The fight against deepfakes isn’t going well for humans. Studies show that people can only spot high-quality fake videos about 24.5% of the time. That’s worse than a coin flip. A 2025 iProov study made things even clearer. Only 0.1% of participants correctly identified all fake and real media they were shown.
What makes this worse is that people think they’re doing much better than they actually are. Around 60% of people believe they can spot deepfakes. But the numbers say otherwise. People claim 73% accuracy with fake audio, yet they’re still frequently fooled. Confidence and actual skill don’t match up.
Technology isn’t helping close the gap. Modern AI-generated videos bypass detection tools over 90% of the time. Detection software that works in a lab loses 45% to 50% of its effectiveness in the real world.
Tools like McAfee’s Deepfake Detector and Trend Micro’s ScamCheck can’t guarantee authenticity. Generative AI models are also being specifically trained to defeat detection algorithms, making it an ongoing arms race that defenders are losing.
Voice cloning has crossed a major line too. It now takes only a few seconds of audio to clone someone’s voice. These clones include natural breathing, pauses, rhythm, and emotion. The old tells are gone. Some major retailers now receive over 1,000 AI-generated scam calls every single day.
One fraud case tied to voice deepfakes reached $11 million in losses.
Fake faces have gotten harder to catch as well. Modern deepfake models no longer flicker or warp around the eyes and jawline. Facial movements, voice, emotion, and lighting all sync up together now.
Older detection tools looked for pixel problems and lip-sync errors. Today’s deepfakes have learned to pass those tests. Some deepfakes can now be created in as little as 27 seconds.
Researchers warn the problem is getting worse, not better. Deepfakes are faster to make, harder to detect, and more convincing than ever. The tools meant to stop them can’t keep up. Competitions like NTIRE specifically focus on building detectors that remain effective against low-quality and degraded deepfakes, highlighting how even intentionally distorted fakes continue to challenge existing tools. In 2025, U.S. financial fraud losses reached $12.5 billion, driven largely by AI-assisted attacks that exploited these very detection gaps.
References
- https://fortune.com/2025/12/27/2026-deepfakes-outlook-forecast/
- https://cyble.com/knowledge-hub/deepfake-as-a-service-exploded-in-2025/
- https://www.codabench.org/competitions/12795/
- https://www.missioncloud.com/blog/how-to-detect-deepfakes-in-2026
- https://keepnetlabs.com/blog/deepfake-statistics-and-trends
- https://www.ekascloud.com/our-blog/deepfake-threats-in-2026-can-we-detect-whatu2019s-fake/3636
- https://vertu.com/guides/deepfake-ai-security-threat-2026-prepare-now/
- https://www.imageclef.org/2026/deepfake-detection-and-generation
