Free tool · Provenance semantics
Inspect AI metadata before publishing
Check embedded provenance signals in your image. Platform labels and disclosure obligations require a separate assessment.
Will this image be labelled AI-generated? This tool cannot determine that. It reads embedded C2PA, XMP and IPTC signals and selected Exif tag IDs, then estimates how a platform might interpret them. For eligible JPEGs it can create a narrow local privacy copy and compare exact files. It cannot observe a platform's final label or served asset, certify signer trust or decide the legal disclosure obligation. Check the disclosure question separately →
Browser-only metadata inspection
Inspect an image before posting
The selected file stays in this browser. Metadata inspection is separate from optional offline C2PA validation in a disposable worker. Disclara does not send the file to an API or store its contents.
Ingredient-level AI usage (optional)
Record the final composition separately from whole-file metadata. These are your production facts, not automatically extracted or authenticated C2PA.
One AI ingredient is not a fully generated asset
A camera video can contain AI-enhanced dialogue or a generated sound effect without being fully AI-generated. Record the final-output components, distinguish inputs used only during production, and assess materiality separately. The optional composition form records your facts; the local image validator can display embedded SDK ingredient relationships and selected explicit AI declarations. Neither establishes what Instagram or TikTok will display.
Adobe documents that supported Premiere/Media Encoder exports can carry credentials from Adobe AI-generated assets in the sequence. Disclara does not inspect MP4, MOV, WAV or MP3 in this version. For video/audio, record the composition manually in the checker and retain the source production records. Audio cleanup is not automatically classified as a synthetic voice.
Required action: preserve ingredient evidence and verify the final served label. A wording candidate such as “Audio enhanced using AI” is not a conclusion that disclosure is mandatory.
Exact-file replay
Recheck a provenance comparison
Choose a candidate-bound comparison capture or local Exif export record (v1 or v2), the same source image, and the exact image the record describes. For a browser-prepared copy, choose that copy as the candidate—not a different file selected during a failed local-copy recheck.
This replays the metadata comparison against newly selected bytes in a disposable browser worker with a 20-second deadline. For a local Exif export record, it also reruns the selective editor and compares the candidate byte-for-byte. A matching record can still contain an inconclusive metadata scan. Approval, policy and disclosure remain caller declarations; the JSON is unsigned and a past recheck or publication cannot be authenticated. The worker is not an operating-system sandbox.
Applies when
- You have an approved JPEG, PNG or WebP and a candidate for publication.
- A mixed asset contains a generated element or AI-assisted edit.
- You need to compare provenance before and after export; a narrow JPEG-only privacy copy is available for some unsigned Exif files.
Doesn’t establish
- Signer trust or revocation status; the optional check assesses embedded credential integrity only.
- That all private metadata was found or removed; MakerNotes, XMP privacy fields and duplicates outside selected standard tags are outside the local export.
- The exact file or label a platform will serve after upload.
Example: remove private fields, preserve provenance
An agency may want to remove Exif GPS coordinates before publication while retaining a C2PA credential and an IPTC AI source declaration. Disclara blocks its local editor for that signed or multi-metadata file: a credential-aware editor must create a new derivative, followed by validation of the exact publication candidate. The comparison then reports lost provenance separately from missing GPS tags—even if the agency reports a visible AI label.
For a simpler JPEG with a single readable Exif segment and no detected C2PA, XMP or IPTC, the browser can instead clear standard GPS and body/lens serial entries and addressed values in a copy. After saving, reselect the local file to compare its exact bytes with the prepared output; a match cannot prove which path it came from or that it was published. If the reselected file differs, the prepared-copy comparison does not describe that file: select it separately as the intended publication candidate. Required action: review the entire output, confirm policy and disclosure, and retain the source snapshot. If client policy prohibits the change, do not publish. The downloadable records are unsigned preflight evidence—not proof of publication, permission or legal compliance.
Why valid provenance can still mislead
C2PA can describe actions, ingredients and regions. A downstream platform may expose a simpler whole-asset label. The credential can therefore be technically valid while the audience-facing interpretation overstates how much of the final image was generated.
Disclara records this separately as PROVENANCE_VALID_BUT_MISLEADING only when the optional local validator finds intact embedded provenance and the inspected scope conflicts with the declared final-asset scope.
The 10 MB verifier runs in a disposable browser worker with a 12-second deadline. It does not fetch external manifests or check signer trust. The 25 MB metadata scan is separate; neither check is an operating-system-grade sandbox or a legal disclosure decision.
Sources
The C2PA specification defines the media-container locations inspected here and notes that hard bindings may cover metadata outside the manifest, including Exif. Its manifest-assertion redaction mechanism is not the same as editing Exif in an existing signed file. CIPA’s Exif standard supports the bounded TIFF tag scan and selective JPEG copy, while IPTC guidance informs digital source-type interpretation. The pinned C2PA web SDK performs the optional offline integrity check; Adobe’s advisory informs the tracked validator-version gate. This version check is not a general security certification.
- C2PA Technical Specification (2.4 series)
- Exif Version 3.1 — CIPA DC-008-Translation-2026
- @contentauth/c2pa-web 0.15.1 release
- Security updates available for Content Credentials SDK | APSB26-147
- IPTC metadata guidance for AI-generated synthetic media
- C2PA User Experience Guidance for Implementers (2.2)
- AI Product Terms
- Labeling AI Content — Meta Transparency Center
- About AI-generated content — TikTok Support
- AI & Transparency Framework v2
Last verified: 2026-09-29.
Minimum necessary, defensible disclosure: describe the material AI contribution precisely without removing authentic provenance merely to avoid a broad platform label.
