There was no doubt in my mind that it was his voice on the phone. It was the exact cadence with which he speaks.
Gary Schildhorn, testifying before the U.S. Senate Special Committee on Aging (govinfo). His son had not called. His son had not been in an accident. His son was not in jail.
The voice told him there had been a crash. A pregnant woman was hurt. There was a test, and the result was bad, and bail would need to be posted before the arraignment. A man identifying himself as a public defender came on the line with a figure and wiring instructions. Schildhorn went to get the money.
He described sitting in his car afterward, motionless, trying to process what had happened.
Schildhorn is an attorney in Philadelphia. He spent a career assessing whether things are what they claim to be. A litigator reads a document and asks who produced it, when, and with what interest. He was not careless. He was not credulous. It worked anyway.
The detail that convinced him was not the story. Stories can be checked. It was the cadence. The rhythm of his son's speech, the way the sentences landed. A machine had reproduced something so particular to one person that his father never thought to question it.
The scam did not target gullibility. It targeted a father who would move fast because his child was in trouble.
What the complaint data does and does not capture
In April 2026 the FBI's Internet Crime Complaint Center published its annual report for 2025. For the first time in the center's twenty-five year history, the report carries a dedicated section on artificial intelligence (IC3 annual reports).
The number in that section is 22,364 complaints citing AI, with losses of roughly $893 million. The report as a whole logged more than a million complaints and more than $20 billion in losses.
Against that total the AI figure reads as a rounding error. It is a floor. The report says so itself.
A complaint enters the AI count only when the person filing it knew AI was involved. Schildhorn found out. Most people do not. Someone who wires money to a voice they believe is family, and never learns otherwise, files a fraud complaint with nothing about AI in it, or files nothing at all. The instrument sees only the victims who solved the puzzle.
| Complaint category | Complaints filed |
|---|---|
| Known AI involvement | 22364 |
| No AI identified by the complainant | 977636 |
Every complaint above sits inside the same annual total. Only the smaller figure is known to involve AI, because IC3 records AI involvement only where the person filing said so. That makes 22,364 a lower bound rather than a measurement.
The scale of the industry behind the calls
In July 2026 the UN Office on Drugs and Crime published its threat assessment for South-East Asia. It is built on field investigation, including compounds that were raided (UNODC announcement).
Delphine Schantz, the agency's regional representative, described the operating model as corporate franchising. Specialized departments for laundering money, trafficking people, smuggling migrants and harvesting data, all drawing on the same network. Inshik Sim, the lead analyst, said the scale is outpacing existing responses.
UNODC put combined annual losses from scam offences across East Asia, South-East Asia, Australia and New Zealand at between $88.3 billion and $114.1 billion for 2025. Two years earlier the same agency, using comparable methods, estimated $18 billion to $37 billion (Inflection Point).
| Assessment year | Low estimate (USD bn) | High estimate (USD bn) |
|---|---|---|
| 2023 | 18 | 37 |
| 2025 | 88.3 | 114.1 |
Both endpoints come from the same agency using comparable methodology, which is why the comparison holds. At the midpoint the increase is 3.1 times in two years. The figures are regional, not global (2026 research brief).
Somewhere inside that hundred billion is a man sitting in a parked car. The voice that put him there was made by people with a payroll, a supply chain, and a quarterly interest in getting better at it.
Where the law stops
When Schildhorn worked out what had happened, he went to law enforcement.
He was told they could not act. He had not sent the money. No loss, no case.
He put the problem to the committee plainly. It is fundamental, he said, that when we are harmed by someone there is a remedy, through the legal system or through law enforcement.
Nobody refused him. The category did not exist. The system answers a completed loss, and what happened to him was a near miss produced by a technology that arrived ahead of the statutes.
He walked out of it whole. That is the reason there was nothing anyone could do.
Why most of the numbers cannot be used
Almost every figure in circulation about synthetic media originates with a company selling protection against it. Detection vendors. Identity verification firms. Fraud prevention services. The estimates are not necessarily wrong. They are produced by parties whose revenue improves as the estimate grows, and that travels with the number no matter who repeats it later.
Two rules governed this piece. No source with a financial stake in the finding. Two independent confirmations for anything load-bearing, where independent means not tracing back to the same origin. Six outlets citing one FBI release is one source.
The rules removed most of what has been written on this subject. What survived is below, with what did not.
| Provenance | Disposition | Reason |
|---|---|---|
| FBI Internet Crime Complaint Center | Kept | Federal collection instrument. Publishes its own methodology and its own limitation. No product sold against the finding. |
| UN Office on Drugs and Crime | Kept | Field investigation including raided compounds. Two assessments two years apart on comparable methodology, which makes the comparison internal rather than assembled. |
| U.S. Senate Special Committee on Aging | Kept | Sworn testimony in the official transcript. The account is the witness's own, given under oath, not a case study supplied by a vendor. |
| Detection and identity-verification vendors | Cut | Revenue rises as the estimate rises. Disqualified regardless of who repeats the number afterward. |
| Trade press citing a single federal release | Cut | Six outlets citing one FBI release is one source, not six. Fails the independence test. |
| Consultancy market-sizing forecasts | Cut | Projected losses sold as measured losses, methodology undisclosed, commissioned by parties in the market being sized. |
What the record actually shows
The IC3 report is an honest instrument. It publishes its own limitation. It counts what people report, and people report what they understood to have happened.
The number cannot say how many people were deceived by a synthetic voice last year. It can say how many worked it out.
Schildhorn is in the record because he worked it out, and then said so under oath.
A woman who wired money to a grandson who never called still believes she helped her grandson. The transaction closed. Nobody is going to tell her.
