Does this file
disclose AI use?

What can this tell you?

Content Credentials are signed records that can describe how an image was created and edited. Some include a declaration of AI generation or AI-assisted editing. This checker reads the active embedded credential and checks its integrity using the C2PA SDK.

It cannot determine whether every image was made by AI. No credentials means the image’s origin is unknown. Screenshots, re-encoding and sharing can remove credentials. A valid credential does not prove that the depicted scene is real or that every declaration is true.

Office document clues

For Excel (.xlsx/.xlsm), Word (.docx/.docm) and PowerPoint (.pptx/.pptm), the checker reads core, application and custom document properties, including author, last editor, application, timestamps and any other fields present in those property parts. It also exposes selected styles, fonts, colours, themes, Excel sheet visibility, package relationships and a complete package file inventory. Template variants are supported too. Macros and external links are never executed or opened.

An AI tool named in metadata is a clue, not verified authorship. A reference to ChatGPT, OpenAI, Claude, Anthropic, Copilot or Gemini could come from a template, description or later editing step. Library names such as openpyxl, XlsxWriter or python-docx indicate possible automation, which is also used without AI.

There is no universal font or colour palette used as a ChatGPT fingerprint in this checker. Common defaults and templates overlap with human-created documents. The openpyxl style documentation shows how ordinary code sets workbook fonts and colours. Matching formatting alone cannot identify the authoring tool.

This is a bounded inspection, not a scan of every possible metadata location: up to 128 selected XML parts, 2 MB per part and 8 MB total. Skipped or damaged parts are listed in the result. Document body text, cell values, slide content, comments, embedded files, macros and digital signatures are not analysed. For PDF files (.pdf), the checker inspects uncompressed /Info dictionary entries (Title, Author, Creator, Producer, CreationDate, ModDate), embedded XMP packets and /BaseFont descriptors. Older .doc/.xls/.ppt files, encrypted documents and OpenDocument files are not supported.

Broader Office hints

The result includes a guide to common Office properties, showing their observed values and whether each field is present, empty, absent or unreadable. It explains authors and last savers, timestamps, Word revision counters, application/version labels, templates, editing time, document statistics and application state flags. Excel application/version and calculation attributes, and Word revision-session identifiers and compatibility settings are also exposed when present.

Office can automatically maintain the last saver, creation date and other hidden properties. The exact fields vary by application, platform, version, template and privacy settings. Some libraries write Office labels too: openpyxl’s application-property defaults include “Microsoft Excel”. Word updates revision counters on save, while python-docx does not automatically increment them.

These are observations about possible saving history. Matching timestamps, an empty author, revision 1, zero editing time or missing properties do not establish AI use. Saving an AI-assisted file in Office can change metadata, and a template or ordinary export can retain library-like defaults. Each interpretation links to documentation; there is no AI probability score.

Additional workflow hints include attached Word template references, stored Word variables, the change-tracking setting, Excel calculation settings, uniform or default-era ZIP entry timestamps, and a PowerPoint slide-statistic/package-part comparison. The inventory also highlights comment/collaboration parts, custom XML, embedded files, macro projects and signature-related paths. Their presence can be consistent with templates, exports or integrations; it does not establish AI authorship. Embedded payloads, comment bodies and signatures are not opened or validated. A tracking setting does not prove tracked edits exist, and a slide XML part is not necessarily a slide referenced by the presentation.

Expand the evidence sections to see exactly what was found. Download the JSON report for extracted properties, package inventory and the inspected source XML. The report may contain personal information from your file; downloading it stays local.

Understanding your result

• AI generation declared: the active credential contains the standard AI-generation source type. Read the integrity result alongside it.
• AI or algorithmic editing declared: the active credential describes AI compositing or algorithmic enhancement. This does not mean the whole image was generated by AI.
• Credentials found; no AI declaration identified: a record exists, but the inspected actions do not disclose AI use. This is not proof of a camera original.
• No embedded credentials found: origin unknown. The image may lack credentials, have lost them, or reference a remote record.
• Credentials failed validation: do not rely on the declarations. The file or credential may be damaged, altered or invalid.

What stays local?

Office XML is inspected locally in your browser. Image credentials are processed in a background worker on your device. We host the checker’s WebAssembly engine ourselves and load it only when you choose an image. Files, filenames and results are not sent as analytics events. Website analytics is described in our privacy overview.

This version does not retrieve remote manifests, load remote thumbnails, check certificate revocation online, or consult a public signer trust list. Signer trust is not established, even when local integrity checks pass. It does not inspect invisible watermarks or guess from visual patterns.

Need a different kind of check?

Use our file type checker to identify an unknown format or misleading extension. Read how to check whether an image discloses AI use for examples and limitations.

For the underlying standard, see the C2PA explanation and the Content Authenticity Initiative’s browser SDK.