Core concepts
IA Checker separates the headline probability from the confidence of the reading and from direct source evidence. Those concepts answer different questions and should not be collapsed into one label.
Probability and confidence are different
AI probability describes how the available signals lean for this input. Confidence describes how much usable evidence the system had, based on factors such as text length, sentence count, language calibration, image quality and preserved source data.
A high probability with limited evidence deserves a closer look. A balanced result with rich evidence can still be useful because it explains which checks agreed and which did not.
Multi-signal analysis
No single surface pattern works across every generator, language, edit and compression path. IA Checker therefore combines several signal families and keeps the strongest contributors visible in the report.
| Signal family | Typical question |
|---|---|
| Text structure | Does the cadence, phrasing and discourse structure resemble generated writing? |
| Visual models | Do semantic and pixel patterns align with generated imagery? |
| Forensics | What do noise, edges, residuals and compression reveal about the file? |
| Provenance | Does the file contain signed or embedded source information? |
| Visible labels | Is an explicit AI label or provider mark visible in the content? |
Detection versus provenance
Detection infers from patterns in the submitted content. Provenance reads source information such as C2PA credentials or metadata. A provenance claim can be direct evidence about the file history, while its absence is simply an absence of that claim.