The subject knows. Adjusting the AI prompt for counter-OSINT

Most OSINT collection assumes a passive subject: someone who does not know they are being observed and whose digital footprint reflects unguarded behaviour. The standard prompt architecture works well in that context: supply the signals, ask the model to assess consistency, receive a confidence tier. The method breaks when the subject becomes active in managing their own visibility, because the signals you are collecting are no longer ambient. They are selected.

Adversarial conditions in OSINT commonly present in one of three ways: sudden and coordinated deletion of content across platforms, rapid account privacy setting changes that restrict previously public information, or the appearance of counter-narrative content designed to create plausible alternative explanations for already-collected evidence. Each of these is a behavioural signal in its own right, but standard prompt architectures treat the absence of data as absence of signal rather than as a signal of concealment. This prompt addresses the first two, deletions and privacy changes. Counter-narrative content needs a different method and is outside its scope.

The difficulty is that deletion and privacy changes also occur for entirely legitimate reasons. A subject who removes location data from their profile after a news report on digital privacy is not necessarily responding to an investigation. The adversarial assessment requires a temporal baseline: what was present before the change, what changed, and what the timing of the change is relative to events in the investigation timeline. Without a before-state, the after-state has no investigative meaning.