Background reading
What a verification tool can actually see.
Since transparency obligations under the EU AI Act became enforceable in August 2026, a lot of tooling has appeared in this space, described in a lot of different ways. This page is an attempt to set out plainly what can be read from a file today, what we would like to add, and where the limits currently sit.
This is not legal advice and not a compliance guide. It describes what verification tooling can and cannot observe, which is a technical question. Whether any obligation applies to you, and what would satisfy it, is a question for a lawyer — and claims circulating about this regulation are worth checking against the text itself rather than against any vendor's summary, including ours.
What you can check today
All of this is free and runs in your browser. Files are not uploaded, which some people find matters when material is unpublished or under embargo.
Content credentials on incoming files
Where a C2PA manifest is present, its signature is validated cryptographically. You see the signing certificate, the timestamp, the recorded edit history, and any AI involvement the producer declared.
Audit a fileHow AI was involved, where it is declared
IPTC digital source types are shown as recorded, which separates fully generated material from composites and from content that was merely enhanced. This only works when a producer chose to declare it.
See the signal matrixA whole set of files at once
Every file in one table with its status, signer, declared origin, and hash — plus totals across the set, and any files that are byte-for-byte identical to each other.
Batch auditA record of what you checked
A timestamped report where each row carries its SHA-256, so it can be matched back against the original files later, by you or by somebody else.
Audit reportWhere your own workflow loses provenance
Drop the same asset as it appears at each stage of your process and see which step stops carrying credentials, metadata, or location data.
Pipeline auditWhat routine processing does to a file
Measured results for resizing, recompression, screenshotting, and cropping, plus a test you can run on your own files.
Survival testThings we are exploring
Ideas we think are worth building, listed so you can see the direction. These are not commitments and none of them has a date — most would need somewhere to keep a record between sessions, which the tools deliberately do not have today.
Records that persist
Audit results that outlive the browser tab, grouped by submission or contributor, so patterns across repeat material could become visible.
Countersigned reports
A version of the audit report stored at a stable address and signed by us, so a third party could confirm a record existed on a given date without receiving your files.
Your own rules
Thresholds you define — flag this, warn on that — checked automatically as files are audited, rather than read off the table by eye.
Prompting for disclosure
Where an audit finds a declared AI origin, prompting the author for more detail at submission time rather than relying on them to volunteer it.
Patterns across contributors
How often material from a given source arrives with credentials intact, calculated from your own audit history rather than assumed.
Testing a tool's output
Generating sample output from an AI tool and checking whether it actually carries machine-readable marking, rather than taking the datasheet's word for it.
If one of these would make the difference for you, it is genuinely useful to know — the early-access list asks what you need for exactly this reason.
Where the limits are
Capabilities we have seen described in this space that we cannot offer, with our understanding of why. Some are limits of the technology, some are a different field wearing this one's name. If you find any of this out of date, we would like to hear about it.
Reading an AI text watermark
Anthropic has begun marking Claude's output but has not yet published the detection documentation. Google's SynthID detector runs through a waitlisted portal with no public API. OpenAI built text watermarking and did not deploy it. As far as we can establish, that leaves no route for an independent tool to read these marks.
This may change. Anthropic has indicated documentation and an API are coming. If a real detector becomes available, we would add it.
Carrying a signature through a file conversion
A C2PA signature covers a hash of the file's bytes. Converting a document to another format produces different bytes, so the original signature no longer corresponds to the file. This is how the mechanism works rather than a shortcoming of any particular software.
A different approach exists: signing the converted output yourself and recording the original as an ingredient. That requires certificates and signing infrastructure, which we do not currently have.
Telling you whether text was written by AI
We are not aware of a detector that does this reliably. Published evaluations of the available ones report false positives, and the burden appears to fall disproportionately on non-native speakers.
Not something we intend to add. A percentage here would be a guess presented as a measurement.
Judging what counts as a matter of public interest
That is a judgement about subject matter, and no provenance signal carries it. A tool that guessed would be substituting its own inference for an editorial decision.
Flagging candidates for a person to consider is possible. The judgement itself is not ours to make.
Adding labels to published content
Applying a visible disclaimer to a published article happens inside the system where the content lives. A verification tool reads files; it does not sit in the publishing path.
Identifying which items a person may want to label is within reach. Applying it belongs to your CMS.
Inspecting software for marking behaviour
Whether an application embeds watermarks is not observable by examining the application. It is only observable in what the application produces.
Framed as testing the output instead, this is more tractable — see above.
Finding manipulation inside an image
Spliced blots, duplicated panels, and altered gels are found by analysing pixel content. These tools read a file's provenance record instead. Our duplicate detection compares hashes, which finds identical files and nothing more subtle.
A different field with dedicated products. Worth using alongside this rather than expecting it here.
A reasonable place to start
If you only look at one thing, we would suggest checking whether your own publishing process is preserving provenance. It is quick, it needs no account, and in our experience the answer surprises people — credentials attached upstream frequently do not survive the trip to a published page.