AI content watermarking
How it works
For LLM text: a cryptographic key plus the context of preceding tokens deterministically split the vocabulary into "green" and "red" tokens; the logits of green tokens are softly boosted so the model statistically favors them. Detection needs no model access — it measures the fraction of green tokens and applies a statistical test (p-value). SynthID-Text (Nature, 2024) generalizes this via "Tournament sampling". For image/audio/video the mark is embedded in the signal representation so it survives common transformations (compression, scaling, cropping). C2PA works differently — it does not modify the content but attaches a signed provenance metadata manifest. A central challenge is robustness against attacks that remove or mask the mark (paraphrasing, editing, compression).
Problem solved
AI-generated content is increasingly indistinguishable from human content, enabling disinformation, fraud, copyright violations and the "poisoning" of training data. Classic post-hoc detectors are unreliable and biased. Watermarking provides a signal built into the content itself at creation time, making detection far more dependable.
Components
During decoding the vocabulary is split into "green" and "red" tokens; green tokens are promoted, leaving a statistical trace in the text.
Algorithm that detects the signature (e.g. the share of green tokens) and returns a statistical test with an interpretable p-value, without access to model weights.
Google DeepMind system embedding an invisible mark in AI content across modalities — image, audio, video and text.
Open standard attaching a cryptographically signed manifest about origin and edit history to a file — complements watermarking with a metadata layer.
The mark’s ability to survive removal or masking attempts: text paraphrasing, editing, compression, image scaling or cropping.
Evolution
Formation of the Coalition for Content Provenance and Authenticity (C2PA) — a content provenance standard based on signed metadata (Content Credentials).
Kirchenbauer et al. publish "A Watermark for Large Language Models" — the green/red token method with statistical-test-based detection.
Launch of SynthID for watermarking AI-generated images, later extended to audio, video and text.
Publication of "Scalable watermarking for identifying large language model outputs" (Nature) — scalable LLM text watermarking based on Tournament sampling.
The EU AI Act enters into force; its Article 50 introduces transparency and marking obligations for AI-generated or manipulated content.