OpenAI has introduced an invisible watermark for text. API customers can now opt in for selected models, while eligible ChatGPT and Codex output in the European Union will begin receiving the watermark over the coming weeks. The change is part of OpenAI's response to the EU AI Act, which requires providers of generative AI systems to make generated content detectable in a machine-readable form.
This sounds like a neat technical answer to a messy problem. A document carries a hidden signal, a detector finds it, and the reader knows AI was involved. OpenAI's own evidence shows why businesses should not treat it that way. The watermark can be weakened by normal editing, it is harder to detect in short or constrained writing, and a positive result says nothing about whether the content is accurate or responsibly published.
For any business using AI to prepare customer guidance, reports, public information or regulated communications, the practical consequence is clear. Provenance is becoming part of the technology, but accountability still has to come from the publishing process around it.
What OpenAI Has Actually Released
OpenAI calls its system textGrain. It changes the statistical pattern of a model's word choices in a way intended to be invisible to the reader. A separate detector looks for that pattern and reports whether it can identify an OpenAI watermark.
According to the OpenAI announcement, watermarking is off by default for API customers and available as an opt-in feature on selected models. OpenAI plans to apply it to eligible ChatGPT and Codex text output in the EU across all plans. Its detector is not being opened to the general public at launch. Access will initially be limited to approved researchers and expert organisations because the result can be misinterpreted.
The published tests show a useful but imperfect signal. At a target false-positive rate of 1%, OpenAI says its detector identified the watermark in about 80% of 200-token passages and about 95% of 400-token passages for one type of general writing. Detection was substantially lower for mathematical content, where the model has less freedom to vary its wording.
Ordinary editing also changes the result. In one test using 400-token passages, replacing 10% of the words with synonyms reduced detection from about 92% to 66%. Replacing a quarter of the words reduced it to 17%. Those figures do not make the technology pointless. They do show that it is evidence with limitations, not a dependable certificate attached to every piece of AI-assisted writing.
A Watermark Cannot Approve a Document
OpenAI is unusually direct about what the signal does not establish. It does not measure how much a person contributed, identify the user, prove ownership, decide whether disclosure was legally required or verify any factual claim. An undetected watermark does not prove that a person wrote the text either. The material may have been shortened, edited, translated, produced with an unsupported model or generated by another provider.
That distinction matters in normal business work. Consider a housing association using an AI assistant to prepare an update about repairs, eligibility and temporary arrangements. The draft may contain a valid watermark, yet still quote an old policy date or omit an exception that affects vulnerable residents. A communications manager could correct the facts and rewrite several passages, weakening the watermark while making the document substantially safer and more useful.
The opposite is also possible. A strongly detected passage may have moved from a model into a document without anybody checking the source material. The provenance signal would be working exactly as designed while the organisation's review process had failed.
Treating detection as an accuracy check would therefore create false confidence. Treating the absence of detection as evidence of human authorship could lead to an unfair accusation. A watermark should be one piece of technical metadata in a wider record of how important content was produced and approved.
The EU Rules Reach Beyond Technology Providers
The legal duties vary according to the organisation's role and the type of content. Article 50 of the EU AI Act requires providers of systems that generate synthetic text, audio, images or video to mark outputs in a machine-readable and detectable form. It also places disclosure duties on organisations deploying certain systems, including where AI-generated or manipulated text is published to inform the public on matters of public interest without human review or editorial responsibility.
The European Commission's Article 50 guidance says these transparency obligations have applied since 2 August 2026, with a limited transition to 2 December for the marking duty on some systems already on the market. The Commission also states that enforcement can involve fines of up to €15 million or 3% of worldwide annual turnover, with proportionality for smaller organisations.
A UK company should not assume that leaving the EU removed every practical concern. A business may sell into EU markets, operate an EU entity, publish information aimed at people there or buy software whose provider changes its product globally or regionally to meet the rules. The precise legal position depends on the activity and should be checked with appropriate advice. Operationally, it is sensible to know which AI tools create public content, which markets receive it and who takes editorial responsibility.
Build the Evidence Into the Workflow
The useful response is not a blanket label on every internal draft. It is a proportionate publishing route based on consequence. A quick summary used by one colleague does not need the same control as public health guidance, financial commentary or a customer notice that could change somebody's decision.
For higher-consequence material, keep the original sources beside the draft and require the reviewer to check claims against them. Record the model or service used, the date, the meaningful edits and the person who accepted editorial responsibility. If the chosen API supports watermarking and it fits the use case, enable it, but preserve this audit trail separately because the text itself may be reformatted, translated or shortened later.
The housing association example can be handled without turning publication into a compliance exercise that nobody follows. The AI assistant prepares a draft only from the current repairs policy, approved service notices and a structured case summary. It links each important statement to a source. The communications manager checks dates, eligibility wording and contact routes, then records approval in the content system. The published notice carries any required disclosure, while the business retains the draft, sources, review and final version.
This design also deals with channels that alter content. A website article may be condensed for email, social media and a customer portal. The watermark may survive in one version and not another. The approval record can show that all four versions came from the same checked source material and identify who was responsible for the final wording.
Use Provenance as a Signal, Not a Verdict
OpenAI's release is important because machine-readable provenance is moving from policy discussion into ordinary AI products. Businesses building with the API now have a choice to make, and teams using ChatGPT or Codex in the EU will start encountering watermarked output by default.
The sensible move is to review the content journey rather than add a detector at the end. Find where AI-generated text enters public or consequential work, decide which material needs disclosure, preserve its sources and give a named person responsibility for approval. Test how editing, translation and channel changes affect any provenance signal before depending on it.
An invisible watermark may help establish that a tool was involved. It cannot establish that the business did the job properly. That proof still comes from a clear, usable process that connects source material, human judgement and the final publication.
Birdcage Tech helps businesses build practical AI workflows with the controls, evidence and human review needed for real operational use.
