Counting fingers and checking shadow directions worked in 2023 and does not work now. The check has to start with the file, not the eye. An image carries several layers beyond the pixels on screen, and those layers say how it was made.
A PNG downloaded straight from a generator very often still carries the prompt, seed and sampler in its text chunks. XMP may name the digital source type, and some files declare themselves AI-generated in a C2PA signature. When any of that is present there is no need to analyse pixels at all. This is the strongest evidence there is, and the first layer to look at.
A real capture carries EXIF describing lens, exposure and time, sensor noise unique to the device, and a JPEG quantisation table that differs by manufacturer. When those survive together they argue for a human capture. Their absence argues for nothing: one pass through a messenger removes most of them.
Regular patterns in the frequency spectrum, compression-history mismatches revealed by error level analysis and unnatural noise distributions are all clues, and all of them move under a single re-compression. So signals of this family are never allowed to carry a verdict alone; they only gain weight alongside evidence from another layer.
A screenshot, or a file re-compressed several times with its metadata stripped, has almost nothing left to read. What is needed then is not a plausible number but the fact that no verdict is possible, and why. If the original file can be obtained, that is the fastest fix by far.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.