Was it AI? What watermarks and metadata can tell you
An image can carry a record of how it was made. Text can contain a statistical watermark. Neither gives you a universal test for AI authorship. Here is how to read the evidence without asking it to prove too much.
Start with the original file
Ask for the original download or camera file if you can. A screenshot is a new image, and exports or social platforms may remove metadata. Inspecting that copy tells you about the copy; it may no longer contain the history of the original.
A checklist to reuse
- Keep the original file and record where it came from.
- Check the evidence type, its source and its validation status separately.
- Distinguish AI generation from AI editing or a composite containing AI elements.
- Treat an empty result as inconclusive, not as proof of human authorship.
Metadata records claims
EXIF, XMP and IPTC fields can record a camera, editing application or digital source type. An explicit source declaration can describe AI generation or the inclusion of AI elements. Ordinary metadata is editable; a camera name or Photoshop entry alone does not establish whether AI was used.
The inspector shows selected source and software fields, rather than every piece of personal metadata. It recognises standard IPTC AI source declarations. It does not infer authorship from filenames, writing style, prompts or a familiar software name.
A file names an editing application but contains no AI declaration. The useful result is “recorded software”; there is no basis for an AI verdict.
Content Credentials: integrity and trust are separate
C2PA Content Credentials attach signed provenance records to a file. They can describe creation and editing actions. Checking the signature and its binding to the file helps establish whether that record is intact. It does not establish that every claim is true or that the signer is someone you trust.
Our tool checks embedded credentials locally with the official SDK. It reports the active credential and local validation results, including reported history failures. It does not fetch remote manifests, trust lists or online certificate status. A successful integrity check therefore leaves signer trust unverified. An unavailable check is not the same as an absent credential.
Claude’s text watermark is not a hidden character
Anthropic describes a statistical watermark based on word choices, using a version of SynthID-Text. Detecting its pattern requires the corresponding key. Looking for zero-width characters cannot detect that watermark.
As checked on 8 September 2026, Anthropic offers its detection API in private preview to eligible organisations and plans broader access. This site does not have that integration. A detected mark would indicate possible Claude involvement, including substantial editing; it would not prove sole authorship. Short passages and constrained text offer less evidence.
Invisible image watermarks need their own detector
Google’s SynthID embeds an invisible signal in generated media, separate from ordinary metadata. Google documents provider-specific checks through Gemini and its SynthID tools. Those checks concern supported Google AI content; they are not universal AI tests. Our local metadata inspector does not perform a SynthID check.
If you choose to use an external provider’s detector, review that service’s data handling before submitting a file. Following a source link from this guide does not send your selected image to that service.
Use the result as a starting point
An AI declaration, a failed credential and no metadata are three different findings. Read the actual field and the validation result together. Conflicting records deserve further checking. For a consequential decision, seek the original source and supporting context rather than turning an incomplete file record into an accusation.
Sources and upkeep
These sources inform the guidance above. The examples and checklists are modeleven’s editorial recommendations, not promises of results. Sources reviewed 8 September 2026.