OpenAI is adding text watermarking in ChatGPT and Codex

Text watermarking is the hardest provenance problem in AI. Images and video carry redundant pixels where a signal can be hidden; text is a sparse sequence of discrete tokens that can be rephrased, translated, or rewritten with a single prompt. If OpenAI's textGrain works as claimed — biasing token selection in statistically detectable but humanly imperceptible ways — it's a meaningful technical achievement. But "matched or exceeded" is a marketing claim, and the real benchmark is robustness against adversarial modification, not lab comparisons. The EU-first rollout is the most revealing detail. OpenAI isn't shipping this as a consumer feature; it's shipping it as a compliance mechanism. The EU AI Act's transparency requirements for AI-generated content make watermarking a legal obligation in Europe, so the EU becomes a live-fire testing ground before any global deployment. This is the emerging pattern: the strictest regulator effectively becomes the beta-testing environment for trust infrastructure. The deeper story is fragmentation. OpenAI has textGrain, Google has SynthID for text, and Anthropic has its own variant built on SynthID's foundation. None of these are interoperable, and none are open standards. If watermarking is meant to be the backbone of AI content provenance, we're building a tower of Babel — every model leaves a different fingerprint, and no unified detector can read them all. And there's an uncomfortable circularity here. The companies generating the flood of synthetic text are now appointing themselves the arbiters of authenticity. That's a concentration of power worth watching, even when the intent is benign. As The Verge reports, the rollout begins in the EU — but the questions it raises are global. Source: https://www.theverge.com/ai-artificial-intelligence/1004880/openai-chatgpt-text-watermarks-eu-ai-act