Deepfake detection is losing ground. New testing shows automated systems dropping from 94% accuracy to 48% against the latest AI-generated video output, a decline that exposes how quickly generation tools are moving beyond older safeguards.
The work, titled DF26: We Cannot Tell Fake From Real Anymore, compares real videos with image-to-video and text-to-video examples created for a new dataset. Its steepest result represents a 49% drop in detection performance across nearly a year of deterioration in AI-video detection efficacy.
The closed-source commercial models used to generate the detector-defeating videos were Grok 1.0, Kling 3.0, Veo 3.1, and Wan 2.6. The finding matters because a detector trained against yesterday’s synthetic footage can lose its value when newer systems produce more convincing material — a familiar problem in AI, where the benchmark ages faster than the press release.
An article by Martin Anderson presents the work against a wider backdrop of deepfake use in fraud and misinformation. The authors identify influencer and informational contexts as prime targets for deepfake activity, where convincing video can damage trust before anyone has time to check whether it is real.
Deepfakes Move From Deception to Infrastructure
Manipulated footage of Richard Nixon has already shown how deepfake technology can simulate his announcement of the death of the Apollo 9 astronauts in 1969. That example points beyond personal scams: fabricated video can place real people in invented events and give false claims the appearance of historical evidence.
The fraud threat now reaches identity systems as well. The Shufti Identity Fraud Report 2026 draws on identity verification checks processed between January and June 2026 across eleven industries, showing how organised crime networks reuse AI-generated documents and identities across borders through shared devices and infrastructure.
Faryam Asif, CTO at Shufti, described the shift directly: “Deepfakes are no longer just an individual fraudster’s tool. Organised crime networks are using the same AI-generated documents and identities across borders, reusing them through shared devices and infrastructure. A document authenticity check has no memory. It tests an artefact against a template, not against the attempts that came before it, so a ring clears onboarding one request at a time.”
The report recorded 2.01% of network fraud spanning more than one country. The typical interval between activity in one country and the next was 9 minutes 33 seconds, while the fastest observed sequence took 38 seconds.
Document Checks Face a Software Problem
Deepfake document fraud accounted for 80.10% of AI-enabled fraud, making it the dominant attack instrument and the primary attack in nine of the eleven industries measured. Identity fraud exposure varied more than fivefold across those industries, with Digital Assets recording the highest exposure at 22.49% of all verification requests and Banking recording a fraud rate of 4.24%.
These attacks divide into presentation attacks and injection attacks. A presentation attack reaches a physical sensor, while an injection attack never reaches one; that makes injection a software integrity problem as much as a document problem.
The report also identifies multi-identity account abuse in iGaming, forex, and lending. Its findings include activity involving 70 identities, 13 devices, and 16 verification events — the kind of pattern a one-off document check is poorly equipped to connect.
Aroosa Virk, Brand and Communications Manager at Shufti, is attached to the report’s wider identity-fraud work, while Shufti describes itself as a Glocal identity verification platform. The combined evidence points to a basic weakness: systems that inspect each document or video in isolation can miss the network using them.
The detection problem is no longer limited to whether a single file looks authentic. It now involves tracking repeated identities, shared devices, cross-border timing, software injection, and synthetic media that can outpace the tools built to identify it.
Based on
- Deepfake Detection Accuracy Falls 49% in a Year of AI Video Progress — unite.ai
- The Threat And Challenges Of Deepfakes In Our Digital Ecosystem — forbes.com
- Deepfakes are wrecking influencers’ credibility, one fake ad at a time | AI (artificial intelligence) | The Guardian — theguardian.com
- Organised Fraud Rings Use Deepfakes and Reused Identities to Target Digital Businesses, Shufti Identity Fraud Report 2026 Finds | Markets Insider — markets.businessinsider.com
