Is AI Generated Image Detection Reliable? Scores, False Positives, and Human Review
An AI image detection score is a review signal, not proof of origin. Learn what pixel classifiers measure, why false positives matter, and how to review compressed, cropped, or partly edited images.
·
AI generated image detection can help identify images that deserve closer review. A high score, however, does not establish how an image was made. Before labeling content, restricting its distribution, or penalizing an account, you need to know what the detector analyzed and what evidence exists beyond its output.
The available documentation explains one provider’s approach. It does not establish independently measured accuracy or a universal enforcement threshold. A suspicious classification and a verified origin are different findings.
A score classifies an input; it does not reconstruct its history
Sightengine’s documentation says its model analyzes image pixels without using metadata. It describes ai_generated as a confidence score between 0 and 1 that an image was produced or edited using generative AI.
That scope matters. A score may relate to a fully generated image or to generative editing within a photograph. By itself, it does not identify the edited region, the operator, the time of creation, or the publishing context. The provider also describes generator-specific scores, but these remain classification outputs—not original records of the generation process.
Do not automatically read 0.9 as a verified 90% probability that this particular image is AI-generated. That interpretation would require evidence about calibration on data comparable to your own. The supplied material does not provide that validation.
| Signal | What it can help assess | What it cannot establish alone |
|---|---|---|
| Pixel-classifier score | Whether an image has features the model associates with AI | Its creator, complete production history, or the truth of the depicted event |
| Checkable original files, publication records, and edit history | Where a file came from and how versions relate | Whether an individual record is authentic and complete |
| Visible texture, text, or structural anomalies | Which details warrant further investigation | That AI caused the anomaly, or that an image without anomalies is authentic |
Source records also need verification. A tool name written into a file is not enough to establish its history. Conversely, missing source records do not establish AI generation.
Why false positives can matter even when a rate looks small
Consider a hypothetical example—not a product benchmark. A batch contains 1,000 images, of which 100 are genuinely AI-generated. Suppose a detector flags 90 of those 100, but also flags 45 of the 900 non-AI images.
There are now 135 flagged images, and 45 of them are false positives: one-third of the flagged set. The false-positive rate among non-AI images is 45 divided by 900, or 5%. That is a different quantity from the share of flagged results that are wrong.
The example shows why the prevalence of AI images in your review queue matters alongside detector performance. A low score is not an authenticity certificate either: failing to detect generation is not the same as establishing camera capture.
Transformations change the question being tested
Sightengine states that its detector is designed for redistribution artifacts such as compression, resizing, and re-encoding. It also notes that heavy degradation or screenshots with substantial interface elements may reduce confidence. These are the vendor’s claims about its own model, not a performance guarantee for all detectors.
| Situation | Evidence gap | Useful follow-up |
|---|---|---|
| Repeated compression, screenshots, or reposting | The tested derivative may not preserve all characteristics of the original | Request the earliest available file and keep versions and results separate |
| Cropping | The input covers a different area; its score says nothing about excluded content | Retain the full frame and document the tested region |
| Generative edits within a photograph | A whole-image score does not locate or quantify the edit | Compare before-and-after files and edit records; do not automatically call the entire image fabricated |
| A new generator or model version | A supported-generator list does not establish performance on new outputs | Check detector version and coverage information; retain uncertainty where validation is missing |
Do not overwrite the received file to make it “easier to detect.” If you test crops or recompressed copies, preserve the original and record each transformation. Scores from different versions are not repeat measurements of the same input.
Make human review traceable
The following is a recommended review method, not a tested detection protocol.
Save the received file, source URL, acquisition time, and accompanying claims. Keep original files separate from screenshots and reposted versions when available. Record the detection service, analysis time, visible model version, raw response, and exact file submitted. A file hash can help confirm that a file has not changed; it does not establish that the depicted scene is real.
Then investigate the publisher and context. What is the earliest publication you can find? Can the contributor supply a camera original, neighboring shots, or an edit history? Do those records actually correspond to this image? Visual anomalies should guide further questions rather than replace source verification.
When evidence conflicts, a useful review note might say:
> The detector returned a high AI-related score, but only a reposted screenshot is available. No original file or verifiable production record has been supplied. Full generation versus partial editing remains unresolved; further evidence is needed.
That is more accurate than “the detector says AI, therefore it is fraudulent.” For penalties or public accusations, do not rely on a detection score alone. Preserve a route for additional evidence and reconsideration. The goal is a conclusion proportionate to the evidence, not certainty manufactured from a decimal.