Pulling off health insurance fraud used to take time and skill.
Until a few years ago, criminals couldn’t fabricate health records without knowing medical terminology and billing codes.
If the scheme involved fooling a call center worker, the fraudster would have to pick up the phone and pose as a patient or doctor.
With artificial intelligence, these barriers have melted away. A simple prompt in a large language model like ChatGPT can generate documentation of a procedure that never happened. AI agents can be told to call an insurance company thousands of times in a single day without human involvement.
“We believed (AI) was something that was going to be leveraged against us as an insurance industry for fraud, and now we’re starting to see that,” said
Kurt Spear, vice president of financial investigation and provider review at Highmark.
The rapid proliferation of artificial intelligence could put commercial health insurers as well as Medicare, Medicaid and other government insurance programs at risk for further losses, the group warned in a 2025 report.
Up to $480 billion is lost each year to healthcare fraud, according to the National Health Care Anti-Fraud Association. Criminal investigations are the only way to recover that money, and recovery is often only cents on the dollar.
A UPMC spokesperson who declined to be named said the organization has not seen a significant number of AI-generated fraud attempts. Neither has the Pennsylvania Department of Human Services, which administers the state’s Medicaid program, though staff do occasionally get fake calls, according to agency spokeswoman Ali Fogarty.
AI can be used to falsify medical records, create fake patient identities, impersonate doctors and scour coverage policies for loopholes, the National Health Care Anti-Fraud Association report said. The scale at which this technology operates is especially concerning to insurance and cybersecurity companies.
“We have customers that have seen 15,000 bot calls in just a couple months,” said Jason Barr, vice president of healthcare for Pindrop. The Atlanta-based firm’s technology is used by some of the nation’s largest health insurers to parse human voices from AI-generated ones.
Pindrop’s AI detection system runs in the background during calls, analyzing the speaker’s voice, behavior, cadence and other attributes to see if they’re human. It also relies on data from carrier signals and the devices being used to make the calls.
About two years ago, many synthetic voices were obviously fake, according to Barr. They’re not perfect these days — Pindrop’s customers have reported robocallers changing accents or adopting the agent’s voice mid-call — but they’re much more realistic.
Highmark has various tools it uses to scan for AI-powered fraud and is in the process of adding another that can detect anomalies in medical imaging down to the pixel level. The human eye may no longer be enough in many cases. A study published this year in the journal Radiology found radiologists had a 75% accuracy rate in telling real and deepfake X-rays apart.
AI has tells in its writing, too, that can be leveraged to prevent fraud. Researchers at the University at Buffalo have developed a tool to sniff out AI-generated radiology reports, the written results of an imaging test. Large language models tend to use polished, elaborate wording, while doctors prefer a more concise approach, the research team found.
For all the high-tech systems being developed, ordinary patients are among the best defenses against fraud. If they receive an email or letter about care that never happened, that’s a strong sign something might be wrong, according to Spear.
“Some of the best referrals come from members,” Spear said.