Nano Banana Image Provenance Test: SynthID, Metadata & Content Credentials

Research questions

• What ordinary metadata is present in Nano Banana-generated image files?
• Are embedded C2PA / Content Credentials readable by AI Watermark Center?
• Does metadata differ by Google product surface?
• Do visible-watermark behaviors differ by documented surface or tier, when voluntarily recorded?
• What can the local inspector verify independently?
• Which provenance claims remain dependent on Google’s SynthID detector?

Why surface differences matter

Gemini Apps, Google AI Studio, and the Gemini API are different products. Visible-mark settings documented for Apps are explicitly not applied to other Google AI features. Merging surfaces would invent a policy that Google does not claim.

Test methodology

Samples must be generated for this research or supplied with permission. For each file we will record provider, product family, exact model or model ID only if known from the UI or API (never guessed from a filename), surface or unknown, optional tier, generation date, generated vs edited, input type, output format, file size, and pixel dimensions.

Supported AWC inspection

The Phase 3 image inspector reports MIME/signature, EXIF, XMP, IPTC, GPS, software metadata, and embedded C2PA claims, actions, and SDK validation status for JPEG, PNG, and WebP. Soft-binding and cloud recovery are out of scope.

SynthID limitation

SynthID status for AWC is not-tested-by-AWC. If an official Google detector is used on a sample, the result is stored as a provider-verified result with verification surface, date, and evidence source = provider tool. That is not independent verification by AI Watermark Center.

Sampling plan

• Gemini App: generated stills exported as original files where the UI allows.
• Google AI Studio: same, recorded as a distinct surface.
• Gemini API: model ID taken from the request, not inferred later.
• No scraping of Google Images, social posts, or unrelated copyrighted photos.
• Private prompts stay private unless a sample is explicitly approved for public display.

Evidence classification

Inspector output is independently observed only after a real file is processed. Google documentation remains official evidence for product claims. Provider detector output remains provider-verified. Unknown cells in any future table will say Unknown or Not tested, never blank.

Planned output

After collection: a results table with sample, surface, model, visible mark, embedded C2PA, optional provider SynthID check, and AWC metadata result. No public CSV until data exists. No “Download dataset” offer while the dataset is pending.

Primary sources

Product facts follow current Google DeepMind, Gemini Apps Help, Gemini API, and Nano Banana family posts listed on this page — not competitor round-ups.