AI is both a boon and a worry. On the one hand, it opens the door to groundbreaking technology, while on the other hand it puts us at new risks. Researchers in Pennsylvania, USA have identified one such risk. They experimented with AI-powered systems that could be wired up on the phone, even when you were a few feet apart. The result has raised new privacy concerns.
This study was presented at the 18th ACM Conference on Wireless and Mobile Network Security and Privacy (Wisec 2025) (via Pennsylvania State Blog Post). In the study, the researchers detailed how to use artificial intelligence and radar technology to intercept private phone conversations without touching the device.
The research team tested a system called “wireless tapping.” This shows that by measuring the small vibrations produced by the smartphone earpiece, phone conversations can be partially transferred from up to 10 feet away. Researchers say that although call tapping technology is still in its early stages, their experiments demonstrate that using a vocabulary of up to 10,000 words can reconstruct transcripts with approximately 60% accuracy.
How AI uses wireless staging
So how does this AI system work? Researchers say that every time you speak on the phone, the caller's voice is played through the earpiece speaker, causing small vibrations to occur on the surface of the phone. These vibrations are usually unaware. However, in this experiment, the researchers were able to use these vibrations to decipher the call. “If you can use remote radar to capture these same vibrations, bring machine learning and help you learn what's being said, you can decide the entire conversation,” says Suryoday Basak, a doctoral researcher at the Faculty of Engineering in Pennsylvania.
Researchers used millimeter-wave radar sensors, the same technology found in self-driving cars, motion detectors and 5G networks. They placed it a few meters away from their smartphone. During the experiment, the radar managed to detect subtle surface movements caused by the smartphone earpiece. Researchers then processed these vibrational signals using a customized version of Whisper, an open source AI speech recognition model developed by OpenAI.
Now, instead of retraining the entire model, researchers reportedly applied a technique called low-rank adaptation, which only allowed them to adjust about 1% of Whisper's parameters. This helped AI systems to interpret “noisy” radar data. This usually looks very different from the whispers of clean audio.
And the results of this experiment were of concern. The system was able to create partial but recognizable transcripts for telephone conversations. Although not perfect, the output was reportedly detailed enough to identify important phrases, causing fear that similar techniques could be exploited for spying.
Researchers emphasize that their systems are strictly built for academic purposes and are not ready-made spy tools. But they warned that the downsizing of radar sensors, coupled with the rapid advances in AI's voice recognition, could make it possible for people to replicate technology in compact, hidden devices.
In fact, the team noted that radar chips, such as pens and smart home gadgets, can already be scaled down to everyday objects, allowing wireless eavesdropping that keeps the very realistic possibilities in the near future.
– end
