How the assessment is built
What Made It? evaluates three evidence channels independently. A result describes patterns present in the submitted file; it does not prove who created the image or why it was made.
The visual channel looks for local texture, edge and consistency patterns. File evidence reads container structure, encoding and available metadata. Provenance checks look for signed Content Credentials. A missing metadata field or missing credentials manifest is not treated as proof of AI generation.
A result can identify technical signals that are consistent with AI generation, editing or camera capture. Each available model channel is shown as a directional signal strength from 0 to 100, together with its threshold state and model version. That number is not the probability that a file was AI-generated. A result cannot establish authorship, intent, copyright ownership or legal admissibility.
The controlled Beta distinguishes one-channel indicators from stronger independent agreement. It returns Inconclusive when model channels are unavailable, the file provides no usable directional score, or trusted provenance conflicts with a model signal. A formal production accuracy claim still requires the stricter frozen-threshold, transformation-robustness and untouched-holdout gates.
The explanation applies to the submitted image and keeps each model or provenance channel separate.
Each selected rendered PDF page or selected video frame receives its own image report. Collection pages remain categorical and never turn selected pages or frames into a whole-file score.
Each provider chunk within the explicitly selected clip receives its own directional explanation. Chunk scores are not averaged into a probability for the clip or the complete recording.
Visual model
LimitHighlights supporting regions—not a generator guess.
File evidence
LimitCan support a result, but absence proves nothing.
Provenance
LimitStrong positive evidence; many valid images have none.
Decision boundary
When available evidence conflicts or stays below a calibrated release threshold, the system stops instead of manufacturing certainty.
Methodology reviewed by the What Made It? product and methodology team · Last updated