Short answer
non-consensual ai images invert almost every assumption in a standard image review. The file matters less than the person in it. Speed matters more than certainty. And the reviewer's job is not to establish whether the image is synthetic but to limit further distribution while a determination is made. Teams that run these cases through a standard authenticity queue are applying the wrong process at the moment when the cost of delay is highest.
The case that shows the gap
In October 2026, Global Voices reported that Bangladesh was weighing tougher cyber crime provisions after AI-generated imagery targeting women circulated online. The reporting focused on a pattern that recurs across jurisdictions: the images spread faster than any review process could evaluate them, and the legal response arrived after the harm had compounded. The technical question of whether the images were synthetic was never the bottleneck.
Why the standard queue does not fit
An authenticity workflow asks whether a file is camera-native, then routes it for review. That sequence is reasonable for misinformation, where the risk is reputational and the timeline is hours or days. It fails here for three reasons:
- Delay is the harm. Every hour a file stays live increases redistribution and archive copies.
- Authenticity is often irrelevant. A fabricated intimate image of a real person causes the same harm whether or not a detector scores it as synthetic.
- The subject has standing the file does not. The person depicted can usually tell you what is real, and their report should drive the process rather than a confidence score.
What the intake process should capture
- Whether a report has been made by the person depicted, and when. This is the trigger, not the detector output.
- The URL or message identifier as delivered, preserved before any takedown request, since removal destroys your own record.
- Distribution surface: public post, private group, or direct message. The escalation path differs and so does the urgency.
- Whether the file is still propagating, which usually matters more than whether it is synthetic.
Where detection still helps
Detection has one narrow use here: distinguishing an edited authentic image from a synthetic one can matter for how a case is described publicly, and for whether a provenance record supports the subject's account. It should never gate the initial response. A low detector score is not proof the image is authentic, and a high score is not proof of fabrication. Where the evidence conflicts and the subject's account is available, escalate to human review with the subject's statement attached rather than resolving it on a model score.
What this workflow cannot do
It cannot undo distribution. It cannot establish who created the image. And it cannot substitute for legal advice, since the obligations that apply depend on the jurisdiction, the age of the subject, and the platform's own published terms. Read the methodology for how evidence layers are combined, or check sample reports to see how a case with an incomplete record is documented.
If you need to inspect provenance and metadata on a file already in hand while a takedown is in progress, start a free review and record what you find before it is removed.