ChatGPT watermark remover

Yes, for a growing share of users. OpenAI announced on 5 October 2026 that eligible ChatGPT and Codex text in the EU gets an invisible watermark over the following weeks, and that API customers anywhere can switch it on for select models.

OpenAI calls the technology textGrain and describes it this way: it "adds an invisible statistical signal to the model's word choices." It is not yet a worldwide default, and the announcement does not define which accounts count as eligible.

The reason is legal. Article 50(2) of the EU AI Act, applicable from 2 August 2026, requires providers of generative AI to mark output "in a machine-readable format and detectable as artificially generated." Systems already on the market before that date have until 2 December 2026 to comply.

What about the hidden spaces people found in 2025?

In April 2025 the startup Rumi reported narrow no-break spaces (U+202F) in text from OpenAI's o3 and o4-mini models and called them watermarks. OpenAI told Rumi, as later reported, that they were "simply a quirk of large-scale reinforcement learning." Those characters are easy to strip, and the free tab above does it, but they were never the real watermark.

Sources: OpenAI, EU text provenance, 5 Oct 2026 · EU AI Act Article 50 · Regulation (EU) 2026/1744

Two different things hide in AI text, and each needs its own fix: invisible characters, which a cleaner deletes, and a word-choice signal, which only new wording changes.

A tool that only deletes characters, however it is named, leaves layer 2 exactly where it was. That is why this page has two tabs.

Before each word, a secret key splits the possible next words into a green half and a red half, and the model leans toward green. Anyone holding the key counts green words: ordinary writing lands near half, watermarked text lands far above it.

Try it below. The paragraph was written by leaning toward green under a demo key. Move the slider to swap words for synonyms and watch the detector's score fall.

This is a teaching model of the green-list method from Kirchenbauer et al., 2023, computed live in your browser with a public demo key. textGrain is a newer, entropy-calibrated design and its key is secret; its detector needs only the text and that key (textGrain technical report).

Less than most people expect. In OpenAI's own test on 400-token passages, replacing 10% of the words with synonyms cut detection from about 92% to 66%, and replacing 25% cut it to 17%.

OpenAI evaluation, 400-token passages, published 5 Oct 2026.

Length matters too. At a 1% false positive target, OpenAI's detector found the mark in about 80% of 200-token passages and about 95% of 400-token passages for content such as psychology, and far less often in mathematics. Shorter, more factual text carries a weaker signal from the start.

One caution worth knowing: research on earlier watermark schemes found that long texts can keep a trace after light paraphrasing, because fragments of the original wording leak through. Swapping a few words is the weak version. The rewrite tab changes the wording sentence by sentence, which is the strong one.

Sources: OpenAI, 5 Oct 2026 · Kirchenbauer et al., 2023b · Krishna et al., 2023

Not the public. OpenAI did not release its text watermark detector at launch; access is limited to approved organizations, such as research and academic institutions, reviewed case by case.

So a website that claims to detect the ChatGPT watermark cannot be checking textGrain, because only key holders can. What anyone can check is the hidden-character layer, which is what the free tab shows you.

OpenAI is also clear about what a result means. A watermark "does not identify the user," and "the absence of a detected watermark does not prove human authorship": text can be too short, edited or translated for detection to work.

Sources: OpenAI, 5 Oct 2026 · OpenAI Content Provenance API docs

The EU marking duty is written for providers: Article 50(2) tells companies that offer generative AI to mark what their systems produce. Editing your own text is a separate matter.

The rules that do reach you are local ones. A school, employer, journal or client may have its own policy on AI use and disclosure, and those still apply whatever the text carries. If a policy asks you to say that AI helped, say so. This page explains the technology; it is not legal advice.

Removing hidden characters is free and unlimited, no account needed. Rewriting comes with four free runs after an email sign-in; after that, plans start at $12 a month billed yearly.

An independent computational linguistics studio built and runs it. The rewriting engine learned its style from more than 2,500 essays written by real students, not from AI output, which is why its sentences read like a person's.

The cleaner is plain code that runs on your device. The rewrite sends your text to our server and the engine, returns the result, and keeps no copy after 24 hours. Every figure on this page links to the page it came from.

Not affiliated with OpenAI. ChatGPT and Codex are trademarks of OpenAI. Questions: support@chatgptwatermark.app