Deepfakes Explained: How They’re Made and How They’re Detected

Deepfakes have moved from a novelty tech demo to a genuine, widely recognized societal concern in a remarkably short span of time. Understanding both how they’re actually made and how they’re actually detected explains why this particular AI application generates such disproportionate concern relative to many other, less controversial AI capabilities.

How Deepfakes Are Actually Created

Modern deepfakes typically use generative AI models trained specifically on images or video of a target person, learning to synthesize new, entirely fabricated content that convincingly places that person’s likeness into situations, actions, or statements they never actually experienced or said. Early deepfake techniques required substantial amounts of source material and specialized technical expertise to produce a genuinely convincing result, but the tools have become progressively more accessible, and the amount of source material actually needed has dropped substantially as the underlying techniques have continued to improve.

Why Voice Cloning Deserves Its Own Specific Concern

Voice cloning technology, a specific and increasingly prominent category of deepfake capability, can now convincingly replicate a specific person’s voice from a comparatively short audio sample, enabling entirely fabricated audio of someone apparently saying things they genuinely never said. This capability has been directly implicated in a growing number of real, documented fraud cases, including impersonation scams that convincingly mimic a family member’s or a company executive’s actual voice to manipulate a targeted victim.

The Legitimate, Non-Deceptive Uses Worth Acknowledging

It’s worth acknowledging that the same underlying technology also enables genuinely legitimate applications: dubbing film and video content convincingly into other languages, historical or educational recreations clearly presented and labeled as such, and various creative and entertainment applications where the synthetic nature of the content is made clear and isn’t intended to deceive anyone about what’s real. The technology itself is fundamentally neutral — its potential for harm depends specifically on intent and on whether a viewer or listener is deceived about what they’re actually seeing or hearing.

How Detection Approaches Actually Work

Deepfake detection tools generally look for subtle technical artifacts that generation processes often leave behind — inconsistencies in lighting or reflections, unnatural blinking or micro-expression patterns, or characteristic statistical signatures in the underlying pixel or audio data that trained detection models learn to recognize as more likely synthetic than genuinely authentic. Some detection approaches also examine metadata and file-level technical properties that can reveal editing history or hint at synthetic origin, beyond just the visible or audible content itself.

Why Detection Is Locked in a Genuine, Ongoing Arms Race

Deepfake detection faces a genuinely difficult, structurally ongoing arms-race dynamic: as detection techniques improve at catching current generation methods, generation techniques evolve specifically to avoid whatever new detectable artifacts those improved detection methods learned to catch. This means no detection approach can realistically claim permanent, durable effectiveness, since detection and generation techniques keep pushing against each other in a continuously escalating cycle with no clear, final stopping point in sight.

Provenance and Watermarking as a Complementary Strategy

Given detection’s inherent limitations, a complementary strategy focuses on proactively establishing content provenance and embedding watermarks directly at the point of original creation, rather than relying entirely on after-the-fact detection of already-circulating deepfakes. This approach shifts part of the burden toward verifying what’s authentic at the source, rather than trying to catch every fabrication after it’s already been created and has started actively circulating.

What This Means for Ordinary Media Consumption

For ordinary people encountering media online, the practical, sober takeaway is that increasingly convincing synthetic audio and video genuinely exist and continue to improve, which argues for a somewhat more skeptical default posture toward unverified, particularly surprising or inflammatory audio and video content, and for placing real added weight on established source credibility and independent corroboration rather than trusting any single piece of striking media content in complete isolation.

Further reading: See our guides to AI and Elections: The Policy Debate Over Political Deepfakes and Disclosure and Watermarking AI Content: How Companies Are Labeling AI-Generated Media. For background, see Wikipedia’s entry on Deepfake.

Want to contribute to Daily AI News? We welcome guest posts, research write-ups, and press releases from the AI community. Write for us →