Safety
Watermarking
Embedding a hidden, machine-detectable signal in AI-generated content so it can later be identified as AI-made.
01 ——
In plain English
Watermarking is the technique of marking AI outputs — text, images, audio, video — with a hidden signal that lets you (or a detector) verify later that the content was AI-generated. It's one of the main proposed defences against misuse of generative AI.
How it works:
- Text watermarking — biases the model toward certain tokens in a statistical pattern undetectable to humans but detectable to the watermark verifier (Google's SynthID-Text, OpenAI's research)
- Image watermarking — embed imperceptible perturbations in pixel patterns (SynthID-Image, Stable Signature, Adobe Content Credentials)
- Audio / video — similar perturbation in spectrograms or frames
- Metadata — C2PA standard adds signed provenance to file headers
Where it's deployed:
- Google SynthID — applied to Imagen, Veo, MusicLM, Gemini text outputs
- Adobe Content Credentials (C2PA) — Photoshop, Firefly
- Meta — applied to AI-generated imagery on their platforms
- OpenAI — DALL-E images include C2PA metadata
Limits: Watermarks can be stripped by re-screenshotting, re-encoding, or simply paraphrasing text. Watermarking buys provenance for honest actors but doesn't stop motivated bad actors.
02 ——
Related terms
Deepfake
AI-generated media — usually video or audio — that convincingly impersonates a real person saying or doing something they didn't.
Generative AI
AI systems that create new content — text, images, audio, video, or code — rather than just classifying or predicting from existing data.
AI Avatar
A photorealistic or stylised digital character driven by AI — used for video presenters, customer service, training, and marketing.
Text-to-Image
AI that generates new images from a written description — the technology behind tools like Midjourney, DALL-E, and Stable Diffusion.
Guardrails
Rules and filters that constrain what an AI model can output — used to block harmful, off-topic, or non-compliant responses.
Alignment
The challenge of making AI systems behave in ways that match human values and intentions — not just their literal instructions.
Back to glossaryLast reviewed May 2026