In February 2024, a finance worker at an engineering firm in Hong Kong joined a video call with the company's CFO and several colleagues, all requesting urgent transfers. The call looked real: familiar faces, office backdrops, synchronized audio. The worker authorized transfers of roughly $25 million. The CFO and the colleagues on that call were deepfakes, generated frame by frame. The Arup case remains the benchmark for a category that has stopped being an edge case: synthetic video has moved from novelty to the default weapon in high-value fraud.
The volume has followed the capability. Industry trackers counted more than 8 million deepfake videos shared globally in 2025, and Resemble AI's year-end threat report verified 1,567 unique deepfake incidents worldwide with over $1.28 billion in documented losses — a figure its authors stress is a floor, since more than 80% of incidents disclose no financial damage at all. Deepfake-related losses in North America exceeded $200 million in Q1 2025 alone, more than the combined total for 2019 through 2023. Sumsub now estimates deepfakes account for 11% of all fraudulent activity worldwide.
The technology that enables this is now consumer-grade. Open-source face-swap pipelines like DeepFaceLab and Deep-Live-Cam, commercial tools that run real-time face swapping, and subscription deepfake kits mean a convincing video of any person — a CEO, a celebrity, a family member — can be produced by someone with no engineering skills. Vendors sell these as services with support channels and money-back guarantees. Real-time face-swap packages from groups like Haotian AI run $1,000 to $10,000 depending on the support tier, and their promotional channels count subscribers in the tens of thousands.
The failure point is not the generators, it is the detectors — including the human ones. In one study of 2,000 people in the UK and US, just 0.1% identified every deepfake in a mixed set correctly, despite being told in advance to hunt for fakes. People held roughly 60% confidence in their answers whether they were right or wrong. Automated detectors are not much better: they are trained on yesterday's fakes, produce inconsistent false-positive and false-negative rates, and many are black boxes whose methodology collapses in a courtroom. Detection is an arms race that structurally favors the generator.
The damage concentrates in three patterns. The first is executive impersonation: a cloned face and voice directing a finance team to transfer funds, the most expensive form per incident. The second is investment fraud, AI-generated video of celebrities and officials promoting fake trading platforms — now the single largest category of deepfake loss, around $1.13 billion and roughly 52% of the total. The third is personal: family-emergency video calls, romance fraud built on AI-generated 'selfies', and nonconsensual imagery. They share a single mechanism: a fabricated 'live presence' used to bypass a trust decision.
The unavoidable conclusion is that 'does this look real?' is the wrong question. A face on a screen is no longer evidence that the person behind it is who they claim to be, and no amount of training will make a panicked viewer an effective verifier. The question that works is 'can this be proven?' — provenance that binds the media to the device and moment of capture, cross-checked against the context in which the request arrives. That is the difference between looking at pixels and verifying reality.
That is the layer ZSure builds. Our media verifier inspects video, audio, and images with DeepfakeJudge-class models — 96.3% detection accuracy at under 400ms latency — and checks the media against source and context rather than relying on the naked eye. For enterprises, we embed verification into the moment of action: onboarding, communication, and transactions. When a 'CFO' appears on a call asking for a transfer, the answer is not whether the video looks real, but whether the caller and the media can be proven. In the deepfake era, verification infrastructure is the only trust signal that scales.
