The Checks That Stopped Working

Written By Matthew Delaney

On 3 December 2024 the FBI's Internet Crime Complaint Center published alert I-120324-PSA, "Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud." It is a short document, and most of it is an inventory of what the technology is being used for.

Criminals use it for translation, the advisory says, "to limit grammatical or spelling errors for foreign criminal actors targeting US victims." It records generated profile photographs and identification documents, "short audio clips containing a loved one's voice," and "videos for real time video chats with alleged company executives, law enforcement, or other authority figures."

Read that list against the advice families have been given for twenty years and it becomes an inventory of retired checks.

Bad English marked a foreign caller. A photograph could be searched for earlier use. A voice belonged to one person, and a face on a live video call was in the room it appeared to be in.

The grammar tell was always the weakest of the four, and it has become the least informative. It never tested whether a message was honest; it tested whether whoever wrote it had learned English as a child. Software removed that difference and put nothing in its place.

The photograph is the clearest case. The old check was not really whether an image looked real - it was whether that image had appeared somewhere else first, under a different name.

An image generated on demand has no earlier life to find. The search that used to surface a stolen stock photo returns nothing instead, and nothing has always read as clean.

The same advisory then offers seven protective tips, and two of them are those retired checks.

It asks readers to "look for subtle imperfections in images and videos, such as distorted hands or feet, unrealistic teeth or eyes," and to "listen closely to the tone and word choice to distinguish between a legitimate phone call from a loved one and an AI-generated vocal cloning."

Two peer-reviewed studies have since measured whether people can do the second of those. Mai, Bray, Davies and Griffin, publishing in PLOS ONE on 2 August 2023, tested 529 listeners in English and Mandarin and found they identified a deepfake clip correctly 73 per cent of the time when judging it on its own.

Training those listeners on examples first improved them by 3.84 percentage points. In Scientific Reports on 31 March 2025, Barrington, Cooper and Farid played 304 participants AI clones built from more than 220 real speakers, and listeners judged a clone and its original to be the same person roughly 80 per cent of the time.

A separate group of 300 in that study correctly labelled a voice as AI-generated 60.8 per cent of the time, with 21 per cent of participants at or below chance. Neither study sampled older adults, and neither tested anyone who had just been told a grandchild was in jail. Both measured calm listeners.

The tips that survive are the ones that ask nobody to perceive anything.

Agree a secret word with your family, the advisory says. Hang up, look up the organisation's number independently, and call that. Send money to nobody met only online or by phone.

What separates those from the two that failed is where the decision sits. A secret word is checked rather than judged, and a number you looked up yourself is verified by the channel it travels on, not by whoever answers it.

Harder to see is the scale. The FBI's press release of 6 April 2026 records that the 2025 Internet Crime Report carries a section on artificial intelligence for the first time, "which accounts for 22,364 complaints, costing Americans nearly $893 million" - all ages, every kind of AI-enabled fraud, calendar year 2025.

That figure is not broken out by age, and it is not broken out for voice cloning. No federal loss figure exists for AI-enabled fraud against adults 60 and over. What can be seen changing is the mechanism, not the total.

For anyone advising an older adult, the consequence is narrow. Take the perception tests out of the script and put a callback rule where they were.

If this is your family

Pick the secret word this week, at a moment when nothing has happened. It should be something a stranger could not guess and could not find on anyone's social media - not a pet, not a street you once lived on.

Then agree the half that matters more. Any call about an emergency ends with a hang-up and a call back to a number you already had, whoever it sounded like. The word is for the times the callback cannot wait.

Sources

Matthew Delaney
Previous

What H.R. 2478 Actually Does
Next

The Entry Is the Exposure