How to Spot AI Videos?

The Detection Gap: Why Your Eyes Are Failing You

If you had to judge whether a video was real, you'd probably trust your own eyes. You shouldn't. Research by Korshunov and Marcel found that humans detect high-quality deepfake video with only about 24.5% accuracy — worse than a coin flip, which would land you at 50% by pure chance.

This isn't a failure of attention. It's a failure of the cues we've been trained to look for. For years, the advice was simple: watch for jittery movement, flickering faces, unnatural warping around the edges. Those tells worked against older GAN-based systems. They don't work anymore. According to 2025–2026 deepfake trend research, diffusion models have now surpassed GANs in both photorealism and temporal consistency — the frame-to-frame coherence that used to betray fakes. Meanwhile, NeRF-based techniques can generate full-body 3D avatars that defeat geometric consistency checks entirely.

The scale of the shift is hard to overstate. Circulating video and voice deepfakes grew from roughly 500,000 in 2023 to an estimated 8 million in 2025, per DeepMedia data reported via Reuters. And it's not just volume — it's quality. Deloitte notes that a convincing 60-second deepfake can now be produced in under 25 minutes, at no cost.

Even the machines are struggling. A systematic review of detection methods found that automated detectors degrade by 10–15% on average when tested outside their training datasets, because they learn dataset-specific compression artifacts rather than generalizable deepfake characteristics. Dr. Hany Farid, a digital forensics expert at UC Berkeley, puts it bluntly: synthetic media is no longer reliably distinguishable from authentic recordings — by human perception or by most automated systems.

So what can you actually rely on? Not a single "tell," but a layered process:

• Multiple visual cues, not one. The MIT Media Lab's Detect Fakes framework checks the face, skin, eyes, glasses, facial hair, moles, blinking, and lip movement — three or more anomalies in one video is a strong indicator.
• Audio-visual sync. Peer-reviewed research shows deepfakes exhibit temporal desynchronization between facial movement and speech, because the audio and video are generated independently.
• Procedural verification. Pre-agreed codes and out-of-band confirmation beat visual inspection — the Hong Kong Arup fraud, where every call participant was a real-time deepfake, cost $25.6 million precisely because no one verified out of band.

The uncomfortable truth is that detection is becoming a process problem, not a perception problem. That same logic applies to the AI tooling businesses use every day: trust has to be verified, not assumed. It's why SkillzBench is building independent benchmarking and security scanning for agent skills — so teams can see what a tool actually does before relying on it.

The Multi-Cue Checklist: What to Actually Look For

If a video looks even slightly "off," your brain is probably picking up on something real — you just need a framework to name it. That's exactly what MIT Media Lab's Detect Fakes project offers: a structured way to turn vague suspicion into a scored judgment.

The MIT team's core insight is that there is "no single tell-tale sign" of AI-manipulated media. High-end deepfakes are almost always facial transformations, so the framework trains your attention on the face and its physical behavior. Their guidance notes that while high-quality fakes aren't easy to discern, "with practice, people can build intuition for identifying what is fake and what is real."

The eight checks to run, in rough order of usefulness:

• Facial boundaries — Cheeks and forehead: does skin look too smooth or too wrinkly, and does its apparent age match the hair and eyes?
• Eye physics — Eyes and eyebrows: do shadows fall where physics says they should? Deepfakes often fail to render natural light behavior.
• Glasses — Is there glare, too much glare, or glare that doesn't shift angle as the head moves?
• Blinking — Does the person blink too little, too much, or at unnatural moments?
• Lip movements — Do the lips sync naturally with speech, or drift out of rhythm?

MIT's list also includes facial hair (does an added or removed mustache or beard look natural?) and facial moles (does a mole look real?). Extended practitioner guidance adds peripheral anatomy — hands, ears, teeth — plus hair rendering and lighting consistency as additional categories worth scanning.

