AI Image Detector Guide: Read the Evidence, Not Just the Score

Correction, October 1, 2026: Unsupported accuracy guarantees and claims that every image check stays entirely on your device have been removed. This guide describes a probabilistic, server-assisted tool and the limits of its evidence.

An AI image detector answers a narrower question than “Is this picture true?” It looks for signals associated with its training and analysis methods. A photograph can depict a staged event; an edited photograph can contain real and generated regions; a generated picture can escape detection. A model output is one piece of evidence, not a certificate of origin.

What AICleanify accepts and sends

The AI image detector accepts JPEG, PNG, and WebP files. It sends the selected image to AICleanify's configured detector service. Successful checks create result and statistics records; when media storage succeeds, the image is retained for history and dashboard previews. Do not upload confidential images. Review the page's processing disclosure and the privacy policy before submitting a file.

A browser file picker displaying a file does not mean the detector supports its format. Export an appropriate copy if necessary, but keep the original and record the conversion. A screenshot or a newly compressed copy is a different input, so a different result should not surprise you.

Interpret the three result states

• Likely AI-generated: the detector found an associated signal. This does not establish the generator, the author, or whether every region is synthetic.
• Likely not AI-generated: the detector did not find a strong signal. This is not proof that the picture is authentic, unedited, or entirely human-made.
• We can't tell reliably: there is insufficient separation for a dependable classification. Keep that uncertainty in your notes rather than converting it into a yes or no.

The displayed AI score is a model score, not certainty. If a hypothetical output says 80%, do not report that as a measured 80% chance of deception or as 80% of the image being generated. This example explains interpretation; it is not a test result or a calibration claim.

Keep model analysis separate from provenance

Use the C2PA Content Credentials viewer when you want to inspect signed file records. Signature validation, the trust status of the signer, and the recorded actions are distinct findings. Editable software tags are a different kind of evidence. The C2PA specification's explainer describes these records without making a universal judgment that the depicted event is true.

A file with no accessible credential cannot be declared human-made on that basis. Likewise, reading a provider's name from metadata is not the same operation as decoding that provider's proprietary watermark. The AI watermark checker states which layers it actually inspected; an unavailable check is not a negative result.

Build a small evidence record

1. Preserve the exact input file. Note whether it is an original export, download, crop, or screenshot.
2. Record the tool, date, result wording, and any reported coverage or failure message.
3. Inspect provenance separately and copy the exact status, including unavailable or untrusted findings.
4. Compare the file with its claimed source and, where appropriate, ask the creator for the original or editing history.
5. Write a conclusion proportionate to the evidence: “The model flagged this copy; origin remains unverified” is different from accusing someone of fabrication.

Why we do not publish a universal accuracy promise

The research paper A Sanity Check for AI-generated Image Detection evaluates generalization on challenging generated images and reports limitations in existing detectors. It is not an evaluation of this deployed AICleanify service. A meaningful accuracy claim would require a described dataset, model version, decision threshold, and error analysis relevant to the intended use.

If the service is unavailable, the image is unsupported, or analysis fails, no successful determination has been made. Do not interpret an error as “real.” For consequential decisions, retain uncertainty and obtain additional evidence rather than relying on a single automated score.