How to read a negative
What a missing ChatGPT watermark actually means
A missing watermark is not a finding that a person wrote the text. It is a finding that a particular detector did not see a particular signal, or that the detector was never run. OpenAI lists the ordinary reasons. Most arguments about “clean” drafts skip them.
The source paragraph is in OpenAI’s provenance help: if no signal is detected, the tool did not find OpenAI-issued provenance signals. The content could still have been generated or exported by OpenAI. The same caution applies, more strongly, when the tool in front of you is not OpenAI’s text detector at all.
Which negative you actually have
“No watermark” is said about at least four different tests. Write down which one returned the negative before you quote it.
- textGrain. Someone with OpenAI’s text detector scored the passage and it fell under the threshold. This is the only negative that speaks to textGrain, and only for that passage and that threshold.
- openai.com/verify on a file. No supported SynthID or trusted C2PA signal in that image or audio file. This says nothing about the sentences in a Doc you did not upload.
- A character scan. No invisible Unicode. The ChatGPT watermark detector on this site reports that kind of negative. It is about codepoints.
- A style detector. A low AI-writing score. That is a classifier going quiet. It is not a watermark result. The split is watermark versus an AI detector.
People stack these. A low GPTZero score plus a zero character count gets summarized as “no ChatGPT watermark.” Neither input was a textGrain test. The summary is false even if both tools worked.
OpenAI’s own list
For a real provenance check that comes back empty, OpenAI’s reasons include:
- The content was created before the signal existed.
- It came from an unsupported product, model, export path, or file type.
- Metadata was stripped during upload, download, editing, conversion, or sharing. This one is about files, not about textGrain.
- A watermark was degraded by compression, cropping, noise, edits, or format conversion. Again, mostly media. For text, the matching degradation is a substantial rewrite.
- The specimen is too short, or for audio, too heavily modified. For text, OpenAI adds the short-answer and code limits, and the case where the model had to be factual or to repeat user-supplied text.
The EU code of practice, as OpenAI describes it, does not require a watermark below about 200 tokens or on code. A missing mark on a caption or a function is the policy working as written, not a loophole you discovered. Long prose is where a miss is informative, and even there the miss is “not detected,” not “human.”
Language and entropy sit under the same heading. OpenAI published detection rates that differ across EU languages at a fixed false-positive rate, and said factual text leaves less room to watermark. A Romanian paragraph and a list of dates are not the same evidence as a long Spanish essay, even before anyone edits.
Rewriting is the reason that gets misused. Substantial paraphrasing or translation can make detection less reliable. That sentence is a limitation of the scheme. It is not an instruction, and a miss after a rewrite does not become proof. The boundary between a character clean, a manifest strip, and a rewrite is can you remove a ChatGPT text watermark.
The detector that was never run
Most “no watermark” claims in 2026 have a simpler cause: the speaker does not have the detector. OpenAI’s public verify page does not accept essays. API customers who opt in do not receive detection with the switch. A free website that never mentions a key has not run textGrain, including this one.
Coverage makes the gap wider. Eligible EU ChatGPT text is the marked consumer path OpenAI describes. API text is unmarked unless the setting is on. A draft from a default API project is supposed to miss. Calling that miss a human result inverts the setting. The map is does ChatGPT watermark text in 2026.
If your organization later gets access, keep the original wording. A check on a version that an editor already rewrote is a check on the rewrite. Record the date, the tool, and the threshold. A PDF that only says “no watermark” without those three is not a record.
Other labs and other files
A textGrain miss says nothing about Claude. Anthropic’s mark is a SynthID-Text variant with its own key, and Anthropic describes it as on for covered models globally. A Gemini paragraph is a third key. The comparison is Claude, Gemini, and ChatGPT. Clearing one lab is not a clearance.
Files fail differently from paragraphs. A JPEG can lose C2PA when someone screenshots it, while SynthID remains in the pixels, or the reverse can be what a given tool is able to see. This site reporting no C2PA means the container segments it looks for were absent. It does not mean openai.com/verify would agree about SynthID, and it does not mean the picture is a photograph. Read what a C2PA watermark is before you promote a clean metadata result into an origin story.
Hidden characters fail in the boring direction. Most model output has none. A zero count is the usual result of the hidden-character check. Publishing “we scanned it and there is no watermark” after that check misnames the tool.
What a positive result does not add
The matching mistake is to over-read a hit. OpenAI says a detected signal means the content contains a supported provenance signal associated with OpenAI. It does not confirm accuracy, ownership, legal responsibility, or who created it. It does not show whether the model wrote some or all of the text. A watermark on one section of a report is compatible with a human author everywhere else.