The critical rule is cumulative scoring. Detection guidance from security researchers states plainly that no single tell is definitive, but three or more anomalies within the same video constitute a strong indicator of manipulation. One odd shadow means nothing; three simultaneous failures across skin, eyes, and lip-sync means treat the video as suspect until proven otherwise.

Two caveats keep this checklist honest. First, human detection accuracy against high-quality deepfake video is only about 24.5% — below a coin flip — so visual inspection alone is never sufficient for high-stakes verification. Second, newer diffusion-based models have eliminated many of the jittery artifacts older cues relied on, meaning some tells may be weaker today than when first documented.

Lip-sync deserves special weight: it's the most scientifically grounded human check, since deepfakes show temporal desynchronization between facial muscle movement and vocalization because visual and audio are generated independently. A Stanford/UC Berkeley study found lip-sync mismatch detection correctly identified 94–97% of lip-sync deepfakes.

For teams building verification into their workflows, agent skills can automate parts of this process — and SkillzBench is building a free directory where you'll be able to browse, benchmark, and connect proven skills to your agents via a free MCP tool.

Audio-Visual Sync: The Most Reliable Scientific Cue

Watch a person speak and your brain performs a silent calculation: it checks whether the lips, jaw, and facial muscles move in perfect harmony with the sounds you hear. That calculation is exactly where most AI-generated video falls apart.

The reason is structural, not cosmetic. In genuine footage, facial muscle movements and vocalization are naturally matched; in deepfakes, visual and audio information are generated independently, producing temporal desynchronization — tiny gaps between sound and pronunciation that the generator cannot fully reconcile. This is documented in a peer-reviewed systematic review of deepfake detection methods, and it explains why lip-sync timing is the single most scientifically grounded cue a human viewer can check.

The numbers back it up. A 2020 Stanford/UC Berkeley study built a detection tool around phoneme-viseme mismatch — the alignment between spoken sounds and lip shapes — and correctly identified 94–97% of lip-sync deepfakes, while misclassifying only about 0.5% of authentic videos as fake, according to a practical detection guide. Multimodal frameworks that combine audio and visual signals outperform video-only methods by 6–8%, reinforcing that the audio track is not a side detail but half the evidence.

How do you apply this without lab equipment? Focus on the moments where sync is hardest to fake:

• Watch the mouth during hard consonants — "p," "b," and "m" require full lip closure; fakes often produce sound without the lips sealing.
• Check whether lip movement lags or leads the audio by even a fraction of a second, especially mid-sentence.
• Compare emotional intensity: does the face express what the voice conveys, or does expression feel detached from tone?
• Listen for audio that is suspiciously clean, with no room echo or background texture, paired with a "live" video.

One caveat: this cue works best against lip-sync-based deepfakes, where a real video's mouth is re-animated. Fully synthetic video generated end-to-end is fundamentally harder to detect, and newer diffusion models keep closing the gap. Treat sync analysis as one strong signal in a multi-cue check, not a verdict on its own.

For teams that want to operationalize this — say, agents that automatically run detection steps or verify content provenance on incoming video — SkillzBench is building a free directory of agent skills and a free MCP connector so you can browse, test, and equip exactly the verification skills your workflow needs.

Procedural Defense: Moving Beyond Visual Inspection

In early 2024, a finance employee at Arup joined a video call with the company's CFO and several colleagues. Every participant on that call was a real-time deepfake, and the meeting ended with $25.6 million lost across 15 wire transfers, according to reports on the incident. No amount of careful "eyeball checking" would have saved that money, because there was nothing visually wrong to find.

That case illustrates why businesses need to move beyond visual inspection entirely. Human detection accuracy against high-quality deepfake video sits at roughly 24.5% — worse than a coin flip, per research by Korshunov and Marcel. Meanwhile, deepfake fraud attempts have surged 2,137% over three years, rising from 0.1% to 6.5% of all fraud attempts according to Signicat's 2024 data.

The procedural answer rests on two pillars: out-of-band confirmation and pre-agreed verification codes.

