Stop checking ChatGPT essays for watermarks
The first file in the queue is almost always a document. A cover letter. A freelance caption. A PDF of a blog draft someone swears came from ChatGPT. You paste it into a watermark tool and wait for the mark.
There is no mark to read. ChatGPT text is not watermarked.
The watermark is on the still and the clip
OpenAI applies Google DeepMind SynthID to ChatGPT images and audio. That is the media that can carry a signal. The essay cannot. Every workflow that starts by pasting prose into a "ChatGPT watermark checker" is running the wrong job and then treating a null as a result.
The workaround is older than the watermark. Photo editors squint for a too-smooth product shot. Brand-safety reviewers reverse-search a still. Marketplace QA folders fill up with "looks generated" notes that never become a field on the asset. Trust-and-safety teams listen to a voice note twice. None of that gives you a row you can keep next to the file.
Stop checking ChatGPT essays for watermarks. Check the export: the PNG, the JPEG, the WebP, the GIF, the MP3, the WAV. That is where the mark can live.
What I actually run
I built the OpenAI ChatGPT Watermark Checker API for that media. You upload the still or the clip. Each file becomes one dataset row: whether it looks AI generated, whether a watermark showed up, confidence, a likely source, and billed token units. There is no ChatGPT login. There is no browser.
I reject PDF, Word, and plain text on purpose. There is no SynthID-Text signal on ChatGPT writing, and accepting those files would only reprint the mistake this article is about.
Three steps: upload the image or the audio. Start the run. Read the AI-generated flag next to the watermark fields, not as a red card on a person.
Prefer the original export. A heavily recompressed screenshot is a worse sample than the file ChatGPT actually wrote to disk.
Who this is for
If you are a newsroom photo editor who needs a provenance field before a still ships, this is for you. If you review UGC photos and voice notes, marketplace product shots, freelance illustrations, or ad creatives, this is for you. If you are labeling a corpus of stills and clips, or attaching an AI-generated flag to produced media, this is for you.
If your queue is essays and take-homes, this API will refuse them. That is the correct refusal. For Gemini writing, use the Gemini SynthID-Text checker. For Claude-shaped copy, use the Claude writing classifier.
Do not treat the score as legal proof of authorship. A watermark estimates whether ChatGPT media was involved. It does not identify who prompted it. Files over fifteen megabytes are rejected. A missing mark on a mashed-down screenshot is not the same fact as a missing mark on the original export.
Input, pricing, and the rest of the caveats live on the Apify Store README for this checker.
If the next job is disclosing the image rather than detecting it, that is a different API: I stamp official EU AI Act labels into the pixels so the mark survives download.
The cost of the old way is a null you file as a no
The old way starts with the essay, gets nothing, and calls the still fine because the text tool had nothing to say. Those are different files. The watermark, when it exists, is on the media.
If you have the original still or clip, upload it. Store the row next to the asset. Leave the documents to the checkers that actually read writing.
The other posts on rainminer cover writing when the source is Gemini or Claude, not ChatGPT text.