--- title: "The invisible ID inside AI images that survives a screenshot" date: 2026-11-09 category: Generative Media site: NeuroAI canonical: https://neuroai.site/a/na-aigc-invisible-watermark-tech language: en --- # The invisible ID inside AI images that survives a screenshot > Watermarks and signed metadata are becoming the machine-readable 'passport' for AI-made content — but each method breaks in a different way. You screenshot an AI-produced image, crop it, and post it elsewhere. The visible "AI-produced" tag is gone. To a human, it now looks like any other picture. To a machine, though, it may still be saying "I was made by an AI." That quiet gap — between what people can see and what software can detect — is why 2025 and 2026 became the years of the invisible watermark. ## Two kinds of label China drew the line first. On September 1, 2025, the Measures for Labelling AI-Produced and Synthetic Content (人工智能生成合成内容标识办法), together with the mandatory national standard GB 45438-2025, took effect. They require providers to add two distinct marks to AI-produced text, images, audio, video and virtual scenes: an explicit label (显式标识) that humans can see, and an implicit label (隐式标识) embedded in the file's data that is not meant to be visible. Crucially, the rules say the implicit mark cannot rely only on file metadata that is easily stripped — watermarking is explicitly encouraged. The European Union followed on a different track. Article 50 of the EU AI Act, its transparency obligation for AI-produced content, began applying on August 2, 2026. Providers must make synthetic text, image, audio and video detectable in a machine-readable form, with penalties reaching €15 million or 3% of global turnover. ## What the technology actually does There are two main technical approaches, and they fail differently. The first is cryptographically signed metadata, most visibly the C2PA Content Credentials standard (now also an ISO standard, ISO 22144). It attaches a signed "birth certificate" to a file: who made it, with what model, what was changed. The strength is that any later edit breaks the signature, proving tampering. The weakness is that metadata is fragile — a screenshot, a re-save, a messaging app, or a social platform routinely strips it. Once it is gone, the provenance is gone. The second is an embedded watermark placed inside the content itself. Google's SynthID, for example, nudges the probabilities a language model uses when picking each token, leaving a statistical pattern in AI-written text that specialized software can later test for — while the text still reads normally. For images and video, similar signals are baked into the pixels or frames. These survive transformations that destroy metadata, but they are not magic: a heavy crop, pasting the image onto a new background, or a heavy re-encode can still erase them. ## Where the two meet In practice, the industry consensus is "both." Alibaba, for instance, applies China's TC260 implicit label to every image produced by its Qwen image model, according to a November 2026 test by the verification firm Koncile across 37 models. OpenAI and Google sign images with both C2PA metadata and their own watermarks. TikTok, which says it has labeled billions of AI videos, began testing a platform-readable invisible watermark in late 2025 to cope with content re-uploaded or edited after the visible tag was lost. The point is not that any single method is perfect. It is that an AI file is starting to carry a machine-readable record of how it came to exist — and that record is designed to survive the things humans do to files. For readers, the quickest check today is a free browser-based provenance tool that reads C2PA or TC260 fields; it only flags what a creator actually embedded, but is fast and decisive when that field is present. ## What readers can do now - If you publish AI-produced visuals, keep the original file with its metadata intact and avoid unnecessary re-saves; a clean export preserves the most detectable signal. - If you need to verify whether an image is AI-made, check the metadata first (C2PA/EXIF/TC260) with a free checker — it is fast and decisive when present, but absence is not proof either way. - If you run a platform or brand, plan for the case where the label is gone: pair watermark detection with pixel-level forensic models, because screenshots and re-encodes are exactly where single-method detection fails. ## Honest limitations This article explains the mainstream techniques (C2PA, SynthID, TC260) and does not cover every vendor product; Aliyun's own AI DeepSign service is one commercial example but is not independently benchmarked here. Detection rates vary by model and by transformation, and no method survives all edits. The EU Article 50 effective date and the China measures' effective date are stated as enacted; implementation details continue to evolve, and open-source models run locally remain outside mandatory labelling. ## Key takeaways - China's GB 45438-2025 and the EU AI Act Article 50 have made implicit, machine-readable labelling a legal requirement in two major markets. - Signed metadata (C2PA) and embedded watermarks (SynthID, TC260) complement each other — metadata is rich but fragile; watermarks survive more but not all edits. - The practical goal is not a perfect tag but a layered record that survives screenshots, crops and re-encodes better than a visible label alone. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-aigc-invisible-watermark-tech Free to quote with attribution and a link to the original.