C2PA Content Credentials
Reads embedded content credentials and manifest data.
- Upload an image to run full server analysis.
AI image detection
All-in-one AI image detector
Upload one image and review C2PA credentials, Google SynthID, Meta watermark signals, generic AI fingerprints, and basic metadata in one workflow.
C2PA
Browser-side read
Watermark
SynthID / Meta
Fingerprint
Generic model
Drop an image here
JPG, PNG, WebP, or AVIF up to 10 MB
Image preview appears after upload
No file selected
JPG, PNG, WebP, or AVIF up to 10 MB
Upload runs browser-side C2PA first. Full server analysis calls Gemini, a Meta adapter, and Replicate only when those API keys are configured.
Summary
Waiting for image upload
Upload an image to run full server analysis.
Detection channels
OpenAI / Google / Adobe provenance
Looks for trusted signer, claim generator, and AI provenance evidence.
- Server check required
Google SynthID
Checks for Google AI invisible watermark signals through Gemini/SynthID capability.
- Server check required
Meta Stable Signature / Seal
Checks Meta/Stable Signature style watermark signals through an optional adapter.
- Server check required
Generic AI fingerprint
Calls a generic AI image detector model as a secondary signal.
- Server check required
Basic metadata signals
Records format, size, dimensions, and cache hash evidence.
- Server check required
Results are an aggregation of signals, not a legal or factual final verdict. Compression, screenshots, reposting, and edits can remove watermarks or metadata.
How the detector works
The workflow starts with the strongest deterministic signals before using heavier model-based checks. It is designed for reviewers who need a fast first pass, not a final legal ruling.
Last updated: May 20, 2026
1. Provenance
Read C2PA first
The browser reads embedded Content Credentials before upload. A valid credential can identify the signer, claim generator, edit history, and whether an AI tool was declared in the media record.
2. Watermark
Check provider marks
Server analysis can call Gemini/SynthID and a Meta signature adapter when configured. These checks look for provider-specific marks that may survive ordinary viewing but can still be damaged by edits.
3. Fingerprint
Use a model fallback
A generic detector adds a probabilistic signal when no watermark or credential is found. It is useful for triage, but it should never override stronger provenance or watermark evidence.
Methodology and signal hierarchy
AI image detection is most reliable when the tool separates cryptographic provenance, watermark signals, model-based guesses, and ordinary metadata. This page uses that hierarchy so a missing signal does not get mistaken for proof that an image is human-made.
High confidence
A valid C2PA credential or a confirmed provider watermark is treated as the strongest signal because it comes from media provenance or a provider-side marking system.
Medium confidence
A trusted provider name, claim generator, or AI-generation assertion inside a manifest is useful, but it still needs context about validation state and signer trust.
Low confidence
Generic AI image classifiers are fallbacks. They can spot visual patterns, but they are vulnerable to compression, style transfer, screenshots, and domain mismatch.
No verified signal
No signal means no supported evidence was found. It does not prove the image is authentic, human-made, or unedited.
What each detection channel means
The tool reports every channel separately because different AI image generators leave different traces. A single image may contain a signed credential, an invisible watermark, both, or neither.
C2PA Content Credentials
Checks for signed provenance data embedded in the file. When present, it can describe origin, edits, ingredients, and claim generator information.
Trusted provenance
Looks for trusted signers or recognizable tool names from OpenAI, Google, Adobe, and related provenance ecosystems.
Google SynthID
Checks whether a Google AI watermark signal can be detected through configured provider capability. A negative result does not rule out other AI systems.
Meta Stable Signature / Seal
Provides an adapter slot for Meta-style and Stable Signature-style watermark detection when a compatible detector service is available.
Generic AI fingerprint
Uses a model-based AI image classifier as a supplemental signal. This channel is best used for triage, moderation queues, and manual review prioritization.
Basic metadata
Records file type, size, dimensions, and server-side hash. Metadata is helpful context, but it is easy to remove or rewrite.
Limits you should expect
AI image detection is a signal aggregation problem. Screenshots, cropped images, social media recompression, format conversion, and manual retouching can remove metadata or weaken invisible watermarks.
Model-based detectors also have false positives and false negatives. A polished illustration, 3D render, stock photo, or heavily edited camera image can resemble generated media even when it is not.
Use high-confidence signals to support review decisions, and keep inconclusive results in a human review workflow when the image has legal, brand safety, editorial, or moderation consequences.
Privacy and upload handling
Browser-side C2PA detection runs before server analysis. The full server check sends the image only after the user chooses to run it, and provider calls are only made when the relevant API keys are configured.
The upload limit is intentionally set to 10 MB even though Cloudflare plan limits can be higher. This keeps memory usage predictable and reduces risk when the Worker has to inspect binary image data.
The API uses a SHA-256 hash for cache lookup so repeated checks of the same image can reuse the prior result without rerunning every provider.
Reference sources
These references explain the standards and platform constraints behind the detector design.
C2PA open standard
Official overview of the provenance standard used for Content Credentials.
C2PA FAQ
Explains how provenance travels with media and why credentials can be lost.
Google DeepMind SynthID
Official description of invisible watermarking for generated media.
Cloudflare Workers limits
Platform limits that inform upload size, memory, and backend processing choices.
AI image detection FAQ
Practical limits and privacy details for image verification.
No detector can prove that in every case. This tool combines provenance, watermark, model, and metadata signals so reviewers can make a better decision with context.
C2PA Content Credentials are signed provenance records embedded in media files. When present, they can show the tool, signer, and process used to create or edit an image.
Screenshots, compression, social platform reposting, and manual editing can strip metadata or damage invisible watermarks. Some generators also do not publish detectable provenance.
The browser reads C2PA locally first. Full server analysis sends the image to the site API and then to configured detector providers only when those provider keys are enabled.