An image retains the “Generative AI” tag after extensive manual editing because Content Credentials (C2PA) operate on an immutable, append-only provenance graph. Once an Adobe Firefly generative tool—such as Generative Fill, Generative Expand, or Text to Image—is used in a document, that action is recorded as a foundational “ingredient.” Subsequent manual edits (painting, cloning, color grading, or rasterizing) do not delete earlier nodes in the cryptographic manifest; they simply append new actions on top of the original generative event.
Fast Fix
If you have fully replaced or painted over all generative content and need a clean manifest, create a new document (File > New), copy your manually created pixel layers into it, ensure Content Credentials is toggled Off in Window > Content Credentials, and export via Save for Web (Legacy) to prevent inheriting the project’s upstream assertion history.
Understanding the C2PA Directed Acyclic Graph (DAG)
To understand why the tag persists, consider how Content Credentials log history. Unlike Photoshop’s local History panel, which clears when you close a file, the C2PA engine builds a permanent tree of digital signatures:
[Camera Capture / Blank Canvas] (Root Node)
│
▼
[Firefly Generative Fill] (Assertion: AI Tool Used) ──► Immutably Sealed
│
▼
[Manual Paint Layer] (Assertion: Brush Tool)
│
▼
[Curves Adjustment] (Assertion: Color Edit)
│
▼
[Merged & Flattened] (Assertion: Layer Merge)
│
▼
[Final Exported Asset] ──► Inherits all parent nodes (Flags AI-Generated)
Because digital signatures verify that an asset has not been tampered with, no user or software is permitted to edit or redact historical assertions within an active manifest tree. As long as the document session remains continuous, the presence of an AI assertion anywhere in the ancestor chain requires the final export to disclose it.
Common Attempts That Fail to Clear the AI Tag
Many standard Photoshop workflows seem like they should eliminate the generative tag, but the C2PA engine treats them as normal derivative steps:
- Rasterizing the Generative Layer: Merging an AI layer down onto a blank layer or rasterizing its Smart Object status simply creates a new pixel layer whose ingredient was the generative layer. The parent link remains unbroken.
- Cloning or Painting Over AI Pixels: Even if you paint over 100% of the pixels generated by Firefly using standard brush and clone tools, Photoshop logs the action as “Retouching applied to Generative Ingredient.” The manifest still lists the AI ingredient as an input.
- Deleting the AI Layer Mid-Session: If you generate an element, decide you do not like it, and delete the layer, the session history may still retain the generative event in its session cache unless purged before signing.
- Using Smart Objects: Embedding a generative asset inside a Smart Object does not quarantine it. The manifest engine parses nested Smart Objects as child ingredients, carrying the AI flag upward into the main document.
For a deeper look at inspecting this assertion tree, see How to View “Edit History” via the Content Credentials Panel.
How Platforms Read the Inherited AI Assertion
When you post the file online, social media networks and verification tools inspect the manifest’s top-level summary:
| Platform / Tool | Inspection Method | Label Applied |
|---|---|---|
| Adobe Verify | Reads entire manifest tree | Shows full timeline (AI creation followed by manual edits) |
| Instagram / Meta | Evaluates C2PA digitalSourceType | May flag as “AI info” or “AI-generated” depending on platform thresholds |
| Desktop C2PA Viewers | Reads embedded JUMBF metadata | Flags asset as composite containing synthetic media |
To understand how third-party platforms evaluate these tags, consult Why Your “AI-Generated” Label is Missing on Instagram/Facebook.
How to Clean the Manifest for Legitimate Manual Art
If an image has been completely redone by hand, or if an accidental AI generation was deleted and you must produce a file that accurately reflects only human effort, you must break the document’s historical container:
Method 1: The Clean Canvas Migration
- Open your project containing the finished manual artwork.
- Ensure all generative layers are hidden or deleted.
- Select your manual artwork layers, right-click, and select Duplicate Layers…
- Set the Destination Document to New. Name it
Clean_Master.psd. - In the new document, open Window > Content Credentials.
- Verify that the activity log shows only manual edits. If the AI tool is absent from the summary, the parent link was successfully severed.
- Export using File > Export > Export As.
+--------------------------------------------------------------+
| Duplicate Layers Dialog |
| |
| Destination: |
| Document: [ New v ] |
| Name: Clean_Master.psd |
+--------------------------------------------------------------+
Method 2: Purging via Legacy Export
If policy restrictions lock your credentials toggle due to Firefly policies (explained in How to Opt-Out of Content Credentials (When Allowed)), route the pixels through the legacy export engine:
- Select File > Export > Save for Web (Legacy).
- In the preset dropdown, select JPEG or PNG-24.
- Under Metadata, select None or Copyright Only.
- Click Save.
The legacy engine strips all JUMBF metadata blocks during encoding, creating a flat raster file with no historical manifest.
Proving Genuine Human Authorship
If you must provide transparency rather than stripping metadata, do not try to hide the AI tag. Instead, use Content Credentials to highlight your manual contributions:
- Ensure Window > Content Credentials is set to record Edits and activity.
- Connect your verified identity so your name is bound to the manual work, as outlined in How to Use Content Credentials for “Proof of Human Effort”.
- When clients or platforms inspect the asset on How to Verify an Image Using the Adobe Verify Website, they will see the precise sequence: an AI element was added, followed by extensive, authenticated human brushwork and compositing.