Deepfake

Deepfakes are produced by models trained on recordings of a person or on large sets of media. Common forms are face swaps in video, synthetic voice that imitates a specific speaker, and fully generated images or clips. Tools that once required specialist skill are now available as consumer apps, which lowers the cost of producing convincing fakes.

Detection is an active technical problem. Some detectors look for artifacts a model leaves behind, such as inconsistencies in lighting, blinking, lip movement or audio spectrum. Others rely on provenance: metadata and cryptographic signing schemes that record where and how a file was made. No detector is fully reliable, and generators improve against them, so verification usually combines technical checks with reporting on where the file came from.

A hypothetical example: a short audio clip circulates in which a chief executive appears to tell staff that a division will be sold. The voice is a close match. The company's first question is whether the recording is real, and the second is who is spreading it and to whom. Another common form is synthetic video used in fraud, where a fake executive on a call asks an employee to approve a payment.

PeakMetrics has published an account of how a fabricated image spread in how a fake narrative spreads.

The harm from a deepfake has two parts: the false content itself and the doubt it introduces about real content. Once fakes are known to exist, genuine recordings can be dismissed as fabricated. Comms teams need a pre-agreed way to confirm authenticity, a channel that stakeholders trust for corrections, and a list of the people inside the company who can verify a file.

Security and finance teams have a related concern, because synthetic voice and video are used to impersonate executives in payment and access requests. That side of the risk is closer to brand impersonation than to reputation.

Teams can reduce the damage of a fake by agreeing in advance on how authenticity will be verified, who signs off on a statement, and where corrections will be posted. Keeping originals of official recordings with their source metadata makes comparison easier. Staff who handle payments or press inquiries should know that a familiar voice or a video call is not sufficient proof of identity for a sensitive request, and that a callback to a known number is the safer step.

Not every manipulated image is a deepfake. Cheap edits, mislabeled real footage and out-of-context clips are more common than machine-generated fakes and are just as damaging. Detection scores are also probabilities, not verdicts, and a report that says a file is "likely synthetic" should be read as one input.