The more specific the question, the easier it becomes to invent something that sounds right
En un momento en que millones de personas han comenzado a tratar los chatbots como si fueran motores de búsqueda, el investigador Ricardo Baeza-Yates —Premio Nacional de Ciencias Aplicadas de Chile y estudioso de la inteligencia artificial durante décadas— advierte que esta confusión conceptual entraña un riesgo profundo: los sistemas de lenguaje no distinguen la verdad de la ficción, sino que predicen patrones, y cuando fallan, lo hacen de manera invisible y convincente. La pregunta que subyace no es tecnológica, sino civilizatoria: ¿puede la confianza colectiva sobrevivir a una herramienta que fabrica realidades con la misma fluidez con que las describe?
- Los chatbots no mienten como mienten los humanos —simplemente generan la siguiente palabra estadísticamente probable, sin ningún acceso real a la verdad.
- Cuanto más específica es la pregunta, mayor es el peligro: casos judiciales inventados, artículos científicos inexistentes y recuerdos falsos emergen vestidos con el lenguaje de la certeza.
- El verdadero punto de quiebre es conductual: los usuarios confían en las respuestas el 99% de las veces sin verificarlas, convirtiendo ese margen de error en un canal masivo de desinformación.
- Baeza-Yates señala que la amenaza trasciende al individuo —la IA generativa sin regulación erosiona el suelo común sobre el que se sostiene el discurso democrático.
- La carrera entre el daño y la regulación está abierta, y por ahora, la tecnología lleva ventaja.
Ricardo Baeza-Yates lleva décadas estudiando cómo funciona la inteligencia artificial y, sobre todo, cómo falla. Con doctorados y cátedras en tres continentes y el Premio Nacional de Ciencias Aplicadas de Chile en 2024, observa desde Silicon Valley una transformación que le preocupa profundamente: el mundo ha empezado a usar los modelos de lenguaje como si fueran buscadores. Es, a su juicio, un error conceptual fundamental —y lo más peligroso es que casi nadie lo reconoce como tal, porque el sistema funciona la mayor parte del tiempo.
Lo que estos sistemas hacen no es buscar ni verificar: predicen. Dada una secuencia de palabras, generan la siguiente con mayor probabilidad estadística, encadenando predicciones hasta construir algo que suena a respuesta. El resultado puede ser coherente, autoritario en su tono, y completamente inventado. El peligro se agudiza con la especificidad: una pregunta vaga puede apoyarse en conocimiento común bien absorbido, pero una consulta sobre un caso judicial concreto, un artículo científico preciso o un evento personal sitúa al sistema en territorio donde no puede saber si algo existe. Entonces hace lo único que sabe hacer: genera palabras en un orden estadísticamente sensato. El resultado es un recuerdo falso con apariencia de hecho.
Baeza-Yates explicó esta dinámica con la precisión de quien la ha pensado durante años. La especificidad misma se convierte en vulnerabilidad, porque desplaza la consulta hacia zonas donde la verificación es difícil o imposible. Decisiones judiciales fabricadas, investigaciones inexistentes, momentos inventados de una vida ajena —todo ello plausible, todo ello erróneo.
El problema de fondo no es la tecnología en sí, sino el comportamiento que ha generado: millones de personas que confían en estas respuestas el 99% de las veces sin contrastarlas. Ese uno por ciento de fallo se convierte en un vector de desinformación a escala, tanto más difícil de detectar cuanto que la información falsa emerge de patrones genuinos y suena verdadera. Para Baeza-Yates, las consecuencias alcanzan al corazón de la democracia: cuando las personas no pueden distinguir la fabricación del hecho, el suelo sobre el que se sostiene la conversación pública comienza a ceder. La pregunta urgente es si la regulación llegará antes de que el daño se vuelva irreversible.
Ricardo Baeza-Yates has spent decades studying how artificial intelligence works and how it fails. He holds doctorates and professorships across three continents—Barcelona, Stockholm, Santiago—and now works from Silicon Valley, where he watches the technology evolve in real time. In 2024, Chile awarded him its National Prize for Applied Sciences and Technology. He is, by any measure, someone who understands what these systems actually do.
