Interpretation guide
How to Interpret AI Detector Scores
Short answer: an AI detector score is a directional output from a particular model, version and threshold applied to a particular file. Unless a provider publishes calibrated evidence to the contrary, “80” should not be read as an 80% probability that AI made the image, and it says nothing by itself about authorship or intent.
Read four fields together
| Field | Question it answers | Why it matters |
|---|---|---|
| Signal strength | How strongly did this model respond? | The scale belongs to that detector; it is not automatically a probability. |
| Threshold state | Did the score cross the configured decision boundary? | A nearby score can change labels after calibration or a version update. |
| Model version | Which detector produced the output? | Outputs from different versions are not safely interchangeable. |
| Evidence scope | Which image, page, frame or clip was checked? | A selected unit does not justify a verdict about the complete document or recording. |
Why two tools can disagree
Detectors can use different training sources, generator coverage, transformations, preprocessing and thresholds. A screenshot may reach one service at a different size or compression level than another. Disagreement is therefore evidence about model uncertainty, not a reason to average the scores into a new percentage.
What Made It? keeps available model channels separate and requires stronger independent agreement for a strong Beta result. A single channel is presented as a lower-confidence review indicator. The current release status is documented on the Research page.
Decision rules for real-world use
- Do not convert a score into a statement about a person’s conduct or intent.
- Check whether the exact generator, camera, edit and compression conditions were represented in evaluation.
- Prefer source-family and transformation-slice results over one aggregate accuracy number.
- Preserve the model version, threshold, file hash and date with any exported report.
- Require human review and independent corroboration for moderation, legal, employment, financial or safety decisions.