The decision to watermark AI-generated text only where regulators demand it reveals something uncomfortable about OpenAI's priorities: transparency is a compliance cost, not a product feature.
Starting this month, ChatGPT outputs in the European Union will carry invisible digital watermarks—cryptographic signatures embedded in the text that allow detection tools to identify AI-written content. Users elsewhere get no such disclosure. The move satisfies the EU AI Act's transparency requirements while preserving the seamless, undetectable experience that has made ChatGPT indispensable to millions of students, marketers, and professionals who would rather not advertise their reliance on the tool.
The technical half-measure
Watermarking AI text is not new; researchers have proposed methods for years, and OpenAI itself developed robust techniques internally as early as 2022. The company long resisted deployment, citing concerns that watermarks could be circumvented and might disadvantage non-native English speakers whose writing already triggers false positives in crude AI detectors. Those concerns apparently evaporate when Brussels writes the rules.
The EU-only rollout also creates an obvious arbitrage: European users who want unmarked text can route requests through VPNs or API endpoints registered elsewhere. OpenAI knows this. The watermark is less a technical safeguard than a legal checkbox—a way to tell regulators the company tried.
Regulatory balkanization accelerates
This is the second major feature divergence driven by jurisdiction. Earlier this year, OpenAI began restricting certain plugins and data-retention practices in Europe while leaving them intact in the United States. Meta, Google, and Anthropic have all hinted at similar regional carve-outs as AI laws proliferate.
The result is an emerging patchwork where the AI you experience depends on where you live—or where your IP address claims you live. For users, that means inconsistent capabilities. For researchers studying AI's societal effects, it means incomparable datasets. For policymakers hoping to set global norms, it means watching companies comply minimally and selectively.
Our take
OpenAI's selective watermarking is rational corporate behavior and a minor public-relations defeat. The company spent years positioning itself as the responsible actor in a reckless industry; now it admits, implicitly, that responsibility is a regional setting. If watermarking were genuinely good for users—reducing misinformation, protecting academic integrity—it would be on everywhere. That it is not tells you what OpenAI actually believes about the feature's value versus its friction. Europe gets the transparency it legislated. Everyone else gets plausible deniability.
