AI Can Guess What You're Typing Just By Listening, Here's How

If someone told you AI could figure out what you're going to say before you finish saying it, you'd probably think of something like autocomplete or Google search results. That's a sensible guess, and indeed, it is based on algorithms determining what you're trying to type based on consistent results. However, at least in theory, it is possible for an AI to guess what you're trying to type even if it can't actually see what you're typing.

How could a machine possibly know such a thing? After all, computers can only parse hard data — they can't read your mind. But that's the thing; it doesn't have to read your mind. All it has to do is interpret your keystrokes. Not the letters and numbers produced by those keystrokes but the sounds of the keystrokes themselves. In other words, it is possible for an AI to guess what you're typing without looking at the text input by simply parsing the sounds of typing coming from your keyboard.

In a recent study conducted by UK researchers Joshua Harrison, Ehsan Toreini, and Marhyam Mehrnezhad, the trio posited that a computer or laptop keyboard's audio, captured through a microphone during something like a Zoom call, could be used to decipher an individual's written communications, even if the theoretical spy couldn't actually see what the individual was typing.

The researchers explored this concept by training a machine-learning algorithm on the distinctive typing sounds of a 2021 MacBook Pro laptop. It might sound nonsensical to guess someone's typings by sound alone, but in actuality, you create many distinctive sounds when you type that an algorithm could latch onto. For instance, the distance of sounds from a microphone could indicate where on a keyboard you're typing, which could be correlated with the slight variations in sound waveforms produced by individual keys.

In the study, after the algorithm was trained, it was able to guess what someone was typing on a MacBook keyboard with 93% accuracy. In instances where the algorithm was wrong, it was usually only off by one or two keys.

While there is definitely potential for this kind of listening algorithm to be put toward malicious purposes, it would take extensive training and testing to replicate the researchers' findings. That's also not considering differences in sound and layout in the many commercially-available keyboards and laptops. As such, the chances of someone monitoring your audio to steal information by keyboard listening are unlikely at best.

Even so, if you're concerned, the researchers made several suggestions for mitigating a theoretical AI's ability to listen to your keyboard. The simplest thing to do would be to change up your typing style, the speed at which you type, and how you hold your hands. You could even try mixing false keystrokes into your text, though this would obviously make it more difficult to actually type.

If you're concerned about passwords specifically, you can try using randomly-generated passwords instead of typing them out yourself or switching to passkeys or other biometric logins.