Check provenance in Gemini image downloads
Compare embedded Content Credentials, invisible SynthID, and visible watermarks on images exported from Gemini Apps.
Gemini image metadata can tell you what a particular download carries. To check whether Google AI made or edited the image, you can also ask Gemini to look for SynthID, a watermark encoded in the image itself. Start with the file downloaded directly from Gemini. A screenshot, an editor export, and a copy saved from a social post can each carry different file metadata. Google’s verification guide
This guide draws on published documentation. Authentic Gemini or Nano Banana exports were not independently tested.
Identify where the image came from
Nano Banana is Google’s name for a family of image models. The original Nano Banana is Gemini 2.5 Flash Image, Nano Banana Pro is Gemini 3 Pro Image, and Nano Banana 2 is Gemini 3.1 Flash Image. Google also lists Nano Banana 2 Lite and Nano Banana 2.1 in its developer guide. Gemini Apps Help describes generating or editing an image with Nano Banana 2 and using Redo with Pro on eligible accounts. If you know the model, note it alongside the app or API and whether the file came from a new generation or an edit. A model name alone cannot tell you which metadata survived into the file you received. Gemini Apps image guide · Gemini API image generation guide
Three provenance signals in Gemini images
The visible Gemini watermark
Where it lives: In the displayed image. How to check: Look at the image. Eligible Gemini Apps users can control this mark separately.
Google’s Media watermark setting in Gemini Apps controls the visible mark. Availability depends on the plan and country, and the setting applies to Gemini Apps on web and mobile. Google says it leaves SynthID and Content Credentials unaffected. An image with no visible mark can therefore still carry both signals. Gemini Apps watermark settings
The invisible SynthID watermark
Where it lives: An invisible signal encoded in the image pixels. How to check: Use Google’s verification tools. Detection can indicate that Google AI generated or edited at least part of the image.
Google describes SynthID as an invisible watermark for media generated or edited in Gemini Apps. For images, the signal is encoded in the pixels and designed to survive common changes such as compression. Detection can still be inconclusive. Google gives simple images and very small edits as examples where a detectable signal may be limited. Removing EXIF, XMP or an embedded credential does not remove a watermark encoded in the image pixels. AiMetaOff has no SynthID detection or removal feature. Google DeepMind’s SynthID overview · Gemini Apps verification guide
Embedded C2PA Content Credentials
Where it lives: A signed provenance record that may be embedded in the file. How to check: Inspect the original download for a manifest. Read what it records about the source or edits and review the verification results.
Google’s Gemini Apps Help describes C2PA metadata as part of its provenance approach. In November 2025, Google specifically announced embedded C2PA for Nano Banana Pro images generated in the Gemini app, Vertex AI and Google Ads. In February 2026, Google reiterated its approach of pairing SynthID with Content Credentials when it announced Nano Banana 2. These statements describe product support. To find out whether a specific download carries a credential, inspect that file. The Gemini API also documents SynthID for generated images, while the Nano Banana Pro announcement names specific surfaces for C2PA. An API response saved by a third-party app can pass through its own export workflow. Gemini Apps watermark settings · Nano Banana Pro C2PA rollout · Nano Banana 2 announcement · Gemini API image generation guide
What EXIF and other file metadata can tell you
EXIF often carries camera information, dates or orientation. XMP can carry descriptive or editing information. A PNG can also have text chunks. These are examples of records a supported format can hold. Their presence and values depend on what wrote the file, so an empty EXIF report tells you little about AI use.
A C2PA manifest can contain signed statements about the image’s origin, ingredients or editing history. A signature check concerns the signed record, while a content-binding check examines whether it matches the covered parts of the file. Trust in the signer is a further consideration. The guide to embedded and durable Content Credentials explains how some credentials can also be stored separately from the file. C2PA Content Credentials explainer
How to inspect a Gemini image locally
Save the direct Gemini download. If you received a forwarded copy, ask for the download too when you need to compare provenance.
In AiMetaOff’s Metadata Viewer, open the file in Inspect File. The file stays in your browser. Look for the format and the ordinary metadata categories reported for that file.
Read the Content Credentials result separately. If the viewer can read an embedded manifest, check the recorded source and compare the claim signature and file content binding. Signer trust may be unverified, and the viewer skips remote manifests. A missing record means the local check found no supported embedded credential. An unsupported check means the viewer could not make that determination.
How to remove supported metadata and check the new copy
Keep the original, particularly if you need to preserve its provenance. In the same workspace, choose the supported metadata categories you intend to clean, run the cleanup and download the new copy. For example, if a supported PNG contains a personal caption in a text chunk, selecting the text category can remove that chunk along with other text in the category. Review what is selected before cleaning. EXIF, XMP, text and embedded C2PA are distinct categories where supported. The AI metadata remover focuses on supported text and XMP. Supported embedded C2PA cleanup depends on the file structure and requires its own selection.
Reopen the downloaded copy in Inspect File and check the same categories against the original. Look for the specific field or record you meant to remove, and open the image to confirm it still displays as expected. A remaining orientation record can be intentional. If you plan to share a later export, inspect that exact copy too. See how to verify metadata removal.
How to check SynthID with Google
For an image you are comfortable uploading to Google, sign in to Gemini, attach the file and ask whether Google AI created or edited it. Read the response for the SynthID result and any uncertainty it reports. A detected watermark can indicate an AI edit to just part of the image. An inconclusive or undetected result provides no proof that the image is human-made.
Google also documents Content Credentials verification on the Gemini web app and Android for compatible credentials. This is a separate result that can summarize source and editing history, including records from supported products outside Google. Use the SynthID result for the watermark question and the credential result to understand the recorded history. Google’s verification instructions and compatibility limits
Will removing metadata remove a social platform’s AI label?
A platform applies labels to posts under its own rules. Meta, for example, describes using industry-shared signals and uploader disclosures for its AI info label. A local metadata report cannot tell you which signal a platform used or whether a label will remain. Follow the platform’s disclosure requirements when sharing generated or edited media. Meta’s AI labeling policy
FAQ
Can a Gemini image carry both SynthID and Content Credentials?
Yes. Google uses both mechanisms. SynthID can flag Google AI involvement in the image, while a compatible credential can show recorded details about its source or how it was edited. Google’s Gemini verification guide
Can I tell which Nano Banana model created an image?
Start with the generation history or export details if you have them. A credential may name a model if the issuer recorded it. A SynthID result can indicate Google AI involvement, but the model name can still be unknown.
Can I check a screenshot of a Gemini image?
Google’s verifier accepts screenshots for SynthID checks, though the result can be inconclusive. A screenshot is a new file. For embedded Content Credentials or EXIF, inspect the original download whenever possible. Google’s verification guide