• Out-of-band confirmation: Verify any unusual payment request, credential change, or urgent instruction through a pre-established secondary channel — a known phone number, an in-person check, or a separate messaging system — never through the same channel the request arrived on.
• Pre-agreed codes: Establish shared verification phrases or code words with key executives and vendors in advance. A deepfake can clone a face and voice in seconds — voice cloning needs as little as 3 seconds of audio for an 85% accurate replica, per McAfee — but it cannot know a code word that was never spoken or written anywhere public.
• Challenge-response protocols: For high-stakes video calls, ask a question only the real person could answer, or request a real-time action through the secondary channel before proceeding.
• Process friction for money movement: Require dual authorization and mandatory callbacks for wire transfers above a defined threshold, so no single video call — however convincing — can trigger payment on its own.

These protocols work because they test something a deepfake cannot fake: shared secret knowledge and independent channel verification. They also align with how the threat is actually evolving. Diffusion-based models have removed many of the "jittery" artifacts humans once relied on, and even automated detection systems degrade 10–15% when tested outside their training datasets, per an MDPI systematic review. As Dr. Hany Farid of UC Berkeley puts it, synthetic media is "no longer reliably distinguishable from authentic recordings by human perception or by most automated detection systems."

For teams building internal verification workflows, agent skills can automate parts of the process — routing unusual requests through confirmation steps, logging verification outcomes, or checking content provenance. SkillzBench is building a free directory of vetted agent skills plus a free MCP connector, so you can browse and test skills for these workflows in one place.

Visual inspection still has a role for casual content scrolling. But when money, credentials, or sensitive data are on the line, the question is not "does this video look real?" — it is "does this request verify through a channel the attacker cannot touch?"

Automating Verification with Agent Skills

Even the best detection tools have a blind spot, and the numbers prove it. A peer-reviewed systematic review found that automated detectors lose 10–15% of their accuracy when tested outside their training datasets, because they learn dataset-specific compression artifacts rather than generalizable deepfake characteristics. Running one tool and trusting its verdict is simply not a defensible workflow.

The practical answer is redundancy. Experts recommend running multiple detection tools in parallel and treating every output as just one signal inside a broader process — alongside frame-by-frame review, metadata verification, Error Level Analysis, and reverse video search. Done manually, that is slow, repetitive, and easy to do inconsistently under pressure.

That is exactly the kind of multi-step, tool-coordinating work where AI agents shine. Instead of a human pasting a video into three detectors and cross-checking provenance by hand, an agent can orchestrate the entire pipeline:

• Run the file through multiple detection tools in parallel and reconcile conflicting verdicts
• Pull metadata and check for provenance credentials or watermark signals
• Execute frame-by-frame and audio-visual sync checks, which the research identifies as the most scientifically grounded cue
• Produce a scored summary so a human reviewer sees the evidence, not just a verdict

This matters because the threat is scaling faster than manual review can. Circulating video and voice deepfakes grew from roughly 500,000 in 2023 to about 8 million in 2025, according to DeepMedia data reported via Reuters. When a convincing 60-second deepfake can be produced in under 25 minutes at no cost, verification needs to move at a similar pace.

Agent skills — reusable SKILL.md instruction packages for tools like Claude Code, Codex, Cursor, and Gemini CLI — are how you package this workflow so any agent can execute it reliably. The catch is quality control: published research found roughly 26% of skills carry at least one vulnerability, so grabbing an unvetted verification skill off GitHub can introduce more risk than it removes.

SkillzBench is building a free, independent answer to that problem. Its directory will crawl and dedupe skills, pin every version to a commit SHA, security-scan and license-check each one, and benchmark skills in a sandbox — testing the same model with and without the skill to measure the actual Lift. A free MCP connector will let you attach chosen skills to any agent, so you can assemble and test your own automated verification pipeline before relying on it. It is pre-launch now, but joining the waitlist is the easiest way to be first in line when those benchmarks go live.