Communicating with neural networks can turn us into overconfident fools
Neural networks provide fast and polished answers. This is convenient, but there is a flip side. A recent study has shown that the more often a person turns to AI, the less they use a very important habit in life. And over time, this can lead to very serious consequences.
How neural networks prevent us from thinking for ourselves
Recently, a team of scientists decided to test how having a smart assistant in the form of a neural network affects people’s honesty with themselves. During the experiment, participants were asked tricky questions about visual details from movies — for example, they were asked to recall the color of the team’s sports uniform in the film “Bend It Like Beckham.”
To ensure the experiment’s integrity, the scientists deliberately used a weak language model that most often made mistakes in such specific tasks. The authors did this on purpose to rule out the phenomenon where a person trusts a genuinely correct answer. The results were striking.
In the control group, where people had no access to artificial intelligence, participants honestly admitted “I don’t know” in 44% of cases, and their overall accuracy was 27%. But as soon as people had the opportunity to ask the neural network, their willingness to admit ignorance plummeted to 3%, which was the main surprise for the study’s authors. Participants began copying the algorithm’s answers, causing their overall accuracy to drop to just 9%. Most frighteningly, people’s confidence in their own correctness rose from 30% to 76%. Even a small monetary reward for correct answers could not restore critical thinking to normal levels.
Why we believe neural network mistakes
What the European scientists discovered perfectly complements a theory proposed earlier this year by researchers from the Wharton School of Business. They introduced a special term: cognitive capitulation. This concept describes a state where a person completely delegates the decision-making process to a machine and turns off their internal filter of skepticism.
In the study, participants agreed with completely incorrect chatbot answers in 80% of cases. Cognitive capitulation syndrome manifests through several clear signs:
- a person accepts the machine’s answer without the slightest attempt to verify the facts in other sources;
- the subjective feeling of being right paradoxically increases;
- slow, analytical thinking skills gradually weaken due to the lack of constant practice.
The publication The Next Web notes that the new European study adds an even more alarming detail to this picture. The main problem is not that people trust factual errors or neural network hallucinations. The primary threat is that the very presence of AI suppresses the habit of doubting and recognizing gaps in one’s own knowledge. We have become so accustomed to receiving instant answers in the form of confident text on a screen that we have forgotten how to stop and reflect.
Why neural network answers are dangerous
According to study co-author Valerio Capraro, the ability to say “I don’t know” is critically important for any intellectual process. This phrase represents an acknowledgment of the boundaries of our mind and serves as a starting point for genuine learning. If we believe the answer is already in our pocket, we permanently stop searching for the truth.
Scientists are particularly concerned about children and teenagers who are growing up surrounded by systems that always produce a ready-made result. The recent integration of AI-generated answers into Google Search only exacerbates the situation, because instead of a familiar list of links for independent study, the user immediately receives a generated summary that is presented as absolute fact.
The organization Common Sense Media has already called such interface design an unacceptable risk for schoolchildren and students. Modern AI products are designed to always provide an answer, and they almost never acknowledge their own incompetence. People who use them simply copy this behavioral model.
Ultimately, neural networks remain an excellent tool for sorting data and routine work, but they should not replace our own thinking. This study serves as a reminder that we need to relearn how to doubt. In a world where an algorithm is ready to deliver any beautiful nonsense with absolute confidence, the ability to honestly say “I don’t know” becomes the main hallmark of a truly alive and strong mind.
