PDF AI Provenance Scanner

Checks a PDF for C2PA Content Credentials (cryptographically verifiable, when present) and scans its Info dictionary and XMP metadata for known AI writing-tool and AI-PDF-generator fingerprints (ChatGPT, Claude, Notion AI, ChatPDF, and others). A signal found is a hint, not proof; no signal found doesn't confirm the document is human-written either — details below.

ℹ️ What does this check, and what are the limits?

C2PA Content Credentials

The same cryptographically verifiable check the AI Provenance Scanner runs on images and video: if a C2PA manifest is embedded in the PDF, its signature is validated (a tampered or untrusted manifest is reported as such, not treated as legitimate), the signing tool is named, and any assertion declaring AI generation is surfaced. Signed PDFs are still rare in the wild — the C2PA SDK reads them but most PDF tools don't write them yet — so "no manifest" is the normal result today and says nothing by itself.

AI-Tool Metadata Fingerprints

Many AI writing assistants and AI-specific PDF generators leave their name in a PDF's own metadata — either because the exporting tool stamps it there by default, or because the document was exported directly from that service.

Info dictionary: the classic Producer, Creator, Author, Subject, Title, and Keywords fields every PDF can carry.
XMP metadata: the newer, XML-based metadata stream — Dublin Core fields (dc:creator, dc:description, dc:subject), the PDF namespace's Producer/Keywords, the XMP namespace's CreatorTool, and any custom properties a tool chose to add.
Not exhaustive, and not proof: new AI tools appear constantly, and any of this metadata can be edited, stripped, or forged in seconds with any PDF editor — a match is a real signal, but a miss proves nothing either way.

What this doesn't do

No writing-style analysis: this doesn't read or judge the document's actual text for "AI-sounding" prose. That's a fundamentally different, far less reliable kind of tool; this one only looks at metadata fields.

What this can and can't tell you: a cryptographically valid C2PA manifest is real, verifiable evidence. Everything else this checks is metadata that the authoring tool chose to write, and anyone can strip or forge with any PDF editor in seconds. So: a signal found is a real hint worth a closer look, not proof; no signal found is not proof the document is human-written, only that no known fingerprint survived in this copy.

💡 Show API usage examples (cURL)

Scan a PDF and get JSON findings

curl -sS -X POST https://devops.majbase.com/pdf-ai-scanner/api \
    -F "file=@document.pdf"