How our AI evidence checks work

What each kind of evidence means

• Signed creation records: we validate C2PA records and report the result. A valid signature proves the integrity of a signed record; the claims inside are the signer’s claims, and we show them as such.
• Saved image details: we read metadata, AI-tool declarations and saved prompts exactly as stored. These details are editable, so the report shows them as statements about a file.
• Statements of AI use: we find statements of AI use in pasted text, code and public pages and show the matching words. A mention of an AI tool is reported separately from a declaration that the content was generated with it.
• Matching copies: we match the file against copies in our records and surface what earlier checks found. A match links copies; our records are a curated corpus, distinct from a search of every image on the internet.

A webpage check covers the material we retrieve and inspect, pinned to that page; it describes that resource, not every author on the domain.

What “No AI evidence found” means

The completed checks read the copy and found no AI evidence: a verified clean copy. It is reported as exactly that, apart from “verified human”, because screenshots, compression, editing and reposting strip creation details. A check that did not run is listed separately from a completed check with a clean result.

Open “How we checked” in the report to see the evidence and which checks ran. For an image, use the original download and compare it with any reposted version. Our image-checking guide walks through the process.

How image quality is measured

Image reports include eight descriptive pixel measurements beside the origin evidence. We correct orientation and sample the first frame at a maximum of 512 × 512 pixels in sRGB. Brightness uses 0.2126 red + 0.7152 green + 0.0722 blue on 8-bit values. Transparency weights brightness statistics and adjacent-pixel edge differences. Near-black means brightness at most 8; near-white means at least 247. Fully transparent samples have no measurable brightness.

Dimensions describe the original frame after orientation correction. Other measurements describe the sampled pixels. Compare the same scene at similar dimensions; texture, artistic choices and resizing move these values. They describe the pixels; the origin answer comes from the evidence.

How accurate is this AI checker?

Text and code checks read AI disclosures and tool mentions and show the exact words, so every finding is verifiable by reading it. They are built for traceable evidence, distinct from a writing-style authorship verdict. Statistical image detection is kept separate from record reading: a provider must pass our release requirements before its estimates can touch the public result, and each report lists which checks ran.

Every result separates signed records, editable labels, model estimates and pixel measurements, so you always know which kind of evidence you are reading. Inspect the finding and the completed checks, then repeat the file-reading examples below. We publish file-behavior results we can reproduce rather than a single blended accuracy percentage across checks that measure different things.

Our hypothetical score-interpretation example shows why 90% evaluation accuracy can coexist with only 50% precision among flagged images, and separates those quantities from confidence scores, edited-pixel area and provenance.

Repeat the file-reading checks

The Image Metadata Lab provides ten controlled, code-drawn PNG files across seven worked cases. Each case includes source files, hashes, observed reader output and scope notes, so you can reproduce missing, changed and conflicting creation records yourself instead of taking a claim on trust. These examples validate specific file behavior, reproducibly.

How we evaluate detector changes

Our evaluation process separates tuning examples from held-out evaluation. Related originals, duplicates and transformed copies stay in the same source group. We measure AI recall, false positives on human content, abstentions, coverage, errors, latency and cost separately for each input type.

A proposed detector must pass the configured release gates and a separate shadow evaluation before it replaces an incumbent. Held-out evaluation data stays held out, because reselecting against one test set would weaken it as independent evidence, and a failed or incomplete candidate never ships as an improvement.

User feedback tells us which cases to investigate. It never becomes a verified origin label or training data on its own. Our privacy policy explains how submitted content and check records are handled.

Sources and corrections

The C2PA explainer describes signed provenance and its scope. The RAID benchmark paper motivates testing AI-text detectors across generators, domains and transformations; its results describe that benchmark, not isGenAI.com.

Send supporting context through our correction process. Our separate EU AI Act Article 50 ratings methodology covers public transparency evidence, a different question from AI authorship detection.