A screenshot is not a copy of the original but a new file redrawn from the screen. It looks the same and keeps almost none of the material a check needs. Knowing this explains why captures so often come back inconclusive.
EXIF, XMP, PNG text chunks and C2PA signatures all go. The prompt a generator left and the capture information a camera left go with them. In other words, the two strongest layers are removed in one step.
They have been resized for display and recompressed. What the forensic analyzers read afterwards is the capture pipeline, not the properties of the original. Basing a verdict on the weak signals left in that state amounts to classifying the capture method.
When a file looks like a screenshot or a re-encode and no strong AI evidence is present, we do not lean towards AI. The rule applies only to verdicts leaning that way; applied in both directions it would turn every twice-saved ordinary photograph into an inconclusive.
That it is a capture can still be established. If a file offered as an original turns out to be a capture, that alone is worth pursuing. And the hash always survives, so it remains useful for proving two files are the same.
This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.