How to Check If a File or Text Has an AI Watermark: Free Methods That Actually Work
Updated September 2026
Before removing anything, look. Most AI provenance is readable by anyone in seconds: a signed manifest in the image, a generator name in the document properties, a pattern of invisible characters in the text. Some of it is readable only by the vendor. Here is how to check each layer and how to know which one you are looking at.
Quick answer
What you can check yourself, for free:
- 1Images and video: C2PA manifests, XMP and IPTC AI tags, EXIF software fields, embedded prompts. Scan with StripShot.
- 2Documents and PDFs: creator, application, producer, session IDs, tool names. Scan with StripShot.
- 3Text: hidden Unicode carriers, look-alike letters, odd spaces, shown in place. Scan with StripShot.
- 4What you cannot check: SynthID in pixels or audio (use Google's own detector portal), and statistical text marks (vendor only).
Layer by layer
| Content | Signal | How to check | Who can read it |
|---|---|---|---|
| Image | C2PA manifest (APP11, caBX, uuid box) | StripShot scan, or the Content Credentials verify site | Anyone |
| Image | XMP DigitalSourceType, EXIF Software, PNG tEXt prompt | StripShot scan | Anyone |
| Image | SynthID in pixels | Google's SynthID detector portal | Google and partners |
| Video | C2PA and XMP boxes, udta software fields | StripShot scan | Anyone |
| Video | SynthID or Video Seal in frames | Vendor detector | Vendor |
| Audio | ID3, RIFF and FLAC tags naming the tool | StripShot scan | Anyone |
| Audio | Waveform watermark | Vendor detector | Vendor |
| Document | Creator, application, producer, rsid, comments | StripShot scan | Anyone |
| Text | Hidden Unicode, look-alike letters | StripShot scan with highlight view | Anyone |
| Text | Statistical token watermark | Vendor detector with key | Vendor |
Reading a scan result
- Made with X. A tool name found in the file. Definitive when present: nobody writes OpenAI or Adobe Firefly into a file by accident.
- C2PA manifest. Signed provenance. Definitive that a C2PA-aware tool produced or edited the file, which includes Photoshop edits of real photos.
- Generation prompt embedded. Stable Diffusion and ComfyUI store the full prompt, seed and model in the PNG. Definitive, and a privacy problem in its own right.
- Invisible characters. Definitive that the characters are there; not proof of where they came from. A count of three is copy residue. A regular pattern of dozens is a fingerprint.
- Writing tells. A style read: phrase markers, sentence-length variation, lexical diversity, em dash density. Useful for editing. Not evidence.
No signal is not proof of anything
A clean scan means no metadata and no hidden characters. It does not mean the content was not AI-made. Screenshots, re-encodes, platform uploads and plain copy-paste all strip metadata without anyone intending it, and the vendor-only marks are invisible to every scan on this page. The honest reading of a clean result is that nothing readable remains, which is all any tool can tell you.
After you check
If the scan shows metadata or hidden characters and you want them gone, the same tool strips them and re-scans the result. If it shows a Google generator, the page tells you SynthID stays. If you are checking someone else's content to decide whether to trust it, treat the readable layer as strong evidence when present and as no evidence when absent, and treat any service that sells a verdict on the unreadable layer with suspicion.
Free scanner
Scan a file or paste text to see every provenance signal
Runs in your browser. Nothing is uploaded.
Drop a file here, or click to choose
Images, video, audio, DOCX, PDF, text
Options
Need unlimited strips?
Scans are free. Pro is unlimited.
Images, video, audio, documents, PDF and text in one tool. Pro adds unlimited strips, whole-folder batches and zip download with a JSON report.
Frequently asked questions
How can I tell if an image has an AI watermark?
Scan its metadata. A C2PA manifest, an XMP DigitalSourceType tag or a generator name in EXIF is a definitive AI signal, and StripShot's checker shows them in seconds. A pixel watermark such as SynthID cannot be seen this way; only Google's detector reads it, and Google offers a SynthID detector portal for that.
How can I tell if text has an AI watermark?
You can check for character-level carriers by scanning for hidden Unicode, which StripShot does and shows in place. You cannot check for a statistical watermark from outside the vendor: SynthID-Text needs Google's key and Anthropic's mark needs Anthropic's detector, which it has described as forthcoming. Any public tool that claims to detect those is reading style, not the mark.
Are AI detectors the same as watermark detectors?
No. A watermark detector reads a mark the model planted, with a key, and is reliable. An AI detector guesses from style and statistics with no key, and is known to produce false positives on human writing. StripShot's writing-tells panel is the second kind, deliberately labeled as a style read, not a verdict.
Can I check a Word document or PDF?
Yes. StripShot lists the author, last-modified-by name, application, company, editing time, revision IDs and any tool name in the properties, plus hidden characters in the text. A generated document often names the generator in the creator field.
Is the file uploaded when I check it?
No. Scanning runs in your browser and is free and unlimited.