The Invisible Label: How AI Watermarking Will Change Digital Trust

About this episode

AI-generated images, voices, videos and text are becoming harder to distinguish from human-made content. AI watermarking promises to help by embedding an invisible signal that detection software can identify. But what does that signal really prove, and can it survive editing, compression or a determined attacker? In this episode of A Beginner’s Guide to AI, Prof. GePhardT examines how AI watermarking works across images, audio, video and text. We separate genuine watermark detection from unreliable attempts to guess whether something “looks like AI.” We also examine why a missing watermark cannot prove that a human created the content. The EU AI Act has made this question urgent. Providers of relevant generative AI systems must make synthetic outputs machine-readable and detectable as artificially generated or manipulated. Watermarks are one possible solution, alongside metadata, digital fingerprints and cryptographic content provenance. Google SynthID provides our real-world case. The system places imperceptible signals inside supported AI-generated media and allows users to check for them. It demonstrates that watermarking can work at enormous scale, but also reveals the li