Use AB 853 Detection Tools to Audit Exports
AB 853 forces large GenAI systems to publish provenance detection tools. Use them to QA your own exports and verify what an agency actually delivered.
The useful thing in California's AB 853 isn't a duty you have to discharge. It's a tool you get handed.
AB 853 became operative on 2 August 2026. Among its requirements: generative AI systems above one million monthly users must offer an AI detection tool, at no charge to the person using it, that surfaces provenance data for a piece of content. The intended audience is anyone who needs to check a file. Nothing stops that being you, checking your own work before it ships, or checking what a contractor actually delivered.
That's the first time the disclosure conversation has had a verification step in it. Everything before this was assertion: your editor claims it preserves credentials, your export settings claim to carry metadata, your freelancer claims the footage was generated on the model in the invoice. Now some of that is checkable.
Scope it precisely before you rely on it, though. The obligation is on generative systems above the user threshold, and each tool reports on that system's own provenance signals. It is not a universal AI detector. Feeding a file into one provider's tool tells you whether that provider's signals are present, not whether the file is synthetic in general.
A second phase, from 1 January 2027, extends duties to large online platforms and to GenAI hosting platforms. That's the part with distribution consequences, and it's the reason to build the habit now rather than in December.
What these tools can and can't tell you
| Question | Answerable | Notes |
|---|---|---|
| Does this file carry provenance data? | Yes | The core case, and the one that matters most for QA |
| What does the manifest say was done to it? | Usually | Depends on whether the chain recorded edits or only origin |
| Is this provider's watermark detectable in the pixels? | Often | More durable than metadata, since it survives re-encoding |
| Is this file AI-generated at all? | No | Absence of a signal is not evidence of absence |
| Which of the many available models made it? | No | Each tool speaks for its own system only |
That fourth row is the one to keep in front of you. A clean result means "this provider's signals aren't detectable here," which has at least three innocent explanations before you get to a dishonest one. Treating a null result as proof is the mistake that turns a useful QA step into an accusation you can't support. The reliability differences between metadata, embedded watermarks and classifier guessing are covered properly in how AI content actually gets identified.
Audit one: QA your own exports
This is the boring, high-value use. Five steps, run on the file you're about to publish.
1. Audit the final export, not the preview and not the project. The distinction matters because the failure mode you're hunting for lives in the export step. A preview pass is for checking the cut. If you're working in an EDL-based editor, the 480p preview is free and carries a short per-user cooldown, and the final export is charged once regardless of how many clips are on the timeline. Audit the thing you'll actually upload.
2. Run it through the detection tool of the system that generated the source. If a video came out of one model and the music came out of another, that's two checks, not one. Bookmark the tools for the systems you actually use rather than looking them up each time.
3. Read the whole chain, not just the origin. A manifest that records "generated by X" but nothing after it means your editing step didn't write into the chain. That's a real finding: it tells you the record is incomplete, which is a different problem from the record being absent.
4. Re-run after anything that touches the file. Compression, format conversion, a round trip through a messaging app, an upload-and-download from cloud storage. Each of those is a place the manifest can vanish, and the whole point of a check is to catch it after the step that breaks it. The full list of what survives and what doesn't is in sign, strip, survive.
5. Record the result alongside the deliverable. A screenshot in the project folder, dated. This costs ten seconds and is the entire difference between having a provenance position and remembering that you probably checked.
Fold steps one through four into whatever pre-publish check you already run. It belongs in the same pass as aspect ratio, caption legibility and audio levels, not in a separate compliance ritual that gets skipped when a deadline moves.
Audit two: verify what a freelancer or agency delivered
This is the use nobody talks about, and it's worth more money than the first one.
When you commission video, you're accepting claims you have historically had no way to test. Four of them are now partially checkable:
Was it generated at all? If the brief specified shot footage and the deliverable carries a generation manifest, that's a conversation. Note the asymmetry: presence of a manifest is informative, absence is not.
Was it generated on what you were billed for? Manifests name the originating system. If the invoice says one model and the file says another, you have a specific question rather than a vague suspicion.
Did the chain survive delivery? A deliverable that arrives with no provenance data at all, from a shop that generated it on a signing model, means someone re-encoded it somewhere. Usually innocently. Occasionally not.
Does the edit history match the story? A manifest recording three edit steps against a revision log describing eleven rounds is worth asking about.
A delivery spec you can put in a contract. Three clauses, all cheap for an honest supplier to meet:
- Final exports are delivered with provenance data intact, and the supplier confirms this by running the originating system's detection tool and attaching the output.
- Source generations are delivered alongside the finished cut, not just the export.
- Any deliberate stripping of provenance data requires written notice and a reason.
A supplier who pushes back hard on clause two is telling you something. A supplier who can't meet clause one usually has a pipeline problem rather than an integrity problem, and it's better to find that out at contract stage.
When the tool returns nothing
Rank the explanations before you act on any of them.
- The generating system is below the threshold or doesn't sign output. Most likely, and completely benign. Plenty of capable models publish nothing.
- The chain broke on a re-encode. Second most likely. Ask which step, then test it.
- The signal is in the pixels rather than the metadata, and this tool doesn't read it. Provider watermarks and metadata manifests are different layers with different durability, which is exactly why the EU's own position is that no single technique currently meets the machine-readable marking standard and a layered approach is expected. The mechanics of the pixel-level layer are covered in how invisible watermarking works and where it stops.
- The file wasn't made the way you were told. Last, not first.
The correct escalation is never "explain this null result." It's "send me the source generation." That request resolves the question in one step and is unreasonable to refuse.
Build it in once
Detection tools are only useful as a habit. Set it up as a fixed step:
- A bookmarks folder with the detection tool for each system in your stack.
- A rule that the check runs on the export, after the last re-encode, before the upload.
- A folder convention where the check result lives next to the file.
- A contract clause for anything you commission.
None of that is a compliance programme. It's a QA step that happens to also produce the evidence you'd want if the AI label on one of your uploads ever needed explaining, or if a client asked what your disclosure position was. The habit is the deliverable.
FAQ
Can I use these tools to check whether a competitor's video is AI-generated?
Partially, and with real limits. A positive result from a specific provider's tool is meaningful. A null result across several tools is not proof of anything, because the file may have come from a system below the threshold, from a model that doesn't sign, or through a re-encode that stripped the record. Public claims built on null results are how people end up wrong in an embarrassing way.
Does the detection tool tell me if someone edited an AI video to hide something?
It can show you the recorded edit chain, which sometimes reveals steps the person describing the work didn't mention. It can't reconstruct edits that were made outside a credential-preserving tool, and it can't tell you intent. Treat it as a source of specific questions rather than as an answer.
Do I need to run this on every single upload?
On anything commissioned, anything client-facing, and anything running as paid media, yes. On a daily short-form post from a pipeline you've already verified end to end, spot-checking after any pipeline change is enough. What you're really validating is the pipeline, not each individual file, so re-check whenever a tool, an export setting or a delivery route changes.
What if my main model isn't covered by AB 853?
Then there's no detection tool for it, and you fall back to what your own watermark and metadata handling can tell you. That's a reason to know which of your tools sign output and which don't, and to keep source generations for anything commercially important. The archive is the fallback when the automated check isn't available.