Accuracy and Limitations

No detector is perfect

AI-image detection is probabilistic. A high score can be wrong, and a low score does not prove that an image came from a camera or human artist. The service is most useful as one signal in a broader verification process.

We do not currently publish a universal accuracy percentage. A single number would hide meaningful differences between generators, image styles, resolutions, and editing conditions. Future benchmark results will include the test set, date, transformations, and false-positive and false-negative rates.

Common sources of error

• Compression and screenshots. Re-encoding can remove or introduce image patterns used during analysis.
• Cropping, resizing, filters, and retouching. Heavy edits can make both AI and camera images harder to classify.
• New or uncommon generators. Detection performance can lag behind newly released systems or unusual workflows.
• Hybrid images. A photograph with an AI-edited region may not behave like a fully generated image.
• Illustration and CGI. Human-made digital art, renders, scans, and very clean studio images can resemble synthetic media.
• Low resolution. Small images contain fewer useful signals.

False positives and false negatives

A false positive is a human-made or camera image labeled as likely AI. A false negative is an AI-generated image labeled as likely human. The uncertain verdict exists to reduce overconfident decisions near the middle of the score range, but it cannot eliminate either type of error.

Appropriate use

Use the result to prioritize review, request the original file, check content credentials, compare metadata, or run a reverse-image search. Do not use a single detector score as the sole basis for disciplinary action, legal claims, payment refusal, account suspension, or public accusations.

If you believe a result is wrong, retry with the highest-quality original image and share the case through support without including confidential material. Read the complete detection methodology.