Updated May 20, 2026

OpenAI Image Verification Tool: C2PA, SynthID, and Limits

Name: GPT Image 2 Status
Author: GPT Image 2 Status

OpenAI now offers a public image verification workflow for supported OpenAI-generated images. The important part is not treating it like a magic AI detector. It checks for provenance signals such as Content Credentials and SynthID, then gives reviewers evidence to interpret.

Try GPT Image 2

C2PA Content Credentials

Signed metadata can describe origin, edits, and generator information when platforms preserve it.

SynthID watermarking

An invisible watermark signal can remain useful when normal metadata is stripped or lost.

Official verification result

Use the official OpenAI verifier for supported OpenAI-generated images instead of copying unofficial detector claims.

Do not sell this as certainty

A positive provenance signal can be useful evidence. A negative result is not proof that an image is human-made. Screenshots, format conversion, platform compression, and metadata stripping can all change what a verifier sees.

Verification workflow for platforms

Step	What to check	Decision rule
1. Preserve the original file	Keep the upload before compression, resizing, or screenshot conversion.	Run verification on the closest original asset available.
2. Check official provenance	Use OpenAI's verifier for Content Credentials and SynthID signals.	Positive signal means supported evidence, not automatic rejection.
3. Hash and log the review	Store file hash, source, timestamp, reviewer outcome, and policy category.	Avoid storing raw private files longer than needed.
4. Escalate inconclusive cases	Ask for source files, edit history, creator attestation, or manual review.	Missing provenance should be classified as inconclusive.

C2PA vs SynthID

C2PA is metadata-centered: it can carry signed information about where a file came from and how it was changed. That is useful, but metadata can be removed by downloads, uploads, screenshots, or transformations.

SynthID is watermark-centered: it embeds a signal into generated media. It can be more durable than metadata, but it still should be treated as one signal in a broader review process.

Policy template

We may check uploaded images for supported provenance signals, including C2PA metadata and watermark signals. Verification results are used as review evidence and may be inconclusive.

Creators remain responsible for disclosure, licensing, and rights. We may request source files or additional context when provenance is missing or unclear.

Primary sources

OpenAI provenance announcement OpenAI verification tool Google SynthID overview

Next step

If you are testing generated images for your own product, start with a controlled prompt, keep the original output file, and record what changes after upload, compression, or editing.

Generate a test image Estimate API cost

FAQ

Can OpenAI's verification tool prove an image is not AI-generated?

No. If no provenance signal is found, the safer reading is inconclusive. Metadata can be removed and watermark signals can fail after some transformations.

What does SynthID add beyond C2PA?

C2PA carries signed metadata. SynthID adds an invisible watermark signal designed to survive more transformations than metadata alone. They are complementary, not interchangeable.

Should marketplaces or publishers rely on this alone?

No. Use it as one evidence source alongside upload rules, creator attestations, manual review, source files, and policy enforcement.