Insurers are starting to see AI used against them in health care fraud, and the tools are getting more industrial. A simple prompt in ChatGPT can generate documentation for a procedure that never happened, while AI agents can be set loose to call an insurance company thousands of times in a single day without a human touching the phone.
Kurt Spear said that is no longer a theoretical risk. “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,” he said, capturing a shift that has pushed commercial health insurers and government programs such as Medicare and Medicaid into a more defensive posture.
The scale of the market helps explain the urgency. The National Health Care Anti-Fraud Association says up to $480 billion is lost each year to healthcare fraud, and it warned in 2025 that the rapid spread of artificial intelligence could deepen those losses. AI can be used to falsify medical records, create fake patient identities, impersonate doctors and scan coverage policies for loopholes, making old scams cheaper to run and harder to spot.
That threat is showing up in the tools insurers are buying. Pindrop is used by some of the nation’s largest health insurers to separate human voices from AI-generated ones, using a system that reads a caller’s voice, behavior, cadence and other attributes, along with carrier signals and the device being used. Jason Barr said some customers have seen 15,000 bot calls in just a couple months, and that fraudsters have become more convincing over time. About two years ago, many synthetic voices were obviously fake, he said. Now robocallers can change accents or even adopt an agent’s voice mid-call.
The picture is not uniform. A UPMC spokesperson said the organization has not seen a significant number of AI-generated fraud attempts, and Ali Fogarty said Pennsylvania Department of Human Services staff do occasionally get fake calls. That gap matters, because it suggests the threat is real but unevenly distributed: some insurers are already bracing for it while others have not yet seen it at scale.
Highmark is adding another layer, in this case a tool that can detect anomalies in medical imaging down to the pixel level. That move follows a year in which a study in Radiology found radiologists had a 75% accuracy rate in telling real and deepfake X-rays apart, and University at Buffalo researchers developed a tool to detect AI-generated radiology reports. Large language models tend to use polished, elaborate wording, the researchers found, while doctors usually write more concisely.
What comes next is not whether AI will matter in health care fraud, but how fast insurers can keep pace. The fraud itself can now be generated at machine speed, while recovery still depends on criminal investigations and often comes back only in cents on the dollar. If the next round of scams is built by AI, the next round of defenses will have to be built just as quickly.
