OpenAI Launches Text Watermarking in Europe, Optional API

OpenAI introduces a text watermark for its models, with access to detectors limited to approved entities and restricted guarantees. The deployment begins in Europe, while the API now offers an optional activation worldwide.

OpenAI restricts access to detectors to approved experts

Approved researchers and expert organizations can request access to the watermark detectors starting today. According to OpenAI, this access will be granted on a case-by-case basis, in accordance with the Code of Practice, to support the assessment and improvement of text provenance. The detection tool will only indicate the presence of an OpenAI watermark, without identifying the user or revealing their queries or conversations. Due to the risk of missed watermarks and false positives, OpenAI is not making the detector public at launch.

Global activation option for the API and arrival via cloud partners

Starting today, API customers worldwide can choose to activate the text watermark for certain models. This option allows customers to decide how to integrate the watermark into their transparency obligations and user experiences. OpenAI is also working with cloud partners to offer the watermark on models accessible through their services in the coming weeks.

European deployment on ChatGPT and Codex, without a global default setting

The invisible, machine-readable text watermark is being deployed for ChatGPT and Codex, initially reserved for eligible users in the European Union and across all plans. OpenAI is not making this watermark a global default setting at launch, preferring a regional approach to leverage real-world usage and feedback. Some users have not welcomed this new feature.

Claimed performance and limitations, in the shadow of the AI Act

OpenAI's watermarking system, named textGrain, is presented by the company as matching or exceeding other approaches such as Google DeepMind's SynthID. According to OpenAI, results from AI benchmarks show equivalent performance levels between watermarked texts and those without. However, the company emphasizes that textGrain does not guarantee infallible detection and that watermarks embedded in texts do not verify accuracy, identify the text owner, assess the degree of human intervention, or establish human authorship. SynthID is also the technology used for the watermark unveiled by Anthropic in August, a move similar to OpenAI's, aimed at meeting the obligations of the European Union's AI Act.