Checking Whether an Image Was AI-Generated: What Detectors Can and Cannot See

Image detection does not work by recognising subjects or spotting extra fingers. It analyses low-level statistical structure: noise distribution, frequency artifacts, and the traces left by a generator's upsampling process. These signals are invisible to a viewer and survive mild editing, which is why a detector can flag an image that looks entirely ordinary, and equally why it can miss one that looks obviously synthetic to a human eye.

The single most common reason for an unreliable image result is that the file has been through re-compression. A screenshot of a screenshot, an image saved from a messaging app, or a photo re-encoded by a social platform has had exactly the low-level structure detectors depend on partially overwritten. Always test the highest-quality original available. If all you have is a downloaded thumbnail, the honest answer is that the file cannot be assessed.

Common formats such as JPG, PNG and WebP carry different amounts of usable signal. Lossless formats retain more; aggressive JPEG compression removes more. Resolution matters for the same reason: a small image contains less statistical evidence. This is why the same picture can produce different results at different sizes, which is not detector inconsistency but a genuine difference in how much evidence each file contains.

Detectors and people fail differently, which makes them useful together. People are good at semantic impossibility: text that dissolves into shapes, reflections that disagree, jewellery that merges into skin, shadows falling in incompatible directions. Detectors are good at statistical residue that no viewer notices. Running both and treating disagreement as a reason to investigate further is far more reliable than trusting either alone.

Where it is available, provenance evidence outranks any score. Where did the file first appear, who published it, does the original poster have a history, is there an earlier version elsewhere, does embedded metadata survive. A reverse image search that finds the picture published two years earlier settles the question in a way no percentage can. Detection is most valuable exactly when provenance is unavailable, which is also when it should be reported with its uncertainty attached.

If a judgment will be shared or acted on, record what you tested and what you found: the file version, its resolution and format, the score, the visual anomalies noted, and any provenance evidence. Tools such as NotGPT return a probability for an uploaded image, and keeping a short history of checks makes results comparable over time. A conclusion that records its own basis can be revisited; a bare percentage in a chat message cannot.