Chatbots that see but do not believe: AI failures in the face of evidence

Published on May 29, 2026 | Translated from Spanish

Artificial intelligence assistants like ChatGPT or Gemini have a dangerous blind spot: they ignore what they see. In recent tests, these systems persisted in their errors even when shown videos demonstrating their failure. For anyone seeking scientific or medical information, this is a clear warning: don't assume they are right just because they sound confident.

cinematic wide shot of a person holding a smartphone displaying a paused video of a clear scientific experiment, while a glowing laptop screen beside it shows a chatbot interface with a red error warning icon and a speech bubble containing a deliberately crossed-out visual proof symbol, the person pointing at the video with a frustrated expression, the laptop keyboard reflecting blue light, the chatbot window showing a persistent false statement in bold, contrasting the real evidence on the phone, dark room with focused desk lamp, photorealistic technical illustration, dramatic shadows highlighting the disconnect between visual data and AI output, ultra-detailed screen reflections, engineering visualization style

Selective Blindness in Code 🤖

The technical problem lies in how these models process information. They are trained on static data and lack a mechanism to update their knowledge in real time. When presented with a video that contradicts their response, they do not interpret it as a correction, but as conflicting data that they ignore. Thus, they repeat the error without learning. This is due to their architecture: they prioritize previous statistical patterns over the new evidence they receive.

The Student Who Never Admits Their Mistake 🧠

It's like that classmate who insists that 2+2 equals 5, and when you put four apples in front of them, they say the apples are lying. Chatbots are experts in digital excuses: if the video shows they are wrong, they respond that the video is wrong or not relevant. At least a human, after seeing the evidence, looks foolish and corrects themselves. These programs do not. So you know: if a chatbot tells you the sky is green, go outside and look before buying new glasses. 😉