Kirchenbauer et al
Partition vocab into 'green' + 'red' per hash of preceding token. Bias toward green. Detector counts green fraction.
Advertisement
Stealth
Human reader doesn't notice. Undetectable without key. Small perplexity increase.
Advertisement
Robustness
Survives paraphrase partially. Removed by heavy rewriting. Not undefeatable.
Deployment
Anthropic, OpenAI have deployed variants. Not universally on. Trade-off with user preference for natural output.