Checker methodology · Reviewed 9 September 2026
How our AI detector works
isGenAI.com checks images, public webpages, text and code for evidence of AI use. It inspects disclosures, saved creation details, signed creation records and matching copies in its records. A finding describes the evidence available to the checks that ran. Missing evidence does not establish human authorship.
What each kind of evidence means
- Signed creation records: inspect supported C2PA records and their validation results. A valid signature can establish the integrity of a signed record; it does not prove that every claim in it is true or that the depicted event happened.
- Saved image details: read supported metadata, AI-tool declarations and saved prompts. These details are editable. They are statements about a file, not independently verified authorship.
- Statements of AI use: look for statements of AI use in pasted text, code and public pages. A mention of an AI tool needs context; it is different from a declaration that the content was generated with it.
- Matching copies: use copies in our records to recover related findings. A match does not itself prove AI origin, and our records are not a search of every image on the internet.
A webpage check covers the material we can retrieve and inspect. It cannot determine who authored an entire website.
What “No signs of AI found” means
The completed checks found no supporting signs of AI. It does not mean “verified human.” Screenshots, compression, editing and reposting can remove creation details. A check that could not run is also different from a completed check that found nothing.
Open “How we checked” in the report to review the evidence and check availability. For an image, use the original download where possible and compare it with any reposted version. Our image-checking guide explains this process.
Do we measure AI authorship from writing style?
Current text and code checks inspect AI disclosures and tool mentions. They do not provide a validated writing-style authorship verdict. Statistical image estimates depend on an available, configured provider; the report identifies which checks ran.
We have not established a public accuracy percentage for isGenAI.com across images, text and code. A model integration, a successful software test or a detector’s self-reported confidence is not an accuracy benchmark. We do not claim to outperform another detector without comparable independent measurements.
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 examine 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 can replace an incumbent. Repeatedly selecting a model against the same test set weakens that set as independent evidence. A failed or incomplete candidate is not a measured improvement.
User feedback helps identify cases to investigate. It does not automatically become a verified origin label or training data. Our privacy policy explains how submitted content and check records are handled.
Sources and corrections
The C2PA explainer describes signed provenance and its limits. The RAID benchmark paper motivates testing AI-text detectors across generators, domains and transformations; its results are not isGenAI.com performance measurements.
Send supporting context through our correction process. Our separate EU AI Act Article 50 ratings methodology concerns public transparency evidence, rather than AI authorship detection.