Instant Studio logo

playbooks

Does your AI content still carry its label?

· Adeh DeSandies · brand governance

What happens to the label between the model and the post Top row: a model writes a signed Content Credential into the file. A routine production step rewrites the file, the signature no longer matches, and the asset is published with no readable label. Bottom row, how Toolkit handles it: the model's signed original is kept untouched, the finished asset is signed with Instant Studio's credential, which lists the original as its source and records the edits, and the published asset passes verification. FROM MODEL TO POST A TYPICAL PIPELINE Model output signed by the model → Resize, re-encode file rewritten → Published no readable label IN TOOLKIT Model output kept untouched → Finished asset signed by Instant Studio → Published verifies, source listed Our credential names the model's original as its source and lists every edit.

When a brand publishes an AI-generated image or video, the file should include a label that says so. The industry standard for that label is C2PA, the specification behind Content Credentials, and the leading image and video models already write it into their output. The label is a signed record inside the file: which model made it, when, and that the content was generated by AI. LinkedIn shows it to viewers, TikTok reads it to apply its AI-generated tag, and anyone can inspect it at contentcredentials.org/verify.

Whether that label is still readable when your asset goes live depends on what happened to the file in between.

Instant Studio is a Contributor member of the C2PA, the coalition that maintains the standard, and Toolkit signs every finished image and video to it.

Why the label breaks

The signature covers the exact bytes A model's output carries a credential signed against the file's fingerprint. Resizing the image writes a new file with a different fingerprint, so verification finds no match and the credential fails. Steps that write a new file include resizing, compressing, re-encoding, adding a logo and the optimisation a content delivery network applies. THE SIGNATURE COVERS THE EXACT BYTES Model output fingerprint 3f9a → Resized fingerprint 8d04 → Verification no match, fails STEPS THAT WRITE A NEW FILE Resize Compress Re-encode Add a logo CDN optimisation Each of these changes the fingerprint the signature was made against.

A Content Credential is signed against the exact bytes of the file. That signature is what makes it trustworthy: nobody can move a credential from one file to another or edit the content without the mismatch showing. It also means any step that rewrites the file breaks the signature, even when nothing visible changes.

Production teams rewrite files at almost every step. Resizing an image for each placement, compressing it for the web, re-encoding a video for a platform's specification, adding a logo or a caption, and the automatic optimisation many content delivery networks apply all produce a new file. After any one of them the credential either disappears or no longer verifies. The asset still looks exactly right, and nothing in a normal review shows that the label has gone.

What we found while building Toolkit

One video, checked at three stages Stage one, the original from the model: credential found, signature valid. Stage two, after the step that lets a video start playing before it has downloaded: no credential, the label is gone. Stage three, the original credential copied back into the processed file: credential found, but the signature fails because the file no longer matches it. ONE VIDEO, CHECKED AT THREE STAGES 1 · ORIGINAL Credential found signature valid → 2 · PROCESSED No credential the label is gone → 3 · COPIED BACK Credential found signature fails The step that lets a video play before it has downloaded removed the label. Copying the label back left a signature that no longer matches the file.

While developing our own pipeline we tested the files at every stage, from the model's output to the asset a client downloads. One routine step, which reorders a video file so it starts playing in a browser before it has fully downloaded, removed the credential entirely. The video was identical to the eye and had lost its AI label.

The obvious repair does not work. Copying the original credential back into the processed file fails verification, because the signature covers the original byte layout. A credential that fails verification is worse than none: it tells a platform or a regulator that the file was altered after signing.

So we built signing into Toolkit itself, at every step a file goes through on the platform.

How Toolkit signs every asset

What a Toolkit credential records An example credential on a finished asset. It is signed by Instant Studio with c2patool and records the edits made, for example resizing to 1080 by 1350, adding a logo and re-encoding for Instagram. Inside it, the model's original is listed as the ingredient: kept byte for byte with the model's own signed credential, created by an AI model and signed by the model provider. WHAT A TOOLKIT CREDENTIAL RECORDS Finished asset Signed by Instant Studio with c2patool EDITS RECORDED Resized to 1080 × 1350, logo added, re-encoded for Instagram INGREDIENT: THE MODEL'S ORIGINAL Kept byte for byte, with the model's own signed credential Created by an AI model and signed by the model provider Anyone can check each link, from the model's output to the file you publish.

Toolkit handles Content Credentials for every image and video produced on the platform:

  • The model's original is kept as generated. We store the signed file each model returns without changing a byte, so the model's own credential always remains verifiable.
  • Every finished asset carries our credential. When a Creative Workflow resizes, re-encodes, adds a logo or adapts an asset for a market, Toolkit signs the result with Instant Studio's own C2PA credential using c2patool, the open-source tool maintained by the C2PA community.
  • Our credential points back to the source. Each signed asset lists the model's original as its ingredient and records the edits made to it, so the chain from generation to the published file can be checked link by link.
  • It applies to images and video alike, including models that do not write a credential of their own. Those assets still leave Toolkit labelled as AI-generated.

For a brand this means every asset a Creative Workflow produces arrives with a verifiable record of how it was made, and your team does not have to remember to add it.

Why it matters to legal and compliance teams

Where the label gets read Four places that read a Content Credential. LinkedIn shows the credential to people who view the post. TikTok reads it to apply its AI-generated tag. Article 50 of the EU AI Act requires a machine-readable mark on AI images and video. Anyone can inspect a file at contentcredentials.org/verify. Without a credential, the label depends on someone disclosing it by hand. WHERE THE LABEL GETS READ LinkedIn Shows the credential to people who view the post TikTok Reads it to apply its AI-generated tag EU AI Act, Article 50 Requires a machine-readable mark on AI images and video Anyone Can inspect a file at contentcredentials.org/verify Without a credential, the label depends on someone disclosing it by hand.

Regulators are moving from asking for disclosure to requiring it in a form machines can read. The EU AI Act's transparency rules, in Article 50, require AI-generated images and video to be marked in a machine-readable way, and C2PA is the standard the major platforms have adopted to read that mark. Platforms increasingly apply their own AI labels automatically when they find a credential. When they cannot find one, the brand is relying on manual disclosure and on every person in the chain remembering to make it.

A verifiable credential also protects the brand in the other direction. If an image is disputed, the signed record shows which model produced it, when, and what was changed before it was published.

Five questions to ask any AI content supplier

The fastest test Three steps. Pick an asset from your last AI campaign. Upload it to the verify page at contentcredentials.org/verify. Read the result: who signed it and what it lists as its source. A file that passes names its signer and lists the model's original as its source. THE FASTEST TEST STEP 1 Pick an asset from your last AI campaign → STEP 2 Upload it to the verify page at contentcredentials.org → STEP 3 Read the result who signed it and what its source is A file that passes names its signer and lists the model's original as its source.

Use these with us, with your agency or with any production partner:

  1. Does the finished file, the one you will publish, carry a Content Credential that verifies at contentcredentials.org/verify?
  2. Who signs it: the model provider only, or the supplier as well?
  3. Does the credential list the model's original output as its source?
  4. Does it survive every format you publish in, including the video versions?
  5. When a model writes no credential of its own, does the asset still leave labelled as AI-generated?

The fastest test is the first one. Take an asset from your last AI campaign, upload it to the verify page and see what it says.

Adeh DeSandies is co-founder and CTO of Instant Studio. Adeh DeSandies on LinkedIn

Let's get started

Ship AI content with its credentials intact

As a Contributor member of the C2PA, we sign every finished image and video on Toolkit to the standard. Get in touch to see how we can enable it for your team.

Start a fixed-price pilot Talk to our team