Video Deepfake Detection: Why Courts Still Reject the Evidence

Last week, a UN report landed with the kind of quiet devastation that policy documents rarely manage. Roughly 98% of all deepfakes circulating online are non-consensual pornographic images of women. The content exists. The harm is documented. And in most countries, there is no law that specifically covers it. Less than half of all nations have any legislation addressing online abuse in this form. The tools to create the content cost nothing. The tools to fight it cost billions. And the courts? They're still figuring out what "proof" means in a world where video evidence is basically untrustworthy.

This week's headlines aren't separate stories. They're the same story told from three different angles: the abuse crisis, the identity verification scramble, and the emerging technology market trying to clean up both messes at once. If you work cases involving digital evidence, and increasingly, who doesn't, all three threads are heading straight for your desk.

---

Deepfake Images: The Scale Is Staggering and Growing

Here's a number that should stop you cold. According to research cited by Deloitte, online deepfakes grew from roughly 500,000 pieces of content in 2023 to approximately 8 million in 2025. That's a 16x increase in under two years. The detection market is growing at 42% annually and is projected to hit $15.7 billion by 2026. Those numbers sound impressive until you do the math: detection investment is scaling linearly while deepfake creation scales exponentially. Offense is still winning.

Video Session Identity: Why Selfie Video Checks Are Multiplying

A growing number of platforms now ask users to complete a short video step before they can open an account or reset access. During that video session, the system asks the person to move their head, blink, or read a number out loud, which helps confirm a real human is present rather than a photo or a recording. This kind of selfie video check is becoming one of the most common ways platforms try to confirm that the person signing up is who they claim to be. The appeal is simple: a live video is much harder to fake convincingly than a single still picture, at least for now.

The UN News report focuses specifically on women as targets, and it's worth sitting with the institutional failure this represents. Survivors aren't just being harmed, they're being disbelieved. According to the UN Women explainer accompanying that report, deepfake images can be difficult to disprove precisely because gender stereotypes already undermine women's credibility. The deepfake doesn't just cause harm, it creates a secondary victimization loop where the target has to prove a negative to a skeptical audience. That's not a technology problem. That's a forensic standards problem that technology has to solve.

Meanwhile, the legislative response is fractured by design. South Dakota just signed a deepfake pornography felony bill. Washington state passed its own identity protection law. Germany is debating criminal statutes after a high-profile deepfake scandal. Minnesota's "anti-deepfake" law is being challenged on free speech grounds by the Liberty Justice Center. Every jurisdiction is writing its own rules, and none of them are synchronized. What's admissible in one courthouse may be inadmissible three states over, or completely unaddressed in the country where the content was created.

---

Biometric Identity Verification Against Deepfake Images

Live Video ID Verification: What the Verification Call Actually Checks

Video identity verification usually works as a short live video call or a brief recorded clip rather than a single uploaded photo. On a verification call, an agent or an automated system asks the person to show a government ID next to their face, then compares the two in real time. This live video id verification step matters because a still photo can be lifted from social media, but a live video is much harder to stage convincingly on short notice. That single difference, live versus static, is why so many identity checks have moved toward video in the last two years.

Thread two this week: proving who you are online is rapidly becoming synonymous with showing your face to an algorithm. Discord reportedly runs 269 separate checks that match user faces against databases during its age verification process. Australia's age verification laws triggered a reported 250% overnight VPN surge, people are willing to route their traffic through another country rather than submit a biometric. And yet the laws keep coming, because the alternative (unverified minors accessing harmful content) is politically untenable.

Korea extended its facial recognition phone activation pilot through June. The UK's Centre for Finance, Innovation and Technology is building frameworks for trusted business digital identity. Pakistan formally accepted digital ID in legal proceedings. Ghana is adding liveness detection to SIM verification. India's Aadhaar system now has 134 crore live holders, that's 1.34 billion people tied to a biometric record. This isn't a trend. It's infrastructure being poured while the concrete is still wet.

