OpenAI Rolls Out Invisible Text Watermarks for ChatGPT and Codex

What Happened

OpenAI is rolling out invisible, machine-readable watermarks for text generated by ChatGPT and its Codex coding tool, the company announced this week. The feature, built on a new internal system called textGrain, is being introduced in part to meet requirements under the European Union's AI Act.

How the Technology Works

The textGrain system embeds watermarks directly into generated text in a way that is not visible to human readers but can be detected by compatible machine-reading tools. OpenAI has not disclosed the precise technical method used to embed the signals, but the approach is designed to allow third parties, regulators, or platform operators to verify whether a given piece of text was produced by a covered OpenAI model.

The watermarks apply to output from both ChatGPT, OpenAI's consumer-facing conversational product, and Codex, its code-generation tool aimed at developers. The Verge reported the rollout is tied to OpenAI's textGrain system, which the company developed internally for this purpose.

Regulatory Context

The EU AI Act, which entered into force in August 2024 and is being phased in through 2026 and 2027, includes provisions requiring providers of general-purpose AI systems to label AI-generated content. The regulation applies to systems made available in EU member states, and major AI developers have been working to bring their products into compliance ahead of applicable deadlines.

OpenAI's watermarking rollout, as reported by The Verge and trendingtopics.eu, is positioned at least in part as a compliance measure for that regulatory framework. The EU market represents a significant share of ChatGPT's international user base.

Industry Comparisons

Anthropichas its own approach to labeling content generated by its Claude models, though the specific technical implementation differs from OpenAI's textGrain method, according to trendingtopics.eu. Google and Meta have also developed or announced content provenance and watermarking tools under the Coalition for Content Provenance and Authenticity, known as C2PA, a broader industry standard that uses cryptographic metadata attached to media files rather than embedded text signals.

Text watermarking presents distinct technical challenges compared to watermarking images or audio. Text can be easily edited, paraphrased, or reformatted, which can degrade or remove embedded signals. OpenAI has not publicly stated the robustness of textGrain against such manipulation, and the limits of the system have not been independently verified based on current reporting.

Company Background

OpenAI is the developer of the GPT series of large language models and operates ChatGPT, which the company has said surpassed 300 million weekly active users as of early 2025. Codex, which underpins the GitHub Copilot coding assistant developed with Microsoft, is one of OpenAI's primary enterprise-facing products. OpenAI completed a funding round valuing the company at $157 billion in late 2024.

The company has faced recurring scrutiny from regulators in the EU, the United States, and elsewhere over questions of AI transparency, content labeling, and the provenance of training data. The textGrain watermarking system represents one of the more concrete technical steps the company has taken in response to content-authenticity concerns.

What It Means in Practice

For users and developers, the watermarks are designed to be unobtrusive during normal use. Detection would require access to a compatible verification tool. OpenAI has not yet publicly released details of a detection API or indicated which third parties will have access to the verification layer.

For publishers, platforms, and regulators, the system would in principle allow automated screening of text to determine whether it originated from a covered OpenAI model, provided the watermark signal remains intact after any editing or reformatting.

OpenAI has not announced a specific public release date for textGrain beyond the current rollout phase, and further technical documentation is expected as the company advances toward full EU AI Act compliance deadlines.