Part of the AI, Future and War series. Analysis and hypothetical examples are identified in the text.
A convincing video is not a verified event
During a crisis, a fabricated statement can reach decision-makers before anyone establishes who published it or when it was recorded. Generative AI adds another way to produce misleading material, but ordinary editing, recycled footage and incorrect captions remain relevant. A verification process must examine the claim and its context, not just whether a detector labels the file as artificial.
NIST's Generative AI Profile identifies information-integrity risks among the challenges organisations need to manage. This article applies that concern to a practical crisis-communications workflow; it does not claim that any named wartime recording is false. NIST: Generative AI Profile, AI 600-1
What provenance can tell you
C2PA Content Credentials provide a structure for recording and checking information about the origin and editing history of digital material. They help establish provenance and detect tampering with recorded assertions. They do not certify that the depicted event or the accompanying political claim is true. C2PA: Content Credentials explainer
A valid credential can therefore be useful without being decisive. A genuine recording can still be presented with a misleading explanation. Conversely, a file without credentials is not automatically fabricated: older equipment, export processes and distribution platforms can leave a legitimate recording without a recoverable history.
Build a chain of verification
Start by preserving the original file or link and recording when it was received. Identify the first available publisher, look for an official statement through an independently located channel and compare the claim with other evidence. Avoid using contact information embedded in the suspicious message as the sole verification route.
For a hypothetical company receiving a voice message that requests an urgent crisis payment, the important control is independent callback and payment authorisation. An audio-detection score may inform the review, but it should not replace that control. The organisation needs a reliable process even when the recording sounds completely authentic.
Communicate uncertainty without spreading the claim
Internal alerts should separate what is known, what remains unverified and what action is necessary now. Repeating dramatic allegations in an email subject line can amplify them. A controlled incident record is often more useful than forwarding the original material widely to request opinions.
Assign one owner for verification and another for communications approval. Preserve changes to the assessment so that an earlier provisional conclusion is not mistaken for the final position. If a false report has already circulated, a correction should identify the specific claim and the current evidence without pretending that all uncertainty has disappeared.
Evaluate the workflow, not just detection accuracy
A proposed exercise could measure time to identify the original source, time to contact an authorised spokesperson and the number of unverified claims shared externally. Include both fabricated and authentic material. Otherwise a cautious team may appear successful simply by rejecting everything, including genuine warnings.
The future of information integrity depends on combining provenance, independent corroboration and disciplined communication. No single label should carry the entire burden of trust. For organisations exposed to geopolitical disruption, the strongest preparation is a routine that continues to work when information is emotional, incomplete and urgent.