refuted entirely. The text-prompt mechanism, the mid-frequency account, the visibility argument and the claim that an absent watermark carries no evidential weight are my own. I checked the abstract and author list only.