"AI-detection tools remain an emerging field where tools and methodologies are often proprietary and can introduce uncertainties in their results, making manual validation essential and reducing courtroom defensibility." Kennedys Law, "86% Fake, 100% Admissible: Rethinking Evidence in the AI Era"

For investigators, this biometric default creates two simultaneous pressures. On one hand, more cases will hinge on identity, was this person actually present, did they actually send this message, is this actually their voice? On the other, the systems being built to answer those questions are themselves under scrutiny. Essex police just paused their facial recognition camera program after a study flagged racial bias. Spain fined identity tool Yoti for privacy violations in its biometric app. The tools exist. Their defensibility in court is still being negotiated. Continue reading: Deepfakes Hit 8 Million Courts Still Cant Trust Th.

---

Deepfake Legislation Gaps: The Courtroom Battleground

Video Identification Is Becoming a Document Verification Standard

Video identification is becoming a document verification standard because it links two things a court can actually weigh: a physical ID document and a live, moving human face. Document verification alone, just scanning a driver's license or passport, can confirm a document is real, but it cannot confirm the person holding it is the same person pictured. Adding a short video step closes that gap. This is the process that verifies a user's identity remotely, and it is quickly becoming the baseline expectation rather than an extra security layer.

Thread three is where the money is moving, and it tells you everything about where the pressure is landing. VeryAI just raised $10 million to launch what it's calling a "Proof of Reality" identity verification platform. Neuramancer landed €1.7 million in pre-seed funding to scale deepfake detection tools. A major energy-backed venture fund invested in Resemble AI specifically to expand deepfake detection in the Middle East. Zoom integrated Pindrop's voice security to flag synthetic audio on calls. A biometric IDV startup just opened US operations and launched an anti-fraud suite simultaneously.

Banks, courts, and platforms are all arriving at the same realization: they need technology that can say, with documented confidence, "this face and this identity belong to the same person" and "this audio or video has not been synthetically generated or tampered with." That's not a product category. That's a new standard of proof, one that will be written by whoever can make their methodology survive cross-examination.

Here's the thing about the Kennedys Law analysis that should keep investigators up at night: the problem isn't whether you can detect a deepfake. The Stimson Center's analysis of AI-facilitated violence lays it out plainly, the real obstacle is what happens when your detection result reaches a courtroom. Judges can't admit what they can't explain to a jury. Defense counsel will challenge the proprietary nature of any tool that produces its findings from a black box. And without a standardized certification framework for facial comparison analysts, which still doesn't formally exist, every expert opinion is vulnerable to being characterized as one person's guess dressed up in technical language.

As Fortune's deepfake outlook makes clear, real-time voice synthesis is no longer a threat on the horizon. It's already here. A CFO at a major firm was defrauded via deepfake video call. A York city councillor had a fabricated video circulating about them. Seniors are losing thousands to AI-generated phone scams that replicate the voices of family members. These aren't edge cases in a distant future, they're this week's headlines. Each one of those incidents becomes a case. Each case demands an investigator who can produce something more than "my gut said it was fake."

Platforms like CaraComp are built specifically for the documented, explainable facial comparison work that courts actually require, not just detection, but defensible methodology with audit trails. That distinction matters more than it ever has.

Video identity verification is not a single product; it is a category that includes selfie checks, live video calls, document verification, and identification steps layered together. Understanding the difference between these methods matters for investigators because each one produces a different kind of record, and each record carries different weight in court. A simple selfie video comparison is not the same as a verification call with an ID document, a timestamp, and a session log.

Identity verification vendors describe their video identity verification process in different ways, but most share a common structure. First, the user submits a photo ID. Second, the system or agent runs a live video or short video capture of the user's face. Third, software or a human reviewer compares the live capture against the ID photo to confirm identity. Some services add identification checks against government or credit bureau databases as a fourth layer, especially for financial accounts.

For investigators building a case file, the paper trail behind a video identity verification matters as much as the result. A verification call that includes a timestamped recording, a document scan, and a match score is a fundamentally different piece of evidence than a vendor's one-line confirmation that "identity verified." Courts weighing digital evidence in deepfake-adjacent cases are increasingly going to ask what stood behind that green checkmark, not just whether it appeared.

Video session identity checks also vary widely in rigor. Some services only ask for a still selfie compared against an ID photo, which is closer to basic identification than true video identity verification. Others require the user to move through a short video sequence, turning their head, speaking a phrase, or following an on-screen prompt, specifically to defeat pre-recorded or synthetic video. That second category is what serious identity verification increasingly means in 2026, and it is the standard courts are likely to expect when video evidence itself is in question.

