Hany Farid Warns AI Makes Reality Indistinguishable

Business Insider published a piece on June 21, 2026 covering Hany Farid, digital forensics expert and co-founder and Chief Science Officer of GetReal Security, who warns that ordinary internet users can no longer reliably distinguish AI-generated images, video, or audio from real content. The article draws on a New York Times profile (June 14, 2026) that described Farid as struggling to verify what is real before the internet makes up its mind (New York Times). The San Francisco Chronicle cites cybersecurity firm DeepStrike's estimate that deepfake content grew by roughly 900% over the past year - from approximately 500,000 online deepfakes in 2023 to around 8 million in 2025 (SF Chronicle; DeepStrike). Farid is departing UC Berkeley for Dartmouth College in the fall.

Per the New York Times profile (as covered by PYMNTS), Farid said "I feel like I'm going blind" when describing the difficulty of keeping pace with modern synthetic media - even as a specialist who has spent decades working in detection. He draws a distinction between what computational forensics can do versus what unaided human perception can achieve: forensic analysts given time and mathematical tools can still detect many manipulations, but the gap matters because detection rarely finishes before content goes viral. He told students: "Within 20 minutes, the whole ballgame's basically over," meaning a fake can become accepted fact before analysis completes (New York Times). Vendor-reported data from DeepStrike puts human detection accuracy for high-quality deepfake video at around 24.5% - well below what people estimate their own ability to be (DeepStrike).

Farid co-founded GetReal Security to translate academic detection methods into scalable tooling; UC Berkeley has highlighted his work spanning image, video, and audio forensics (news.berkeley.edu; ischool.berkeley.edu). The underlying challenge for practitioners is that generative model improvements eliminate visible artifacting faster than single-frame classifiers can track. Parallel reporting by PYMNTS notes that deepfake-enabled financial fraud is accelerating - synthetic borrowers built from deepfake video, cloned voice, and fabricated financial histories are being used to pass automated lending underwriting models (PYMNTS).

Editorial analysis - practitioner implications: Robust detection pipelines are shifting toward provenance metadata standards (such as C2PA), model fingerprinting, multi-modal ensemble approaches, and cross-source corroboration rather than single-frame visual anomaly detection. Detection tool accuracy can drop by roughly 50% when moved from controlled lab conditions to real-world deepfakes, reinforcing the need for procedural verification layers alongside algorithmic ones - vendor-reported from DeepStrike (DeepStrike).

Key indicators for practitioners: uptake of digital provenance standards such as C2PA; open adversarial benchmarks; and regulatory standards on AI-detection admissibility. Farid's GetReal Security and academic programs such as DARPA's Semantic Forensics are two signals of where tooling investment is concentrating. Platform moderation systems publishing synthetic-content prevalence metrics and commercial vendors disclosing false-positive/false-negative tradeoffs will also be worth tracking.