Provenance Week: OpenAI Watermarks ChatGPT Text in the EU, Google Opens SynthID Detector to Everyone -- and Both Say a Missing Mark Proves Nothing

Three provenance moves landed in four days. They are worth reading together, because each covers what the others leave out.

OpenAI: text watermarking, EU first (October 5)

OpenAI's post ties the change to European law: "The EU AI Act requires generative AI providers to make generated text identifiable in a machine-readable way." What it is doing:

• ChatGPT and Codex, EU only. "Over the coming weeks, we will add an invisible watermark to eligible ChatGPT and Codex text output in the European Union," across all plans. "We are not making text watermarking a global default at launch."
• API, opt-in worldwide. "API customers globally will be able to opt in to text watermarking for select models. Text watermarking will remain off by default in the API." OpenAI says it is working with cloud partners to offer the same for its models served through their platforms.
• How it works. textGrain "adds an invisible statistical signal to the model's word choices," and a detector looks for it. OpenAI plans to open-source the technology, and says it saw no meaningful benchmark differences for its Astra model with and without watermarking.
• Restricted detector. Access "will initially be limited to approved researchers and expert organizations." The tool reports whether a watermark is present "without identifying the user or revealing their prompts or conversations."

The published numbers explain the caution. At a 1% false-positive target, the detector found watermarks in "about 80% of 200-token passages, compared with about 95% of 400-token passages" for psychology content, and "substantially lower" rates for mathematics. Swapping "10% of words with synonyms reduced detection from about 92% to 66%. Replacing 25% of words reduced it to 17%." OpenAI is unusually direct about what a hit means: "A watermark does not measure human contribution," it "does not establish ownership or responsibility," and "The absence of a detected watermark does not prove human authorship."

Google: a public detector for media (October 7)

Google DeepMind said it has "watermarked over 180 billion images and videos, along with 240,000 years of audio content" since SynthID launched in 2023, and is now opening SynthID Detector to everyone, "available globally in English starting today." Anyone can check whether "an image, video, or audio file was made with AI from Google or our partners, including OpenAI, NVIDIA, Kakao, and soon, Apple." Built-in verification in Search, the Gemini app and Chrome already handles "over 1 million requests daily." MacRumors notes the site requires sign-in, "cannot distinguish between AI-created and AI-edited media," and only detects media that uses SynthID.

India: labelling by advisory (October 8)

MeitY's October 8 advisory asks platforms to label or contextualise "manipulated, synthetically generated or misleading content." MediaNama points out that India's notified rules on synthetic content apply only to audio, visual or audio-visual material, "so text-only content falls outside" them (see this digest's separate entry on the advisory).

Why it matters

Put side by side, the coverage is patchy. OpenAI's text mark is on by default only in the EU, and its detector is closed. Google's detector is open, but for media, not text, and only for SynthID partners. India's rule targets audio and video. A team that has to show content is AI-generated, or prove it is not, still has no single check that answers both questions.

Analysis: treat provenance signals as one input, never a verdict. A positive result shows a model was involved, not how much; a negative proves nothing. Teams using OpenAI's API for EU users should decide now whether to opt in, and record why. Schools and employers should not discipline anyone on a detector result alone. OpenAI's own limits make that case.