Each signal has a narrow meaning
Visual-model output can identify patterns that are consistent with AI generation, but it is not a probability of authorship. Metadata can describe an available file history, but can be removed, changed or absent. Content Credentials can provide provenance when present and valid, but their absence does not prove a file was AI-generated or deceptive.
Where the available evidence is missing, weak or conflicting, the appropriate result is inconclusive. A strong signal is still evidence to investigate, not proof.
Missing metadata is not a verdict
Platforms, editors, screenshots, exports and recompression often remove metadata and Content Credentials. A screenshot can also change pixels and discard provenance information. Treat absence as an absence of that signal, not as evidence about who made an image.
Known failure modes
Repeated compression, resizing, cropping, screenshots, subtle generative edits and generator families not represented in evaluation can weaken or change detection signals. New generators, editing tools and distribution platforms can create failure modes that have not yet been measured. A Beta detector can change as release controls, providers or model versions change.
The service does not publish an independently validated public accuracy percentage. Research and release status explains the evidence needed before such a claim could be made.
Do not use one result as the sole basis for a high-impact decision
Do not rely on a result alone for employment, education discipline, credit, insurance, legal, moderation, safety or other decisions that materially affect a person. Seek source context, corroborating evidence and appropriate human review.
Use the result as a starting point
Read the visible evidence, compare it with source context, and preserve original files where possible. Methodology explains how the service separates signals. Privacy explains processing. Questions, corrections and security concerns can be sent to support@whatmadeit.com.