OpenAI Introduces Invisible Text Watermarks
OpenAI recently announced its commitment to integrating machine-readable invisible watermarks into eligible ChatGPT and Codex text outputs across the European Union over the forthcoming weeks. This strategic initiative ensures strict compliance with the transparency mandates outlined within the European Union Artificial Intelligence Act.
The Innovative textGrain Technology
The underlying technology, elegantly named textGrain, seamlessly embeds imperceptible statistical signals directly into the word selection process of the model. Subsequently, specialized detectors identify these subtle markers to ascertain whether a specific passage bears the authentic OpenAI watermark.
Moving forward, the enterprise will prioritize deploying this cryptographic feature throughout the European continent. In the upcoming weeks, the development team will exclusively introduce this functionality to eligible ChatGPT and Codex users across all subscription tiers within the European Union. Notably, the organization will not establish this feature as a universal, global default upon its initial release.
Opt-in Features for Global API Clients
Furthermore, OpenAI enables an opt-in text provenance feature for global API clients. Starting immediately, international API consumers can elect to integrate watermarked text outputs for selected models. This crucial flexibility empowers organizations to determine precisely how these markers align with their specific transparency obligations and desired user experiences. Concurrently, the engineering team is actively collaborating with cloud partners. They aim to deliver robust watermarking capabilities for model outputs accessed through third-party services in the approaching weeks.
Providing Access to Specialized Detectors
Finally, the firm will grant exclusive detector access to esteemed researchers and specialized expert organizations. Approved academic and professional institutions may submit their formal applications starting today. Initially, administrators will grant access on a strict case-by-case basis, governed by a comprehensive code of conduct, to foster the continuous evaluation and refinement of text provenance methodologies. While this sophisticated tool reliably reports the presence of a cryptographic watermark, it absolutely will not disclose the user identity, original prompt instructions, or sensitive conversational context.
