A 99% AI image score is an output from a particular detection system. It is not a measurement that 99% of the picture was made by AI. Whether it can be interpreted as a probability for your use case depends on how the system defines and validates that score.
Before acting on the number, identify the detector, the exact file, the decision threshold and the population on which the score was evaluated.
A score is not a label
Many systems produce a continuous value and then compare it with a cutoff to assign a label. Change the cutoff and the label can change while the score stays fixed.
Hypothetical classifier example. These numbers are not Hive or Sightengine results.
Score
What this means
A cutoff converts a continuous score into a decision.
Different policy
What this means
The score did not change. The decision rule did.
Calibration
What this means
Only a suitable calibration study can connect score values to observed proportions in a population.
The example uses hypothetical values. It illustrates the relationship without claiming to reproduce a vendor’s policy. Real systems may also use multiple scores, uncertain regions or additional rules.
| Term | Plain meaning | Common misreading |
|---|---|---|
| Score | A numerical output of the model | Percentage of generated pixels |
| Threshold | A rule for turning a score into a decision | A universal boundary between real and fake |
| Confidence or probability label | The vendor’s interpretation of its output | Guaranteed correctness on any image |
| Calibration | Agreement between scores and observed frequencies under a defined evaluation | Proof of accuracy in every new setting |
| Accuracy | Performance over a specified labelled sample | Certainty about this one file |
When is a probability interpretation justified?
Imagine many independently labelled images receiving scores near 0.8. If about eight out of ten belong to the target class under the evaluation’s definition, that supports calibration around that score on that population. It does not automatically transfer to a new generator, a new editing workflow or a different mix of camera and generated images.
This is a statistical explanation, not a new calibration result. Without a suitable evaluation, it is safer to report the displayed score and label directly than to translate them into personal certainty about an image’s origin.
Vendor documentation still matters. Sightengine, for example, documents its global AI-generated score and other outputs. Read the named field’s definition rather than assuming that every percentage in a response refers to the same task.
Why two vendors’ numbers should not be averaged
A 20% score from one classifier and an 80% score from another do not necessarily produce a meaningful 50% conclusion. The models can use different training data, scales, thresholds and definitions.
The disagreement itself is useful. Record both results, check the exact file submitted and ask whether each model is designed for the same task. A deepfake or face-manipulation score is not interchangeable with a whole-image generation score.
In the selected Deprint portrait check, Sightengine’s GenAI result and its separate face-manipulation result were recorded separately. Combining them would obscure what each measurement meant. The methods preserve that distinction.
Why 99.9% to 90.9% may still be a failure for your goal
A score can fall substantially while the system’s label stays positive. In Deprint’s fixed-chain desk example, Hive moved from 99.9% to 90.9% and remained positive. That is a score change, not a successful negative classification.
Conversely, a small score movement near a cutoff can flip a label. Reporting only “passed” can hide how close the output was to the decision boundary. Report the number and label when available, along with the tool and date.
Also avoid language such as “removed 9% of the AI.” The difference between two classifier scores is not an amount of generated content removed from the image.
What to do with a surprising result
Confirm that the uploaded file is the one you intended. Check for an accidental thumbnail, screenshot, altered export or a different processing stage. Record dimensions and a hash when exact comparison matters.
Then inspect origin evidence independently: the original capture or generation record, editing history and any relevant credentials. A detector score can inform a decision, but it cannot retroactively create missing source history.
If evaluating a treatment, compare with a baseline and a matched export control. If evaluating a detector, use a labelled sample that reflects the images you actually expect, including difficult real photographs and edited generated images.
How to report the number responsibly
Write “the named detector displayed 99% on this file on this date,” then explain its label and the test’s scope. For a collection, provide denominators and error counts. Keep failures, uncertain responses and missing checks visible.
The aim is not to discard detector scores. It is to use them at the level the evidence supports. The detector disagreement guide shows why this discipline matters when several tools inspect the same image.