AI is both a boon and ban. On the one hand, it is opening doors for ground technologies, while on the other hand, it is exposing us to new risks. Researchers at Pen State in the US have identified such a risk. They have experimented with an AI-mangoing system that can evcely on phone calls, even when you are several feet away. And the results have now raised new concerns on confidentiality.

The study was presented at the 18th ACM conference on security and privacy at Wireless and Mobile Network (WISEC 2025) (through Pen State Blog Post). In the study, researchers expanded how they used artificial intelligence and radar technology to disrupt private phone conversations without touching the device.

The research team tested a system that they call “wireless-tapping”, showing that the phone conversation can be partially moved from 10 feet away by measuring the small vibrations generated by the earpiece of a smartphone. According to the researchers, while the call-tapping technique is still in its early stages, their use displayed that tapes can be rebuilt with about 60 percent accuracy when using the 10,000 words vocabulary.

How wireless-tapping works using AI

So how does this AI system work? According to the researchers, whenever we talk on the phone, the sound of a collar is played through the earpiece speaker, causing small vibrations on the surface of the phone. Usually, this vibration is not noticed by anyone. But in this experiment, researchers managed to understand the call using these vibrations. A doctorate researcher at Pen States College of Engineering, Sunrise Basak said, “If we occupy these same vibrations using a remote radar and bring us to help us in the machine learning to know what is being said, we can determine the entire conversation.”

Researchers used a millimeter-wave radar sensor, the same technology that is found in self-driving cars, motion detectors and 5G networks. He deployed it a few meters from the smartphone. During the experiment, the radar managed to detect the microscopic surface movements caused by the earpiece on the smartphone. Researchers then processed these vibration signals using a customized version of the whipper, an open-source AI speech recognition model developed by Openai.

Now instead of withdrawing the entire model, the researchers allegedly implemented a technique called low-rank adaptation, which allowed them to adjust only 1 percent of whipper parameters. This helped the AI system explain the “noise” radar data, which looks very different from the clean audio whipper, usually trained.

And the results of this experiment were related. The system was capable of producing partial but recognizable tape of phone conversations. While far away, the output was allegedly detailed enough to identify major phrases, increasing the apprehension that similar techniques could be exploited for espionage.

Researchers throw light on the fact that their system was strictly created for academic purposes and is not a ready detective device. However, he warned that the AI speech can make it possible to repeat technology in hidden equipment, compact, compact for people, in collaboration with rapid progress in speech recognition.

In fact, the team mentioned that radar chips may already shrink into everyday objects, such as pen or smart home gadgets, which are very real possibilities in the near future.