Provenance signals also do not include the prompt or the user. A hit is not an account takeover story and not a plagiarism percentage. Turnitin’s percentage is still a different instrument, covered in can Turnitin detect ChatGPT watermarks.
Decisions a negative can still support
A negative is useful when the decision matches the test.
- No invisible characters: the paste will not break a diff for that reason. You may still clean other files.
- No C2PA in this JPEG: a viewer that only reads manifests will not show this file’s credential. Pixels may still carry a signal.
- textGrain under threshold, on the original, from an approved detector: this specimen did not meet that threshold. Ask whether the path was even in scope before you say more.
- Low style score: that classifier was quiet. Record the vendor and the date, because the model will change.
A decision the negative cannot support is “tell the reader a model was not used.” If disclosure matters, disclosure is a sentence. The absence of a mark is not that sentence. Editors who need a sequence that stays inside what the tests mean can follow how to check a blog writer and the tool list in free checkers compared.
Run a listed-carrier check and read the negative narrowly →The mechanism, if you need to explain why a real mark can be absent from text that a model did write, is how the OpenAI text watermark works.
Four cases that look like proof
The contractor. A client says the draft “came back clean,” meaning a free classifier sat in the green band and this site found no hidden characters. The writer later says they used ChatGPT outside the EU path. Both tool results can be accurate. Neither one tested textGrain, and the path they named is one OpenAI does not describe as marked by default. The clean report proved the tests that ran. The disclosure proved the origin. Believe the disclosure.
The appeal. A student submits a PDF export. openai.com/verify finds no image signal because the upload is not a supported image. The department treats that dialog as “OpenAI says it is not theirs.” The dialog said the file did not contain the signals that page looks for. The sentences were never the specimen. Export the text and, if the institution has text-detector access, score the text. Until then, the verify page has not spoken.
The translation. A marked English answer is translated for a local edition. The local detector, when someone eventually runs it, misses. OpenAI listed translation among the changes that can drop the pattern. The translation can still be model output, or a human translator’s work, or both. The miss does not pick among those. A byline does.
The screenshot. An editor photographs a chat on a phone and sends the JPEG through a metadata cleaner. The JPEG has no C2PA because a screenshot often never had one, and it has no text watermark because a photograph of letters is not the token string. The cleaner reports success. Success at what? At removing carriers that were not the words. The words in the picture cannot be scored for textGrain until someone types them out again, and that typed copy is a new object.
How to write the negative down
Use a sentence that names the test. “No listed invisible characters in file draft-v3.txt on 5 October 2026, checked on this site.” “No C2PA manifest in header.png after export from the CMS; original export not retained.” “textGrain detector, institution access, under threshold, original paste, threshold as printed on the report.” Each of those can be true. The forbidden compression is “no watermark, so human.”
If you do not have the third sentence, do not imply it. Point readers who want the mechanism at how the mark works and readers who want the coverage holes at where ChatGPT marks text. A negative is a narrow tool. Written narrowly, it is still worth having. Written as a clearance, it will be quoted in a place the original checker never saw, which is how a character scan becomes a press line about authorship.
The same discipline applies when the negative is on an image. “Verify found nothing” belongs next to the filename and the fact that the page accepts images and audio. It does not travel to the article the picture illustrated. Store the report PDF with the asset, not in a shared drive folder called “cleared.”
If you only remember one substitution, make it this: replace “no watermark” with the name of the test that was quiet. The rest of the argument can be rebuilt from that noun. It cannot be rebuilt from the adjective “clean.” Store the report beside the file hash or the Doc revision id so a later copy cannot inherit a clearance it did not earn. If the only copy left is a screenshot, say that too: a photograph of a report is not the report, and it cannot be scored again. Retype the passage if you need a text check, and label that retyped copy as a new specimen.
FAQ
Does no watermark prove a human wrote it?
No. The text may predate the rollout, come from an unmarked path, be too short, be code, be tightly factual, or have been rewritten. Or the text detector was never run.
Does a clean result here mean no ChatGPT watermark?
No. A clean result means no listed hidden characters or, for a file, no listed metadata. textGrain is not in that list.
Does a hit mean the person did nothing?
No. A hit is evidence of model processing. It does not measure the human share, the user, or the prompt.
Why is a short answer a bad example?
There are not enough token choices to carry a reliable pattern. OpenAI points at a floor around 200 tokens in the EU code of practice, and at code as a separate exclusion.