AI Watermark Detector

No. An AI watermark is often invisible. Some signals are embedded in the image itself, like SynthID, while others live in metadata, like C2PA Content Credentials or IPTC disclosure fields.

C2PA and IPTC metadata are easier to strip than SynthID. Screenshots often remove metadata fields, while SynthID is embedded in the image. Repeated heavy edits or severe recompression can still make a check inconclusive.

Slop or Not still checks the pixels and reports a probability when the file has no readable provenance. A missing watermark is not proof that a person made the image.

Yes, for the kind that lives in characters. The Text tab flags invisible Unicode, smart punctuation, and lookalike letters, and shows each one in place. Statistical marks are a separate matter: Anthropic has announced a public API for Claude's SynthID-Text watermark but has not released that API, so only Anthropic can read that mark. So the same tab reads the writing itself, which is why a missing mark does not blind it: 99.5% of AI text caught at a 5% false-positive rate on RAID. Anthropic has not documented the rollout model by model, and a missing signal proves nothing.

Anthropic says future Claude models will generate text with a version of SynthID-Text and older models will be updated over the coming months, but it has not documented the rollout model by model. Its public text-watermark API is announced but not released. Claude images use C2PA Content Credentials, not image SynthID, and a missing signal proves nothing.

You can remove the mechanical part, and one click here does it: invisible Unicode, smart punctuation, and lookalike letters all go, and no word is rewritten. What no cleaner can take out is a statistical mark such as SynthID-Text, which is woven into the word choices themselves, or the writing style, which is the evidence a detector actually reads. Anything sold as an AI watermark remover is doing the first job and charging for the second.

Copy the passage out of Word, paste it into the Text tab above, then paste the cleaned version back. Word's formatting-marks button stops at paragraph marks, tabs, and ordinary spaces, so zero-width characters and lookalike letters stay invisible to it. If the document also holds a generated picture, save that picture out and run it through the Photo tab, where C2PA Content Credentials and IPTC fields sometimes survive the trip.

Copy the text out of Google Docs and paste it into the Text tab above, then paste the cleaned version back over the original. Docs has no view for zero-width characters or lookalike letters, and its own autocorrect curls quotes and lengthens hyphens as you type, so a hit tells you a typesetting tool was involved rather than which one. Images are a separate question: download one out of the doc and check it in the Photo tab, where provenance metadata sometimes survives and often does not.

Yes, if the text is selectable: copy the passage out and paste it in, because the scan reads characters rather than page layout. If the PDF holds an image of a page, export or screenshot that image and use the Photo tab instead, where C2PA Content Credentials and IPTC fields sometimes survive the trip. A scanned page carries no character data at all until you run OCR on it, and OCR output brings its own transcription errors rather than the original marks.

Not in the sense most people mean. OpenAI developed a statistical text watermark and discussed it publicly, but never deployed it or released a checker for it, so there is no ChatGPT mark anyone outside OpenAI can decode. What ChatGPT output does carry is mechanical: typographic characters the chat window bakes into whatever you copy out, including the narrow no-break space at U+202F. The ChatGPT watermark detector page goes through those in detail.

Metadata comes off easily. A screenshot, an export, or a re-encoding upload can drop C2PA Content Credentials and IPTC fields, and tools exist that strip them on purpose. SynthID is embedded in the image itself, which makes it far harder to remove than a metadata field. None of it changes the pixels: a file with its provenance stripped is still the same picture, and the visual check still runs on it.

Yes, on Apple devices, and for both mediums in the same app. The iPhone and Mac apps run the character scan, the text check, and the image check on-device with no network involved, and they add spelling conversion between US and UK English that the online scan does not do. In the browser, the character scan runs on Numen's server in memory and is discarded with the response. The AI text and image checks travel to Numen's private Mac server, away from third-party AI providers, and are deleted after processing.

Yes. Checking a file or a passage you are responsible for is exactly what the page is for, whether that is a classroom submission, a freelance delivery, or a contributed pull request. On the Mac, Pro also ships a slop command and an MCP server, so Claude, Codex, Cursor, and your own scripts can scan, clean, and re-check text locally.

Yes, in the browser. The hidden-character scan sends the passage to Numen's server, where it is processed in memory and discarded with the response; nothing is stored. The AI text and image checks travel to Numen's private Mac server, away from third-party AI providers, and are deleted after processing. The iPhone and Mac apps keep everything on-device.