Counterfeit

Matthew Grimes

In early 2024, a finance worker in Hong Kong joined a video call with his company's chief financial officer and several colleagues. He was asked to move money, and he did: about $25.6 million across a series of transfers. Every person on that call was a deepfake. He found out only when he checked with head office afterward.

That story used to be an outlier. This year, the FBI started counting AI-related scams as their own category for the first time. In 2025, Americans filed 22,364 such complaints and reported nearly $893 million in losses. The FBI says the real figure is likely higher, because it counts only the complaints where victims described AI being used.

Most of the advice about deepfakes is some version of "learn to spot them." Look for odd blinking. Listen for a flat voice. Check the hands. I think that advice is well meant and almost entirely wrong. We've fought this war before, with money, and we didn't win it by training people's eyes.

Funny money

Before the Civil War, the United States didn't have one paper currency. Hundreds of state-chartered banks printed their own notes, each with its own designs. A note from a bank in one state was often taken at a discount, or not at all, in the next.

Counterfeiters loved it. Nobody could know what every genuine note looked like, so a convincing fake from a distant bank was easy to pass. By one count, about 5,400 different counterfeit bills were in circulation in the 1860s. By the end of the war, about one-third of all currency in circulation was counterfeit.

Business adapted the way we're adapting now. Merchants subscribed to counterfeit detectors, periodicals that described every bank's genuine notes and listed the known fakes. Historians count at least 72 of them. One claimed up to 100,000 subscribers. Every shopkeeper became an amateur forensic examiner, checking bills against a booklet behind the counter.

It didn't work. The fakes kept coming. And every booklet that described a genuine note in detail was, in effect, also a guide for the forger.

The fix

What worked was changing the money, not the merchants.

The National Banking Acts of 1863 and 1864 created a single national currency, and a federal tax on state bank notes drove them out of circulation. By 1867, state notes in circulation had fallen from $143 million to $4 million. In 1865, the government created the Secret Service, whose first job wasn't protecting presidents. It was hunting counterfeiters.

Uniform, hard-to-copy currency plus people whose job was enforcement did what a hundred thousand booklets couldn't. Today, counterfeit notes make up an estimated 3/100ths of 1% of US currency.

Share of US currency that was counterfeit

Source: Federal Reserve Bank of Atlanta

1865

~33%

Today

~0.03%

Nobody checks a $20 bill against a booklet anymore. They just use it. The genuine article became easy to prove, so the fakes stopped mattering.

A coin flip

Now look at where we are with deepfakes.

In 2024, researchers pooled 56 studies of human deepfake detection covering more than 86,000 participants. Average accuracy across images, video, audio and text was 55.5%. Statistically, that's no better than chance. Their conclusion was blunt: "Human guesses on whether AI-generated content is fake or real are at chance levels."

People spotting deepfakes vs. a coin flip

Source: Diel et al., 2024

Audio

62%

Video

57%

Image

53%

Text

52%

Coin flip, 50%

0%

100%

Detection software does better than people, and it's worth having. But it's still a counterfeit detector. Every model built to catch fakes is a test the next generator can learn to pass. That's an arms race the forger tends to win, for the same reason it was in 1860.

Prove the real thing

Proving who someone is has become the central problem in cybersecurity. Stolen logins are among the most common ways attackers get in. Deepfakes just extend that problem from passwords to faces and voices. The answer looks a lot like the answer to funny money: make the genuine article provable, and stop asking people to judge fakes by eye.

Three pieces are already in place.

The first is passkeys, the hard-to-copy currency of identity. A passkey is a cryptographic key stored on your device. There's no password to phish and nothing to imitate, so a cloned voice can't talk it out of you. It's also tied to the real website it was made for, so a convincing fake site can't use it. The proof is math, not appearance. The FIDO Alliance estimates 5 billion are now in use. Three-quarters of consumers it surveyed have turned one on somewhere, but only about half use them regularly.

Passkeys: turned on vs. used regularly

Source: FIDO Alliance, May 2026

75%

Turned on

49%

Used regularly

The second is content that carries its origin. Since August 2, the EU AI Act has required providers of AI systems to mark generated content in a machine-readable way, and realistic deepfakes must be disclosed. It's the digital version of a watermark on a banknote.

The third is the oldest trick in security: check through a second channel. The FBI tells families to "create a secret word or phrase" and to hang up and call back on a number they look up themselves. Inside a company, the same rule means no payment moves on the strength of a face or a voice. It moves when someone approves it through a channel the attacker doesn't control, and the record shows who did. The worker in Hong Kong eventually did exactly that, by checking with head office. He just did it after the money had gone.

What's at stake

This isn't only about fraud losses. If nobody can trust a voice on the phone or a face on a screen, a lot gets more expensive. Remote work, telehealth and online banking all run on the assumption that the person on the other end is who they seem. So does a grandparent's call with a grandchild. Americans over 60 already reported about $7.7 billion in fraud losses last year, up 37%. A world where every call needs a second call is a world paying a very high trust tax.

Inside companies, the cost shows up differently. A finance team that can't trust a video call slows down every payment. A help desk that can't trust a voice can't reset a password. Nobody should ban video calls. A face or a voice just shouldn't be the thing that approves anything.

The good news is that the fix doesn't depend on winning the arms race. It depends on making the real thing provable, which we already know how to do.

What I'm not saying

I'm not saying detection is useless. It's a worthwhile layer, and it will catch a lot. It just can't be the foundation.

I'm not saying passkeys solve everything. FIDO found that 57% of organizations using them still let employees sign in with methods that can be phished. And labels can be stripped from content, as watermarks can be cut from a photo.

I work in cybersecurity, so I'm inclined to see a security answer to most problems. Discount that as you see fit.

What would change my mind: if people or machines got reliably good at spotting fakes, and stayed good as the generators improved. Nothing in the last 160 years suggests that's how this goes.

What kind of world?

There's a world where we keep training eyes. Every call is suspect, every video needs a second look, and people learn to distrust the voices of the people they love. The forgers keep improving, and so does the cost of doubt.

There's another world, where proof is built in. Your sign-in can't be phished, content carries its origin, and a family has a word no clone knows. In that world, a deepfake is a nuisance instead of a disaster, because nobody was relying on their eyes to begin with.

By 1866, the counterfeit detectors were disappearing. Nobody needed them anymore. That's the goal for deepfakes: not better booklets, but money worth trusting.

Thanks for reading,
Matthew