Illustration of a magnifying glass scanning a photo, a fingerprint, a verified shield badge, and a document with an AI tag

AI Provenance Scanner

One place for every format: drop in an image, video, audio file, PDF or Office document and it is checked for C2PA Content Credentials and AI-tool metadata fingerprints, or paste text to find hidden Unicode markers. The scanner works out what the file is. A signal found is a hint, not proof; no signal found doesn't confirm authenticity either — details below.

Checks for C2PA Content Credentials in images (JPEG, PNG, WebP, TIFF, GIF, AVIF, HEIC, SVG, DNG), video (MP4, MOV, M4V, AVI), audio (MP3, M4A, WAV, FLAC) and PDF. Images are additionally checked for known AI-generator metadata fingerprints (Stable Diffusion, ComfyUI, Midjourney, and others). Documents — PDF, Word, Excel, PowerPoint, OpenDocument — get the same check as the PDF AI Scanner: AI-tool names in the metadata, graded strong or weak, and hidden characters in the text. A plain-text file (.txt, .md, .csv…) is scanned like pasted text.

Drag & drop a file here, or ·

Preview of the selected image

Paste text to find hidden Unicode markers — zero-width characters, Unicode tag characters (a steganography channel), bidirectional controls, variation selectors, unusual whitespace, and mixed-script homoglyph substitution. Nothing is removed; the invisible parts are surfaced and made visible. This does not read statistical token-distribution watermarks (e.g. SynthID-Text) — those need the vendor's own detector.


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

C2PA Content Credentials

Content Credentials are the industry standard (backed by Adobe, Google, Microsoft, and the broader Content Authenticity Initiative) for cryptographically attaching provenance data to a file at the moment it's created or edited. Specification: C2PA technical specification; background: Content Authenticity Initiative. This tool reads and validates manifests with the open-source c2pa-python bindings (version 0.37.7) of the C2PA Rust SDK.

  • Every container the C2PA SDK can read: images (JPEG, PNG, WebP, TIFF, GIF, AVIF, HEIC/HEIF, SVG, DNG, JPEG XL), video (MP4, MOV, M4V, AVI), audio (MP3, M4A, WAV, FLAC) and PDF. Video and audio files get this check only — the metadata-fingerprint tier below is image-specific. Word, Excel and PowerPoint files are not containers the SDK reads, so they get no C2PA check.
  • Cryptographic validation, not just presence: a manifest's signature is actually verified, so a forged or tampered-with manifest is reported as invalid, not treated as legitimate just because a manifest exists.
  • Claim generator identified: the specific tool or service that created or last signed the file, when the manifest declares one.
  • AI declarations surfaced directly: assertions describing the content as AI-generated, AI-composited, or used in AI training/mining are pulled out and shown plainly, not buried in raw manifest JSON.

AI-Generator Metadata Fingerprints

Most AI image tools in everyday use — especially local, non-hosted ones — don't attach C2PA credentials at all, but many still leave their own fingerprints behind in ordinary file metadata.

  • EXIF fields: the Software, UserComment, ImageDescription, and Artist tags, where tools like AUTOMATIC1111's Stable Diffusion WebUI write full generation parameters (steps, sampler, CFG scale, seed).
  • PNG text chunks: ComfyUI embeds its entire workflow graph as JSON directly in the file; Stable Diffusion WebUI writes a parameters chunk with the prompt and settings.
  • XMP fields: the standardized IPTC DigitalSourceType property, which Midjourney and a growing number of other generators set to explicitly flag AI-generated or AI-composited output.
  • Not exhaustive, and not proof: new generators appear constantly, and any of this metadata can be stripped or forged in seconds with a tool like exiftool — a match is a real signal, but a miss proves nothing either way.

Documents: PDF, Word, Excel, PowerPoint, OpenDocument

A document is handed to the same analysis as the PDF AI Scanner, and the result looks the same on both pages.

  • AI-tool names in the metadata: the PDF Info dictionary and XMP packet, or the Office document properties, are searched for ChatGPT, Claude, Gemini, Copilot and other tools. A match in a field that names the software is strong; a match in the author, title or keywords is weak, because a document can simply be about the tool.
  • Notes about the metadata: no metadata at all, no software named, or an Office file built by a programming library (python-docx, openpyxl, PptxGenJS…) rather than an office program. These say how a file was made or handled and never count as an AI signal.
  • Hidden characters: the document's text gets the same check as the Text tab.
  • Not read: the old binary .doc, .xls, .ppt formats and password-protected Office files.