The practical challenge is that verifying identity and detecting a deepfake are two different jobs that keep getting bundled together. Video identity verification confirms that the person present during a specific video session is who they claim to be. Deepfake detection asks a different question: whether a separate piece of video or audio was manipulated after the fact. An investigator who confuses these two jobs risks presenting a strong identity verification result as if it also settles a deepfake question it was never designed to answer.

That distinction is exactly why documentation matters so much right now. A short video capture used only for identity verification, on its own, says nothing about whether some other clip circulating online is real or synthetic. Treating identity verification records and deepfake detection records as interchangeable evidence is a mistake that a competent defense attorney will find quickly, and it is one more reason investigators need to be precise about which process produced which record before it reaches a courtroom.

Detecting Deepfakes: What a Deepfake Detector Actually Looks For

Detecting deepfakes is a different discipline from verifying identity, and it deserves its own explanation. A deepfake detector examines a video frame by video frame, looking for artifacts that a human eye tends to miss, unnatural blinking patterns, inconsistent lighting between a face and its background, or blending errors around the hairline and jaw. Deepfake detection tools also compare audio and video tracks for sync errors, since a fabricated voice layered onto real footage often drifts out of step with lip movement over time. None of these signals prove fabrication by themselves, but taken together they help a video detector build a confidence score rather than a guess.

Detect deepfakes efforts generally fall into two camps: tools that flag a single video frame as suspicious, and tools that track a signal across an entire clip. A model built to detect ai-generated faces might, for example, watch how skin texture behaves under changing light across many frames rather than judging one still image. Deepfake content that passes a single-frame check can still fail a frame-by-frame video review, which is one reason investigators should ask which method a given video deepfake detection tool actually uses before relying on its output. Detect fakes work is only as strong as the weakest frame it lets through.

Deepfake detection tools rarely work in isolation from provenance data. Provenance simply means the trail of information showing where a piece of media came from and whether it has been altered since, a timestamp, a device signature, an unbroken chain of custody. When a video detector's output is paired with provenance data, an investigator can show a court not just that something looks synthetic, but where the file originated and whether it was edited afterward. That pairing is often what turns a technical finding into evidence a judge will actually admit.

Why Deepfake Detection Alone Rarely Settles a Case

A deepfake detection result is a starting point, not a conclusion, in most investigations. Software that flags deepfake content as likely manipulated still needs a trained analyst to explain, in plain language, which frame-level signals drove that score and why they matter. Courts have shown they will admit expert testimony about a deepfake detector's findings more readily when the expert can walk through the underlying video frame evidence step by step, rather than simply reading out a percentage. That is the same explainability gap this article keeps returning to: detect deepfake tools can flag a problem, but only documented human review turns the flag into something a jury can trust.

Media literacy plays a role here too, separate from the software itself. Audio and video that reaches the public rarely comes with a label explaining how it was checked, so juries and judges are often encountering deepfake detection concepts for the first time inside the courtroom. An investigator who can explain, in plain terms, how video detection works, what a model looks for, what it cannot rule out, and why provenance data strengthens a finding, gives a court a way to trust the process even without a technical background. That plain-language bridge between the model's output and the judge's understanding is often the difference between evidence that gets admitted and evidence that gets excluded.

Images present their own version of this problem. A single manipulated image can be checked with many of the same techniques used on video frame analysis, since a still image is really just one frame stripped of its surrounding context. Investigators who only learn to evaluate moving video can be caught off guard when a case turns on a single deepfake image instead of a clip, so building fluency in both formats matters. The underlying models behind most detection tools are trained on both formats for exactly this reason.

Videos submitted as evidence also raise a practical chain-of-custody question that deepfake detection software cannot answer on its own. Knowing that a clip was likely generated or altered by a model does not tell an investigator who altered it, when, or why, which means detection results have to be paired with the same documentation practices used for any other piece of digital evidence. Treating a positive detection result as the end of the investigation, rather than the start of a documentation process, is the single most common mistake investigators make with this technology right now.