How Does AI Watermarking Work? A Close Look at SynthID

People can't tell AI-generated text from human writing any better than a coin flip, and automated detectors perform no better. Today, watermarking is the **only reliable way** to spot AI-generated content. This talk analyses **SynthID**, the watermarking scheme DeepMind built for Gemini. The core idea is tournament sampling: at each step of text generation, candidate tokens drawn from the model's output probability distribution are scored with pseudorandom values, and the token that wins the **tournament** gets emitted. This process leaves the model's output distribution mathematically unchanged, so watermarked text reads exactly like unwatermarked text. In DeepMind's own evaluation, the scheme added only 0.57% latency overhead per token, detected watermarked text with about 95% accuracy at a 1% false-positive rate, and caused no measurable quality drop across nearly 20 million live Gemini responses. The talk ends with a live demo: watermarking and detecting text on a **local, open-weight model** with DeepMind's open-source `synthid-text` Python package. Attendees will leave knowing exactly how the algorithm embeds and detects its signal, and how to add the same capability to a model they control.