How to Test C2PA, SynthID and AI Detection After Screenshots

A screenshot creates a new image file. That simple fact can remove embedded provenance even when visible pixels still resemble the original. Cropping, resizing, JPEG recompression, metadata removal, and social-platform reposting can change different signals in different ways.

Start by checking the best available file with the AI Image Detector, but do not treat one score as a verdict. This guide explains documented C2PA and SynthID behavior, then provides a controlled protocol for testing transformed images. It is a testing method, not a report of measured survival rates. Product and standards details were last verified on July 30, 2026.

What can each signal actually tell you?

C2PA, SynthID, and pixel classifiers answer different questions. C2PA authenticates signed provenance claims. SynthID indicates supported provider-generated content when an authorized detector finds its watermark. A pixel classifier estimates whether visual patterns resemble its training categories. None alone establishes that an image is factually true or shown in context.

The current C2PA explainer says Content Credentials are opt-in, can be incomplete, and can be removed. Durable credentials may combine a cryptographic hard binding with watermark or fingerprint-based soft bindings to rediscover an external manifest. Recovery still depends on the implementation.

Does a screenshot remove C2PA Content Credentials?

A screenshot normally captures rendered pixels into a new asset. It does not copy the source image's embedded C2PA manifest into the new screenshot. The screenshot may receive its own metadata or credential from the capture software, but that would describe the new capture, not automatically preserve the original asset's signed history.

This is why the original download matters. A screenshot can show a Content Credentials icon or caption, but a picture of a label is not a machine-verifiable manifest.

C2PA also distinguishes incomplete from invalid provenance. A later C2PA-aware application can add a credential after an unrecorded crop, but reviewers must read what the active manifest actually attests.

Can SynthID survive screenshots, crops, and compression?

Google DeepMind says SynthID embeds an invisible watermark into generated media and is designed to withstand cropping, filters, and lossy compression. Gemini help says it will usually remain after resizing, recoloring, or compression, while warning that multiple alterations can make it undetectable.

For provider-specific evidence, use that provider's official verifier. The SynthID Checker separates C2PA and general pixel evidence, but does not use Google's private SynthID verifier.

OpenAI says supported generated images include C2PA and SynthID, while OpenAI Verify checks supported OpenAI signals. It also says watermarking can be more durable than metadata through screenshots. This is a design expectation, not a guarantee.

Say “designed to survive some transformations,” not “survives screenshots.” Detection depends on the source, edits, output, and verifier.

Why can an AI image detector score change after editing?

Pixel detectors analyze submitted pixels, not the lost original. Screenshots can introduce scaling, sharpening, color conversion, interface elements, or compression. Cropping removes context, resizing changes texture, and recompression adds artifacts. Scores may move in either direction.

Record the file variant, score, provider, date, and transformation history. Never reinterpret a broad “AI” label as SynthID, C2PA, or proof of a specific generator.

For a wider view of available workflows, see the AI image detector comparison. For source tracing beyond detector scores, use the AI-generated image identification guide.

How do you run a reproducible six-step verification protocol?

Start from a known original and change one variable at a time. The output is a comparison record, not a universal accuracy claim. Keep watermark, C2PA, metadata, and pixel results separate.

Step 1: Define the question and select source files

Write the exact claim, such as “Does this derivative retain an OpenAI provenance signal?” Obtain the original export or download. Record its application, date, format, dimensions, and acquisition path.

Use known-provenance files. An unknown viral image cannot reveal what signal originally existed.

Step 2: Preserve and fingerprint the originals

Copy originals into a read-only folder, keep their filenames, and calculate SHA-256 hashes before editing.

Do not overwrite the original or send it through a messaging app. Preserve its download receipt, creation page, prompt record, credential display, or source URL separately.

Step 3: Create controlled derivatives

Create every derivative from the untouched original. Example parameters are a tightly cropped screenshot at 100% display scale, a 10% crop on each edge, a 1024-pixel longest-edge resize, one quality-75 JPEG, a metadata-stripped copy, and one platform-repost download. These are comparison settings, not industry standards.

Record software and settings. Add a chained variant only after single-change variants, so cumulative degradation remains identifiable.

Step 4: Validate C2PA and inspect metadata

Inspect every file with a C2PA-aware tool. Record manifest presence, validation, signer trust, and any structured generative-AI assertion. A brand name in ordinary text does not prove AI use.

The AI Watermark Checker separates C2PA, metadata, and pixel evidence. Its model result is not a private provider-watermark conclusion. For the underlying evidence model, read C2PA vs SynthID.

Step 5: Run the appropriate official watermark check

For suspected Google AI content, upload one file at a time to Gemini. Google advises tightly cropping screenshots and avoiding collages. For OpenAI content, use OpenAI Verify and record watermark, trusted C2PA, neither, or error.

Provider tools cover different generators. The Google SynthID vs OpenAI Verify comparison explains which official path matches each suspected source, while the SynthID guide covers Google's result boundaries.

Step 6: Run pixel analysis and interpret the matrix

Submit all variants to one pixel detector under consistent conditions. Record the numeric score, not just a color or label. Repeat only when planned in advance.

Compare variants within each source. Report signals as present, absent, invalid, inconclusive, or unavailable. Do not force unsupported “real” or “fake” verdicts.

What should your results worksheet include?

Separate tool output from interpretation. Use fixed statuses and retain raw reports.

Why is “no signal found” not proof that an image is real?

An image may predate a watermark rollout, come from an unsupported generator, lose metadata, accumulate edits, or use an unsupported format. A service may also fail or lack access to a remote C2PA manifest.

OpenAI says a file may still be OpenAI-generated after metadata loss, watermark degradation, an unsupported source, or creation before signal deployment. C2PA likewise warns against distrusting media purely because credentials are absent.

Write: “This checker did not find a supported signal in this file,” not “This image is authentic.”

Which tools should you use with this protocol?

Use one tool per evidence question and keep the outputs separate. Start with the AI Image Detector for multi-signal screening, use the AI Watermark Checker for provenance and metadata evidence, and use the SynthID Checker to organize Google-related clues before official verification.

Upload the untouched original first, then submit one derivative with clearly recorded settings. A score change is meaningful only when you can identify which file transformation occurred between the two checks.

FAQ: Screenshot and compression testing

Does taking a screenshot always destroy SynthID?

No official source promises one outcome for every screenshot. SynthID is designed for some transformations, but detection depends on the file, capture, edits, and verifier.

Does a screenshot preserve C2PA?

The source manifest normally does not transfer. Capture software may add a new credential, or a durable system may rediscover provenance, but each mechanism requires verification.

Is a C2PA-positive image definitely AI-generated?

No. C2PA covers AI and non-AI media. Check the signature, signer, active manifest, actions, and structured digitalSourceType assertions.

Can social media remove AI provenance?

A platform may strip metadata, resize, crop, or transcode. Download its served derivative and test it.

Which file should I preserve first?

Preserve the earliest export or download before other tools touch it. Keep its hash, filename, metadata, source record, and verification results.