As generative AI becomes increasingly powerful, one question grows more urgent: Is the face, the text, or the video I'm seeing real or AI-generated?
Both technical and regulatory fronts are now trying to provide answers, but progress is uneven.
Research: Invisible Watermarks
Arizona State University (ASU) computer science professor Yingzhen Yang released a decentralized attribution technique this year — essentially a method of hiding information inside AI-generated content.
Rather than slapping an AI-generated label on an image, it works by directly modifying the internal weights of the AI model.
After processing, every image produced by the model carries a unique hidden signature embedded in pixel clusters — completely invisible to the naked eye but readable by detectors. Yang's analogy: Like snowflakes, each model generates a unique pattern.
The system can tell you three things:
- Whether an image is AI-generated
- Which model generated it
- (In certain scenarios) Which user requested it
This has major implications for deepfake tracing — not just labeling, but tracking.
Legislation: Arizona Leads the Way
Yang's research is nearly synchronized with Arizona legislation: local lawmakers are advancing SB 1786, a bill requiring AI tool manufacturers to build invisible watermarking technology into their products.
The bill has passed the state Senate and is awaiting a vote in the House.
EU Legal Deadline: August 2026
Greater pressure comes from the European Union.
Under Article 50 of the EU AI Act, effective August 2026:
- AI systems must inform users when they are interacting with AI
- Synthetic audio, video, and images must be labeled as AI-generated
- Deepfake content must be clearly marked
Penalties: up to €15 million or 3% of annual global turnover, whichever is higher.
The EU is currently drafting specific implementation rules, with a final version expected by May-June 2026. These rules will prescribe a multi-layered combination of watermarks, metadata embedding, and invisible markers — because the EU believes no single technology can reliably solve the problem alone.
India Is Also Moving
India updated its IT regulations this year, requiring synthetic media to include watermarks, C2PA manifests, or equivalent protocols by default. Major platforms must collect uploader declarations on whether content is AI-generated or has a legitimate source.
Where the Difficulty Lies
Watermarking technology itself has a fundamental vulnerability: it requires active cooperation from model providers.
For open-source models, anyone can release a version without the watermarking mechanism. There is currently no good regulatory solution for this loophole.
In other words, watermarks can govern companies willing to be governed, but not malicious actors. This doesn't mean watermarks are useless — they are a necessary condition, but not a sufficient one.
The endgame for content authenticity is probably not a single technology, but a combination: watermarks + provenance metadata + platform verification + legal liability. That's essentially the direction the C2PA (Content Authenticity Initiative) is pushing.
Sources: CocoLoop, ASU researcher develops invisible watermark system to detect AI deepfakes (KOLD News); EU AI Act: First Draft Code of Practice on Transparency and Watermarking (Cooley LLP); What the EU's New AI Code of Practice Means for Labeling Deepfakes (TechPolicy.Press)