The Pre-Upload Checklist Every AI Creator Needs

By Calabi Labs Editorial Team · 2026-08-05

The creators who win with AI content aren't the ones with the best prompts — they're the ones who never upload a raw export. A two-minute file-level QA pass, done every single time, is the difference between content that competes and content that gets silently capped. Here's the complete checklist.

Think of this like a pilot's pre-flight walkthrough. You don't skip it because you're in a hurry — you run it because skipping it is exactly when things go wrong. Every fresh generator export re-introduces the same fingerprints, so this isn't a one-time cleanup; it's a standing habit.

Step 1: Clear the provenance layer

Start with the biggest tell. Modern generators often embed C2PA / Content Credentials — a signed manifest (JUMBF atoms) documenting the file as AI-made. It's invisible in any player and survives naive cleanup, so it's the thing most creators ship by accident.

Confirm no C2PA / Content Credentials manifest is embedded
Confirm the JUMBF atoms and C2PA references are actually gone — not reduced, gone

This is non-negotiable. If the manifest survives, the rest barely matters.

Step 2: Strip the explicit AI flags

Next, the literal labels. XMP fields like trainedAlgorithmicMedia flatly declare the file synthetic, and generator tags name the tool outright.

Remove trainedAlgorithmicMedia and any AI-source flags
Remove generator/model tags
Clear app name and version from Software and comment/description fields
Check digital-source-type fields for synthetic-origin markers

Step 3: Fix the encoder fingerprint

Now the part almost everyone forgets. Even a metadata-clean video can carry an encoder signature — Lavc, x264 SEI data — that reads as server-rendered rather than phone-captured.

Confirm the encoder signature reads like a real device, not a render pipeline
Make sure your "casual footage" isn't encoded like a render farm — that contradiction is a flag on its own

Step 4: Inject authentic, consistent identity

Don't stop at empty. A blank file is suspicious too — real captures are dense with identity. Replace synthetic signals with a coherent phone-capture profile.

Make, Model, and Software present and from the same real device family
GPS coordinates present and plausible
Capture timestamp present and believable (and not identical across a whole batch)
Timezone consistent with the GPS location
Device model not newer or older than it should be for the encoder
The whole set internally consistent — every field agrees with every other

Incoherent identity (iPhone Make, Android software, wrong-hemisphere GPS) is worse than none.

Step 5: Verify with a forensic tool

The step that separates guessing from knowing. Don't trust that the previous steps worked — confirm it.

Run the file through ExifTool (or an equivalent inspector)
Confirm AI atoms are at zero
Confirm the tag count looks phone-normal — rich, not scrubbed near-empty
Compare against a real phone file if you want a reference point

"I ran a cleanup" is a hope. "ExifTool shows the atoms at zero and a phone-normal tag profile" is proof.

The file-level checklist (the whole thing, at a glance)

[ ] No C2PA / Content Credentials manifest (atoms verified at zero)
[ ] No trainedAlgorithmicMedia or AI-source flags
[ ] No generator/model/app name in Software or comment fields
[ ] Encoder signature reads like a device, not Lavc/x264
[ ] Make/Model/Software present and from one consistent device family
[ ] GPS + timezone + timestamp all mutually plausible
[ ] Metadata looks phone-rich, not scrubbed empty
[ ] ExifTool-verified — you looked, you didn't assume

FAQ

How long should this actually take?

By hand, a careful pass runs a few minutes per file once you know what to look for — longer if you're cross-checking device profiles for consistency. That's why most serious creators automate it: same rigor, seconds instead of minutes, and no human slip-ups on the consistency checks.

Do I really need to verify with ExifTool every time?

If reach matters, yes. Cleanup tools fail silently — a converter can leave the C2PA manifest untouched while looking like it "worked." Verification is the only way to know. At minimum, spot-check regularly so you catch a tool that quietly stopped stripping something.

What can this checklist NOT fix?

Anything in the pixels. Visible watermarks and logos only come off by cropping (which costs framing), and invisible pixel-level watermarks are inconsistent — results vary by platform and source model, and no honest tool guarantees removal. This checklist handles the file-level signals that cause false-positive flags, nothing more.

Calabi Sanitizer runs this file-level cleanup automatically — try it free at calabilabs.com.