Berkeley Expert: McConnell Photo Was Real, AI Fuels Fraud

At the ACFE Global Fraud Conference in Boston this week, UC Berkeley professor Hany Farid examined a widely circulated photo of Senator Mitch McConnell and concluded the image is authentic. He warned that generative AI is eroding trust and enabling new forms of fraud.

Farid opened his keynote by saying “things are getting weird out there” and described a rapid acceleration in AI development, from months-long cycles to changes occurring in days. He demonstrated that consumer tools can produce photorealistic images, videos and voices from simple text prompts and that cloning an identity can require only a single photograph and about 15 seconds of voice audio.

He outlined several uses of synthetic media that are troubling for investigators: coordinated disinformation, financial crime, hiring fraud, identity theft and fabricated evidence in legal cases. Farid highlighted live impersonation on video calls, where face-swapping and voice-cloning tools can make a caller appear to be a legitimate participant on platforms such as Zoom, Teams and FaceTime.

Presenting research from his team, Farid reported that people have difficulty distinguishing real from synthetic media. Participants shown mixes of genuine and AI-generated images, audio and video performed only slightly better than chance even after training, and confidence did not predict accuracy.

He rejected common detection heuristics as unreliable, noting “the day of six fingers is over” to illustrate that obvious visual glitches are no longer dependable signals of manipulation.

Farid described a phenomenon he calls the “liar’s dividend,” where authentic material is dismissed as fake. After debate online about a McConnell photo, he spent hours analyzing the image and found it genuine. He warned that growing scepticism is complicating the handling of digital evidence in courts.

For assessing authenticity he recommended technical controls rather than human judgment alone. Methods include metadata-based content credentials, invisible watermarks embedded by AI providers, digital signatures and forensic checks of lighting, shadows, three-dimensional geometry, facial biometrics and voice patterns. He explained many current AI models do not model physical reality, so those inconsistencies can be used to detect manipulated content.

Farid described tools in development to monitor video conferences in real time and detect virtual cameras, voice modulation and known deepfake markers. He referenced federal warnings about AI-driven fraud and cited instances where synthetic identities were used to obtain remote employment, adding that many organisations do not have a clear owner for this risk.

He urged organisations to assign responsibility and provide training on AI-related fraud and identity risks. For individuals he recommended simple verification procedures, such as agreed code words, after his own voice was cloned and used to contact a lawyer on a confidential matter. He closed by noting that creating convincing synthetic media now requires only a keyboard and an internet connection.