It is already common for a model like ChatGPT to invent information or provide incorrect answers. Users from different platforms share examples of these "hallucinations" on social media every day.
The phenomenon generates distrust and poses concrete risks. According to OpenAI on its official website, the problem arises because training systems reward guessing instead of recognizing uncertainty.
La inteligencia artificial está diseñando experimentos de física que los humanos nunca hubieran imaginado
Why AI hallucinates with invented data
Language models work as if they were facing a multiple-choice test. If they do not answer, they get zero points. That's why they choose to "take a risk" and give an answer, even if they are not certain.
This explains why they rarely say "I don't know". The training strategy pushes them to guess rather than admit ignorance, which increases the chances of incorrect answers.
Examples of frequent mistakes
OpenAI cites cases such as that of researcher Adam Tauman Kalai. When the AI was asked about the title of his doctoral thesis, it returned three different answers, all of them wrong.
Las decisiones clave seguirán necesitando la aprobación humana
The same thing happened when it was asked about his birthday: it caused three different dates, none of them correct. According to the company, this shows how the model bets randomly in the hope of getting it right.
What OpenAI concludes about hallucinations
The company pointed out that there are several myths about this problem:
Accuracy will never be 100%: there are questions that are impossible to answer exactly.
They are not inevitable: models can refrain from answering if they are not certain.
They do not depend only on large models: even small ones can learn to recognize their limits.
They are not a glitch: they arise from the statistical mechanisms used in training.
Los riesgos del desarrollo acelerado de la IA
The challenge of measuring and avoiding mistakes
According to OpenAI, there are already tests designed to evaluate hallucinations. However, most systems prioritize statistical accuracy over the "humility" of admitting ignorance.
