OpenAI watermarks: ChatGPT, Codex and the API

OpenAI calls its text watermark textGrain. The October 2026 rollout distinguishes eligible ChatGPT and Codex output in the EU from optional watermarking on supported API models.

Does ChatGPT watermark text?

On 5 October 2026, OpenAI announced phased text watermarking for eligible ChatGPT and Codex output in the EU. It also announced a global API opt-in for select models, with watermarking off by default in the API. Product, region, model, and generation date all matter. OpenAI: EU text provenance rollout.

This is why an old answer to “Does ChatGPT watermark text?” can be misleading today. Check a dated provider announcement and the current support documentation before drawing a conclusion about a particular passage.

How textGrain works

textGrain creates a statistical pattern in word and token choices. It does not insert invisible spaces, hidden characters, or special punctuation. The watermark belongs to the wording rather than a separate label you can select and delete. OpenAI Help: provenance signals.

AI-sounding phrases are a separate editorial issue. Removing filler may make a paragraph clearer, but readability is not a watermark test. Likewise, copying into a plain-text editor does not establish anything about a statistical pattern in the words.

Can the watermark show how much you wrote?

OpenAI says a watermark does not measure human contribution, establish ownership, identify the user, or verify accuracy. Short or constrained passages are harder to detect, and editing can weaken the signal. Absence of a detected mark is not proof of human authorship. OpenAI: EU text provenance rollout.

A newsletter may begin as your field notes, become a model-assisted draft, and go through another round of your edits. To explain that process, keep the versions. A label about provider involvement cannot replace the specific example, opinion, or correction a reader may want to inspect.

Is there a public OpenAI text-watermark checker?

The public Content Provenance API documents image and audio checks. Text verification requires separate approval; it is not a general text-scanning endpoint available to every developer. The public verification tool’s media support should not be advertised as universal text detection. OpenAI: Content Provenance API documentation.

Overwride’s checker has a narrower, visible purpose: it looks up matching Overwride records. A no-match result from that checker says nothing about whether OpenAI generated a passage.

What a rewrite and record can show

Overwride keeps the draft it receives next to the rewrite it returns. Share the record when you want readers to compare those versions. The record supports a concrete explanation of your workflow without assigning a “human-written” percentage.

Substantial paraphrasing may make a watermark harder to detect, according to OpenAI. Overwride does not verify that outcome for an individual rewrite or guarantee that output is watermark-free. OpenAI Help: provenance signals. Read the comparison with OpenAI watermarking to choose the evidence that fits your purpose.