Methodology · How we classify
Why we don't use AI detection
DisclosureGuard never runs an AI detector on your content. No confidence scores, no "87% likely AI-generated" verdicts. This page explains the engineering and legal reasoning: why detection-guessing is the wrong evidence for compliance, and what we do instead.
What an AI detector actually outputs
An AI detector takes pixels or audio and returns a guess: a probability-like score that the content is machine-generated. These scores are not calibrated probabilities, they shift between tools and model versions, and they degrade as generators improve. A score is a statistical opinion about the past — it says "content like this was often AI-made," not "this content is AI-made."
The false-negative liability
Now put that guess inside a compliance workflow. A detector scores an AI-generated product photo as "human" with 92% confidence. The merchant, trusting the number, ships it without a disclosure. That asset is now an undisclosed AI-generated publication — exactly the violation Article 50 fines. The merchant never sees it, because the tool's job was to detect, and the tool was confident and wrong.
This is the core problem: a detection-based workflow converts detector error into legal exposure, silently. The failure mode is invisible by design — you cannot audit what the detector decided not to flag. And the fine lands on the publisher, not on the detector vendor.
The false-positive cost
The error cuts the other way too. A detector flags a photographer's genuinely shot campaign as "AI-generated," the team labels it anyway, and the brand tells its audience its real work is synthetic. Trust damage, wasted review cycles, and a disclosure record that now asserts something untrue. Guessing wrong in either direction has a cost; compliance tooling should not introduce new ways to be wrong.
What regulators actually ask for
Article 50 of the EU AI Act (transparency obligations in force since 2 August 2026) requires machine-readable marking of synthetic content and visible disclosure of its AI origin to recipients. China's labeling measures (effective 1 September 2025) require explicit visible labels plus implicit machine markings. Google Ads requires synthetic-content declarations in the ad.
None of them ask "what did the detector say?" Every regime asks for demonstrable actions taken by the publisher: markings embedded, disclosures shown, records kept. A confidence score is not a marking, not a disclosure, and not a record. It is not evidence a regulator can inspect — it is an opinion that evaporates the moment it is rendered.
What we do instead: read declared provenance
DisclosureGuard reads what the asset declares about itself, deterministically — the same bytes every time, no model, no guessing:
- C2PA content-credentials manifests — signed claims about the asset's creation and edit history, verified cryptographically where the manifest is intact.
- XMP packets and EXIF fields — generator identifiers and creation metadata embedded at export.
- Generator text indicators — known, declared markers emitted by generation tooling.
- Frame-level video scans — provenance reads applied across video frames, not just a thumbnail.
The compliance score is then computed against versioned jurisdiction rules — EU Art. 50, US/Google Ads, China — measuring disclosure completeness: is visible disclosure present, is machine-readable marking present, does the combination satisfy the rule pack? A score of 100 means "every required disclosure element is demonstrably present." A low score means "required elements are missing" — and it names them.
When provenance is absent, we say so
No provenance and no visible label means we report exactly that: "no declared provenance found." We do not invent a verdict. An unknown is flagged as a review item for the publisher — who usually knows whether the asset is AI-made — rather than resolved by a guess. If you know an asset is AI-generated, label it; no tool can know what nobody declared.
The honest limitation
Deterministic provenance reading has a boundary: a fully stripped, unlabeled asset with no declared provenance cannot be identified as AI-made by reading. We consider that a property of the method, not a bug — it is the same boundary regulators drew when they required marking and disclosure rather than detection. Our answer to stripped assets is the operational checklist (preserve provenance, label at first exposure, re-verify after publishing) and the Drift Monitor that catches labels dying after publication — not a guesser that papers over the gap.
Scan an asset — see deterministic scoring Compare detection vs provenance →