There is no universal detector. An invisible watermark is only readable by a system that knows what pattern it embedded, so "check this image for watermarks" really means "check this image against each system that might have marked it." The good news is that the list of systems worth checking is short, and every one of them has a free checker.
First: which kind of mark are you looking for?
- A visible watermark — a logo or text overlay. You can see it. Nothing to detect.
- Metadata credentials (C2PA) — a signed manifest in the file container. Readable by any C2PA viewer, and gone the moment a platform re-encodes the file.
- A pixel watermark — a sub-perceptual pattern in the image data itself. Invisible, survives screenshots and re-uploads, and requires the matching detector.
Checking for AI-generation marks
If the question is "was this made by AI," check two things. Google's SynthID Detector portal reads the SynthID watermark, which since May 2026 covers both Google's models and OpenAI's ChatGPT and API images. A C2PA viewer reads Content Credentials, which most major generators now attach. Run both — they fail differently, and agreement between them is much stronger evidence than either alone.
Interpret the result carefully. A detected mark is strong evidence. No detected mark is close to no evidence at all — plenty of generators participate in neither scheme, and marks degrade under heavy editing. We go through SynthID's real accuracy and failure modes here.
Checking for an ownership watermark
This is the other direction: not "did a machine make this" but "does this belong to someone." Creator-side systems embed an identity signature into original photos and video before publishing, so that a copy can be traced back later even after cropping, re-encoding, or AI editing.
To check a file against HALLMARK.AI, upload it to our watermark checker. It runs the detector and returns one of four states rather than a yes/no, because the honest answer is a confidence band:
- Intact — the signature recovered cleanly. The file is a faithful copy of protected content.
- Tampered — the signature is present but degraded. Something was edited, recompressed, or partially regenerated. Bit accuracy degrades smoothly with damage, so the degree of degradation is itself a signal about how much was changed.
- Noisy — a weak residual signal, below the level we are willing to call a match.
- Absent — no signature. Either the file was never marked, or the transformation was severe enough to destroy the mark entirely.
Why a band and not a yes/no
Because a real detector operating on a real file is doing statistics, not lookup. In our published benchmarks the watermark survives generative inpainting up to substantial mask coverage, with bit accuracy declining gradually as more of the frame is repainted rather than failing off a cliff. An unwatermarked file sits near the chance floor, which is what makes the distinction meaningful in the first place. Collapsing that into a binary would throw away the most useful part of the answer — how much the file was altered.
One counterintuitive result worth knowing: running a stolen file through an AI enhancer sometimes raises recovery rather than lowering it, because the enhancer suppresses high-frequency capture noise more than it disturbs the embedded signal. Laundering a file can make it easier to attribute. Full benchmark data is public here, including the transforms where we do worse.
What no checker can do
- Tell you a file is unmarked by every system. Absence of a detection is absence of evidence, nothing more.
- Recover a watermark that was never embedded. If you want your work checkable later, it has to be marked before you publish it — there is no retroactive option.
- Establish who a mark belongs to without the issuing system. The detector and the registry are the same infrastructure.
If you are checking because you think something was stolen
Detection is one input among several. Reverse image search finds copies, a watermark check proves the copy came from your file specifically, and the two together are much stronger than either alone. Our guide to finding stolen photos covers the search side, and the DMCA takedown template covers what to do once you have found it.