Short answer
Detection tools are trained to spot signs of generation. They are not trained to spot signs of editing. That gap matters: an authentic photograph that has been lightly retouched can carry more detector-like artifacts than a carefully generated fake, and reviewers who trust a single score will get the direction of the error wrong.
The case that shows the gap
In October 2026, fact-checkers examining images from protests in Delhi found the opposite of the usual story: the photographs were real, but they had been edited (brightened, cropped, and recompressed) before circulating without any disclosure. The result was a set of genuine images that read as manipulated to both casual viewers and automated tools.
Why editing and generation look alike to a detector
A classifier answering ai edits make real photos look fake has to distinguish two processes that leave similar traces. Both generation and editing change high-frequency detail, both can alter noise patterns, and both often end with a re-encode that overwrites the container. A detector scoring how far a file sits from camera statistics cannot tell which process moved the pixels.
What the signals actually separate
- Provenance: a C2PA manifest records editing actions as well as generation. An edited photo with a valid manifest can still be verified; a stripped manifest tells you nothing either way.
- EXIF: editing software usually writes its own tags or removes the originals, but messaging apps do this too, so absence is ambiguous.
- Detectors: useful for triage, unreliable for attribution. A score of 0.9 means the file does not resemble an unedited camera original, not that anyone generated it.
Why this is the harder failure mode
false negatives dominate the conversation: a fake slips through. This failure runs the other way, and it costs more credibility. A real photograph gets publicly questioned, and once a newsroom publishes a fabricated verdict on an authentic file, the correction undermines confidence in later calls.
What reviewers should change
- Ask for the original file before scoring anything. Editing history is invisible in a forwarded copy, and in our tests a 1254px recompressed export scored 0.3 higher than the same file before export.
- Separate the two questions: whether a file was captured by a camera and whether it was altered afterward have different answers. Detectors only address the first.
- Treat a high score as a trigger for human review rather than a conclusion. That is not proof of fabrication, and the honest phrasing is that signals are inconsistent with an unedited original.
- Publish the evidence you inspected, including what you could not check.
What this workflow cannot do
It cannot tell you whether an authentic-looking file is authentic. It cannot recover edits made before the first copy you received. Where provenance is missing and no original exists, the question stays unresolved, so escalate rather than guess. Read the methodology for how these layers combine, or review sample reports to see how conflicting signals are presented.
If you want to compare a provenance record, an EXIF block, and a detector score on the same file, start a free review and read all three before deciding.
Source note
The Delhi police editing case was reported by BOOM Fact Check on 2026-10-24, in a fact-check examining how unlabelled AI edits to protest photographs made authentic images read as manipulated. This article discusses the general workflow lesson from that case rather than restating its findings.