Deepfake Detection: what today's tools can and cannot catch

Overview

Deepfake detection spans three distinct modalities - audio, video, and image - each with its own forensic techniques and failure modes. Unlike text detection, the stakes here are often immediate: a cloned voice authorizing a wire transfer, or a fabricated video surfacing hours before an election.

This guide surveys what the current generation of tools can actually do, drawing on hands-on evaluation rather than vendor demos, and connects to the provenance approaches covered in the Content Authenticity guide.

Audio, video, and image

Audio is the frontier. Voice cloning needs only seconds of reference audio, and live detection is immature - making phone-based fraud the most pressing threat. Video forensics look for inter-frame inconsistencies, unnatural blinking, and lighting artifacts, but quality is improving fast. Image detection examines compression and frequency-domain artifacts that generators leave behind.

Nobody asks "is this fake?" in the abstract. The real question is always "can we verify this fast enough to matter?"

The real-time gap

The hardest problem is detection at the speed of a live interaction. Post-hoc analysis of a recorded file is tractable; flagging a synthetic voice or face during a video call isn't yet reliable at scale. Organizations should assume live media can be spoofed and build verification steps - call-backs, code words, secondary channels - that don't depend on catching the fake in the moment.

Building a response plan

The organizations that handle synthetic-media incidents well prepare before one happens. A workable plan defines who verifies suspect media, what tools and provenance checks are used, how findings are escalated, and how the organization communicates. I help teams build exactly this - see the newsroom case study.