A machine just watches the muscles. It doesn't get fooled.
Desde tiempos inmemoriales, la humanidad ha buscado una manera de distinguir la verdad del engaño, ese umbral invisible entre lo que somos y lo que fingimos ser. Investigadores de la Universidad de Tel Aviv han dado un paso significativo en esa búsqueda ancestral: un sistema de inteligencia artificial que lee los movimientos involuntarios del rostro —las microexpresiones que escapan al control consciente— y detecta mentiras con un 73% de precisión, superando al polígrafo tradicional. La máquina no juzga ni condena; simplemente observa lo que el cuerpo no puede ocultar.
- El polígrafo, durante décadas el estándar del interrogatorio, tiene una grieta conocida: quienes se entrenan pueden controlar su respiración y pulso lo suficiente para engañarlo.
- Los músculos del rostro son más difíciles de domesticar — las microexpresiones en mejillas y cejas ocurren más rápido de lo que la voluntad puede suprimirlas.
- Un equipo de Tel Aviv entrenó a 40 voluntarios con electrodos faciales, alimentando un algoritmo de aprendizaje automático con los patrones precisos que separan la verdad de la mentira.
- El sistema alcanzó un 73% de precisión, una cifra imperfecta pero que supera con margen significativo cualquier tecnología de detección disponible hoy.
- Los investigadores planean probar el sistema con mentiras más complejas y de mayor peso, con miras a transformar los protocolos de interrogación y seguridad a escala institucional.
La mayoría de las personas son malas detectando mentiras, incluso cuando el mentiroso está sentado frente a ellas. Los investigadores de la Universidad de Tel Aviv decidieron construir una máquina que no cometiera ese error.
Desarrollaron un sistema que analiza los movimientos involuntarios en mejillas y cejas —microexpresiones que ocurren más rápido de lo que el control consciente puede suprimirlas— y usa inteligencia artificial para identificar el engaño. Un algoritmo de aprendizaje automático fue entrenado con datos recopilados de 40 voluntarios equipados con electrodos faciales, aprendiendo a distinguir los patrones del rostro honesto del rostro que miente.
El resultado fue un 73% de precisión. No es perfecto, pero supera con claridad a las alternativas existentes. El polígrafo tradicional mide respuestas fisiológicas —ritmo cardíaco, presión arterial, respiración— que una persona disciplinada o entrenada puede aprender a controlar. Los músculos del rostro son mucho más difíciles de gobernar.
El equipo reconoce que aún está en etapas tempranas. Su próximo paso es probar el sistema con mentiras más complejas y de mayor consecuencia, buscando un espectro más amplio de microexpresiones. Si esas pruebas tienen éxito, las implicaciones podrían redefinir cómo se conducen los interrogatorios y cómo las instituciones verifican la verdad cuando el riesgo es alto.
Most of us are terrible at spotting a liar, even when we're sitting across from one. We miss the tells. We want to believe. Researchers at Tel Aviv University in Israel decided to build a machine that wouldn't.
They developed a system that watches your face—specifically, the tiny involuntary movements in your cheeks and eyebrows—and uses artificial intelligence to flag when you're being dishonest. The system works by analyzing micro-expressions, those fleeting muscle contractions that happen faster than conscious control can suppress them. A machine learning algorithm was trained to recognize the difference between the facial patterns of someone telling the truth and someone lying.
The results were striking: 73 percent accuracy. That's not perfect, but it's substantially better than anything else currently available. Dino Levy, a behavioral neuroscientist on the team, put it plainly: the technology isn't flawless, but it outperforms existing alternatives by a meaningful margin.
The traditional polygraph, the lie detector most people know from crime shows and interrogation rooms, works on a different principle entirely. It measures physiological responses—heart rate, blood pressure, breathing patterns—that typically spike when someone lies. The problem is that these responses can be managed. A person with enough self-discipline, or with proper training, can control their breathing and heart rate well enough to fool the machine. It's a vulnerability that has long frustrated law enforcement and security professionals.
Facial muscles are harder to game. The research team began by fitting 40 volunteers with portable electrodes that measured the precise movements of their facial muscles as they either told the truth or lied. That data fed the machine learning algorithm, teaching it to recognize the signature patterns of deception written across a human face.
The researchers acknowledge they're still in early stages. They want to test their system on people telling more complex, consequential lies—the kind that might trigger a richer, more varied spectrum of micro-expressions. If those tests succeed, the implications could reshape how interrogations are conducted, how security screening works, and how institutions verify truthfulness when stakes are high. For now, the team has built something that works better than what came before, and they're working to understand whether it can work even better still.
Notable Quotes
Not perfect, but much better than any existing technology— Dino Levy, behavioral neuroscientist
The Hearth Conversation Another angle on the story
Why does a machine see what humans miss?
Because we're distracted by context, by what we want to believe, by the person's words. A machine just watches the muscles. It doesn't get fooled by tone or eye contact or confidence.
But 73 percent—that means one in four times it gets it wrong.
True. But a polygraph isn't much better, and it's easier to cheat. You can slow your breathing. You can't easily control your cheeks.
So this is harder to lie your way through?
Much harder. The micro-expressions happen in milliseconds. Your conscious mind doesn't have time to manage them.
What happens when someone knows the system exists? Does that change how they lie?
That's the real question. Right now they're testing on simple lies. When they move to complex ones, to stakes that matter, we'll learn if awareness of the system itself becomes another variable.
And if it works?
Then interrogation rooms change. Security screening changes. The nature of what we can verify about each other changes.