How to Detect a Deepfake: What Works and What Fails in 2026

The visual checks that stopped working

For a few years, the human eye had a fighting chance. Early generators struggled with blinking cadence, teeth, earrings, glasses reflections, hair wisps, and the boundary where a swapped face met the original neck. Guides taught those tells, and they worked, briefly.

Every one of those artifacts was a training target. Modern diffusion and video models render blink physiology, dental geometry, and reflective surfaces convincingly, and face-swap pipelines blend boundaries below perceptual thresholds. Under test conditions, trained reviewers now perform near chance on high-quality synthetic video, and confidence has no correlation with accuracy. The full research record is covered in [the human guessing fallacy](/media/smarthub/human-guessing-fallacy-deepfake-audits), but the operational conclusion fits in one line: if your deepfake defense is a person looking carefully, you do not have a deepfake defense.

The checks that still help a human

Context checks outlive pixel checks. Verify the channel: did this request arrive through an expected route, or a new number, account, or meeting link? Verify independently: call back on a known-good number before acting on a video instruction. Verify the ask: urgency plus secrecy plus payment change is the social fingerprint of deepfake fraud regardless of media quality. These habits stop real attacks, including the ones with flawless video, because they refuse to let the video be the authentication.

The signals that still expose deepfakes

Machines detect deepfakes by examining what generators do not control: the physics of capture and the physiology of living tissue.

Frequency and texture forensics

Generated imagery carries statistical fingerprints, frequency-domain patterns, texture priors, and consistency signatures inherited from the model that produced it. Detection models trained on these artifacts flag content that looks perfect to the eye. This is the workhorse of media forensics, with one caveat: it is an arms race, and each generator generation shrinks some fingerprints while leaving others.

Structural and subdermal analysis

A real face is a three-dimensional object with blood flow beneath translucent skin. Light striking it scatters subsurface before returning; a screen, a mask, or an injected stream returns nothing of the kind. deepidv's deepeye engine projects and reads structural light patterns and analyzes subdermal reflectance during a passive capture, verifying properties that only exist in living tissue physically present at the sensor. This class of check does not degrade as renderers improve, because it tests presence rather than appearance. The pipeline is documented on the [deepidv technology page](/technology).

Capture-path and provenance forensics

Most fraudulent media never touches a camera: it enters through virtual camera drivers or injection at the stream level. Detection at this layer verifies the session, not the picture: driver provenance, sensor noise lineage, hardware-signed capture, and the telemetry coherence covered in [the telemetry forensic framework](/media/smarthub/telemetry-forensic-framework). A perfect deepfake delivered through an impossible capture path is rejected without its pixels ever being judged.

Voice dimensions

Synthetic voice detection analyzes spectral characteristics and vocal biomarkers, jitter, shimmer, harmonic structure, that cloned voices reproduce imperfectly. In call-based fraud, voice and video forensics corroborate each other; a call that passes one and fails the other is a call to escalate.

How detection software reaches a verdict

Professional deepfake detection software does not run one clever test. It fuses layers: media forensics scoring the content, structural liveness testing presence, capture forensics validating the path, and provenance correlation checking whether this same face, document, or device fingerprint has appeared across other sessions. Fusion is what defeats the single-point counterattack: a fraud ring that beats one signal disturbs two others, and the correlated verdict holds.

The fusion architecture also explains a pattern in this year's fraud data: rings reuse assets, with 65.68 percent of linked fraud attempts running recycled forged documents. Detection with a persistent forensic index recognizes a reused artifact on sight, which converts detection from a per-session gamble into an accumulating advantage. That correlation layer, plus red-team pressure from [Arbiter](/arbiter) attacking the stack with current fraud tooling, is what separates production-grade defense from a demo.

What to look for in deepfake detection tools

Buyers comparing deepfake detection tools can cut through the category's marketing with four questions. What layers run: content forensics alone, or structural liveness and capture-path verification with it? Single-layer tools are the ones the next generator release quietly breaks. Where does it run: batch analysis of uploaded files suits investigations, but fraud defense needs in-session verdicts at login, onboarding, and payment speed. What does it learn: a tool with a persistent forensic index compounds, recognizing reused assets and kit lineage, while a stateless scorer meets every attack fresh. And who attacks it: ask for the vendor's most recent red-team results with current generation tooling, because a detector that has not been adversarially tested this quarter is a hypothesis.

Deployment context matters as much as the engine. Consumer-facing protection, screening calls and media for individuals, prioritizes coverage and simplicity. Enterprise verification prioritizes integration depth: the detector must sit inside the identity flow, share evidence with fraud and compliance systems, and produce records an auditor can read. The same underlying science serves both, but the packaging differs enough that a strong consumer tool can be a weak KYC control and vice versa.

A practical playbook by role

For individuals: authenticate the channel and the request, never the face. Independent callback on a known number defeats the CFO deepfake that fooled a diligent employee out of $25 million.

For fraud and compliance teams: deploy layered automated detection at every camera session, retire visual review as a control (keep humans for adjudicating machine evidence), and demand red-team results from any vendor claiming detection.

For executives: assume every video call, voice call, and document image can be synthetic at zero marginal cost to the attacker, and fund the verification layer accordingly. Budget for the arms race, not the snapshot: detection is a capability with a maintenance curve, and the line item that matters is continuous adversarial testing, because a detector nobody attacks internally is being attacked externally on your production traffic instead. Regulators already assume it: Hong Kong's supervisory circular requires AI deepfake detection across remote banking video KYC, and equivalent expectations are spreading.