Text and text files

Text pasted into the Text tab, or uploaded as a plain-text file, is checked character by character for invisible and look-alike characters. The wording itself is not judged: detectors that guess “AI-written” from style are wrong about people's own writing too often to sit next to verifiable facts.

What this can and can't tell you: a cryptographically-signed C2PA manifest is real, verifiable evidence. Everything else here — EXIF tags, PNG text chunks, XMP fields — is metadata that any AI tool chose to write and anyone can strip or forge with a tool like exiftool in seconds. So: a signal found is a strong hint, not proof; no signal found is not proof the image is authentic or human-made, only that no known fingerprint survived in this copy. This also does not detect Photoshop edits, splicing, or deepfakes — that's a different (and far less reliable) kind of forensic analysis this tool does not attempt.
💡 Show API usage examples (cURL)

Scan an image, video, audio file or document and get JSON findings

curl -sS -X POST https://devops.majbase.com/ai-image-scanner/api \
    -F "file=@photo.jpg"      # or clip.mp4, track.mp3, report.pdf, letter.docx, notes.txt

Example response

{
  "filename": "photo.jpg",
  "format": "PNG",
  "media_kind": "image",
  "size_kb": 1842.5,
  "sha256": "a82f3c1e…9d4b",
  "dimensions": "1024x1024",
  "verdict": "unverified_ai_signal",
  "c2pa": {
    "manifest_found": false,
    "validation_state": null,
    "claim_generator": null,
    "ai_assertion": null,
    "error": null
  },
  "metadata_signals": ["ComfyUI", "Stable Diffusion"],
  "exif_texts": {"Software": "…"},
  "png_texts": {"parameters": "…", "workflow": "…"},
  "camera_details": {},
  "camera_metadata_present": false
}

verdict is one of verified_ai, verified_provenance, unverified_manifest, unverified_ai_signal, no_signal (meanings in "How to read this verdict" above the result, or in the legend that appears with any scan). media_kind is image, video, audio or document; for anything but image, dimensions is null and the image fingerprint fields are empty. A document (PDF, Word, Excel, PowerPoint, OpenDocument) also carries everything the PDF AI Scanner API returns: kind, signal_strength, hits, metadata_notes, text_scan. A plain-text file returns the text findings shown further down, with media_kind text. On a server-side error the response is {"error": "…"} with a non-200 status. No API key is needed; one file per request, 20 MB maximum. This runs on a small server, so keep automated use to a modest rate.

The same call from Python

import requests

with open("photo.jpg", "rb") as f:
    r = requests.post("https://devops.majbase.com/ai-image-scanner/api",
                      files={"file": ("photo.jpg", f)}, timeout=60)
r.raise_for_status()
report = r.json()
print(report["verdict"], report["metadata_signals"], report["c2pa"]["validation_state"])

And from JavaScript (browser or Node 18+)

const form = new FormData();
form.append("file", fileOrBlob, "photo.jpg");
const res = await fetch("https://devops.majbase.com/ai-image-scanner/api", { method: "POST", body: form });
if (!res.ok) throw new Error(`scan failed: ${res.status}`);
const report = await res.json();
console.log(report.verdict, report.metadata_signals);

Scan text for hidden Unicode markers

curl -sS -X POST https://devops.majbase.com/ai-image-scanner/text-api \
    -H "Content-Type: application/json" \
    -d '{"text": "paste text here"}'

Example response

{
  "char_count": 42,
  "verdict": "hidden_characters",
  "summary": "Hidden invisible characters found — …",
  "hidden_total": 3,
  "categories": [
    {"key": "zero_width", "label": "Zero-width / invisible characters",
     "severity": "high", "count": 3,
     "codepoints": [{"cp": "U+200B", "name": "ZERO WIDTH SPACE",
                     "count": 3, "positions": [5, 11, 20]}]}
  ],
  "mixed_script_words": [],
  "annotated": "hello⟨U+200B⟩world…",
  "annotated_truncated": false
}

verdict is one of hidden_characters, mixed_scripts, unusual_only, clean. Nothing is stored.

🤖 Use this API as a Claude Skill

Want an AI agent (Claude Code, claude.ai, or anything else that supports the Agent Skills format) to use this tool for you on request? Download the skill below and add it:

⬇️ Download SKILL.md · View raw