A fraud investigator in Dallas runs a facial comparison this morning. By November, the rules for presenting that evidence in court could change three times. Twenty-six states already regulate election deepfakes. There's still no federal law. California tried, and a judge struck it down. For investigators, that means the same comparison result faces wildly different challenges depending on the state line it crosses.
According to Biometric Update, more than a thousand A.I. bills have been introduced across all fifty states. A new federal evidence rule won't take effect until late 2027 at the earliest.
That's an eighteen-month gap where courts are using rules written before generative A.I. existed. The evidence isn't just what you present. It's how well you can prove you understood what you were presenting.
The legal chaos is one problem. The fraud already exploiting it is another.
Ireland's Deputy Prime Minister, Simon Harris, watched a video of himself endorsing investment products he'd never heard of. He told reporters he had to watch it twice to be sure it wasn't him. If the actual person in the video can't tell, what chance does a bank reviewer have? Meanwhile, in Ahmedabad, police arrested a gang that used A.I. videos with fake blinks and expressions to defeat India's national biometric system, opening loans in victims' names.
According to The Irish Times, the fabrication was convincing enough to require a second viewing from its own subject. Deepfake attacks are now recorded, on average, every five minutes globally.
Most fraud workflows still treat video as real until proven fake. That assumption was reasonable in 2019. It's a liability in 2026.
And the fastest-growing version of this attack isn't visual at all.
A finance worker at the engineering firm Arup joined a video call with his C.F.O. and senior colleagues. The voices were familiar. The faces matched. He authorized a wire of twenty-five million dollars. Every person on that call, except him, was a deepfake. The voices didn't have to be perfect. They just had to sound right enough.
According to the American Bar Association, scammers need only three seconds of audio to clone a voice with eighty-five percent accuracy. Human listeners catch high-quality fakes about a quarter of the time.
The defense isn't better detection software. It's a code word. A callback. A pause before the wire goes out. The cheapest fraud control left is friction.
Three stories, one shift. The technology to fake a person costs almost nothing. The systems built to catch it, legal, biometric, and human, were all designed for a world where faking was hard. That world is gone. Verification is no longer a final check. It's the first one.
Links to every story and today's podcast deep-dives are in the description. See you tomorrow.
