OpenAI is rolling out "textGrain," an invisible watermark for ChatGPT and Codex text in the EU, according to reporting in The Neuron. The watermark is opt-in via API worldwide but off by default; the EU rollout is happening over the coming weeks, with the detector kept private for now. Reported detection performance: around 80% on short passages and roughly 95% on longer text at a 1% false-alarm rate, with benchmark scores described as essentially unchanged, and the method to be published. Google's SynthID is cited as the comparable approach. Whatever you think of watermarking as a technology, the regulatory signal is unmistakable: the EU AI Act's machine-readable content obligations are now operational reality, and EU AI Act watermark compliance is a shipping requirement, not a roadmap item.
What happened
Per The Neuron's reporting: OpenAI's textGrain watermark embeds an invisible signal in generated text. In the EU it is rolling out over the coming weeks; through the API it is available worldwide but opt-in and off by default. The detector remains private for now — meaning third parties cannot yet independently verify watermarked text — with detection rates of roughly 80% on short passages versus about 95% on longer ones at a 1% false-positive rate. Benchmark performance is reportedly unaffected, and OpenAI plans to publish the method. The move is widely read as OpenAI getting ahead of the EU AI Act's transparency obligations for AI-generated content.
The founder's compliance checklist
If you ship AI products in Europe, treat provenance as infrastructure. Here is the practical checklist:
- Know your disclosure duties. The EU AI Act requires that AI-generated content be marked in a machine-readable format so it can be identified as artificially generated. Map exactly where your product generates text, images, audio, or video, and confirm each output path carries the required marking.
- Pick a watermarking or provenance approach now. Evaluate textGrain-style watermarking, Google's SynthID, and the open C2PA standard for content credentials. The right choice depends on your modality and whether your customers need third-party verifiability — note that OpenAI's detector is currently private, which limits independent verification today.
- Make it a feature, not a tax. Enterprise buyers in Europe will ask about provenance in procurement. A clear "compliance and provenance" section in your docs — what you watermark, how it is detected, what your false-positive rate is — shortens sales cycles in regulated industries. Compliance that is documented becomes a differentiator against competitors who treat it as an afterthought.
- Mind the false-positive math. A 1% false-alarm rate sounds small until you apply it to millions of outputs or, worse, to accusations against human authors. Design your product's UX around detection uncertainty: never present a watermark hit as proof of AI generation without human review, and document the limitation for customers.
- Build it in early. Retrofitting provenance into a mature pipeline — touching every output path, backfilling metadata, re-certifying with customers — is an order of magnitude more expensive than building it into the generation layer from the start. If you are pre-launch in the EU, this is a week of work now versus a quarter of work later.
What it means for investors
Regulatory readiness is now a diligence item for any AI startup touching the European market. Add it to the checklist alongside data licensing and model evaluations: does the company have a provenance story, is it documented, and has anyone tested it? Startups that built provenance in early will show you a docs page and a detection pipeline; startups that didn't will show you a roadmap slide. Price the difference — the retrofit cost lands on someone, and it should land on the valuation, not on your post-investment runway.
There is also a positive investment signal here. Compliance infrastructure — watermarking, detection, content credentials, audit tooling — is becoming a category with mandated demand. The EU AI Act is effectively creating a market for provenance tooling by regulatory fiat. Early-stage investors should be looking at the picks-and-shovels layer: detection APIs, C2PA tooling, and compliance automation for AI-generated content.
The takeaway
OpenAI watermarking ChatGPT text in the EU is the moment AI-content regulation moved from legal text to shipped product. For founders, the move is simple: build provenance into your generation layer now, document it as a feature, and never treat a detection score as a verdict. For investors, regulatory readiness belongs in diligence, and the compliance-tooling layer is a newly mandated market. The EU AI Act doesn't reward the companies with the best models — it rewards the ones that can prove what their models made.
Angeled.com connects founders with investors.
Read Next
Open Weights Are Not Open Business Models: Mistral Large 4 and the Licensing Trap
Courts Are Writing AI's Rules Now: Three Rulings Founders Should Read
Delhi High Court Orders 24-Hour Takedown of AI Deepfake Videos Targeting Chief Minister
EU Court Adviser Says AI Training Is Legal — but Gemini's Answers May Still Infringe Copyright
AI Regulation After the Midterms: What Founders Should Do Before the January Bills Arrive
DOJ Launches Super Intelligence Task Force to Police AI Misuse