What troubles him most is not what chatbots are designed to do, but what people have begun asking them to do. Over the past year or so, he has watched the world treat these language models as search engines. It is, he believes, a fundamental conceptual mistake—one that most people do not recognize as a mistake because the systems work most of the time. When they fail, the failure is often invisible.
Chatbots do not lie in the way humans lie. They do not know the difference between truth and falsehood. What they do is predict patterns. Given a sequence of words, they generate the next word most likely to follow, based on the vast ocean of text they were trained on. String enough of these predictions together, and you get something that reads like an answer. It sounds coherent. It feels authoritative. And it can be entirely fabricated.
The danger sharpens as questions become more specific. A vague query might draw from common knowledge the system has absorbed reliably. But ask a chatbot about a particular legal case, a specific scientific paper, or an event from your own life—something narrow enough that few people could verify it—and the system faces a different problem. It has no way to know whether such a thing exists. So it does what it does best: it generates words in a statistically sensible order. The result is a false memory, dressed up in the language of fact. It is plausible. It is wrong. And the person reading it has no reason to doubt.
Baeza-Yates explained this dynamic to Mundo Deportivo with the precision of someone who has thought about it carefully. The more specific the question, he said, the easier it becomes for the system to invent something that sounds right—something that fits the statistical patterns it learned, even though it does not exist. Fabricated court decisions. Fabricated research. Fabricated moments from someone's past. The specificity itself becomes a liability, because it moves the query into territory where verification is difficult or impossible.
The real problem is not the technology. It is what happens when millions of people begin using it as a replacement for search engines, trusting responses 99 percent of the time without checking them. That one percent of failure becomes a vector for misinformation at scale. And because the false information is plausible—because it emerges from genuine statistical patterns—it is harder to catch than an obvious lie.
Baeza-Yates has also been thinking about what this means for democracy and public discourse. Unregulated generative AI, he believes, poses a threat not just to individual users but to the fabric of shared reality itself. When people cannot distinguish fabrication from fact, when they trust systems that have no concept of truth, the ground beneath democratic conversation begins to shift. The question now is whether regulation will catch up before the damage becomes irreversible.
Notable Quotes
The world now uses these chatbots as search engines, which is a grave conceptual error, but people don't notice because if it works 99% of the time, it seems fine. Sometimes it invents things, because it doesn't know if what it says is true or not.— Ricardo Baeza-Yates
The more specific the question, the easier it is for the system to invent something that has statistical coherence in word sequences but doesn't actually exist—like legal cases, scientific articles, or events from someone's life that few people can verify.— Ricardo Baeza-Yates
The Hearth Conversation Another angle on the story
When you say chatbots don't lie, what do you mean? Isn't generating false information the same thing?
Not quite. A lie requires intent—you know something is false and you say it anyway. A chatbot has no knowledge, no intention. It's pattern-matching at scale. It generates the next word that statistically fits, and if that sequence happens to describe something that never existed, the system has no way to know.
So the user is the one who gets deceived, not the system doing the deceiving.
Exactly. And that's why treating these systems like search engines is so dangerous. A search engine retrieves things that exist. A chatbot generates plausible sequences. When you ask Google for a legal case, it finds documents. When you ask a chatbot, it might invent one that sounds entirely real.
Why does specificity make it worse?
Because vague questions draw from common knowledge the system absorbed reliably. But ask about something narrow—a particular ruling, a particular study—and the system has no anchor. It just generates words that fit the pattern. The specificity is what makes the fabrication convincing.
And people don't verify because it works 99 percent of the time.
Right. That's the trap. One percent failure rate feels acceptable until you realize that one percent is where the misinformation lives. And because it's plausible, it spreads.
What worries you most about this?
That we're building a world where people can't trust what they read, because they can't tell what came from reality and what came from a statistical pattern. That's not just a technology problem. That's a democracy problem.