The Spectacular Hallucinations of AI: Threat and Solutions
AI Confidential · March 23, 2025 · by Sergio & Nox
AI Hallucinations: A Real Problem and Its Solutions
Artificial intelligence has conquered every sector, from healthcare to finance, through education and creativity. But behind this revolutionary force hides an anomaly as fascinating as it is dangerous: the so-called AI hallucinations. These are moments when a model, with the confidence of someone who believes they know everything, invents facts, distorts data, proposes absurd or misleading solutions. And it does so without the slightest doubt, as if the truth were always and in any case on its side.
It is not a detail. When it comes to medical diagnoses, financial analysis or public data, an error of this kind can turn into a disaster. It is as if a surgeon, in the middle of an operation, relied on a wrong map drawn by someone who has never set foot in an operating room.
The good – and at the same time the limiting – thing is that we humans, even as children, have an extraordinary ability: that of grasping the sense of things from very few clues.
AIs, by contrast, often stumble when they must abstract, generalize or simply “understand” for real. Even the most advanced models – like GPT-4 or Claude 2 – can fall into these AI hallucinations. Some behave better: Claude 2, for example, showed only 5% of wrong answers in the most recent tests. But just move up to less refined models like GPT-3.5 or LLaMA 13B, and the error rate can shoot up to 21%. And it is not only a question of data: it is a question of structure, of cognitive architecture, of what distinguishes intelligence from pure computational power.
Research has begun to react. How do we solve the problem of AI hallucinations? Some propose “feeding” the AI with verifiable external sources, some work on the way we word our questions, some build self-criticism mechanisms inside the model itself. But the truth is that, today, the risk remains. And understanding it is the first step to not suffering it.
If you want to go deeper, the report published by Visual Capitalist offers a detailed analysis of the most reliable AI models and of the current techniques to mitigate the phenomenon of AI hallucinations.
Read IBM's view on AI hallucinations
And while artificial intelligence evolves, we must stay more awake than ever. Because behind every perfect answer there may hide a well-built lie, that is, a perfect AI hallucination.
If you are curious about the limits of AI, read this post
“Artificial intelligence is not a substitute for human intelligence; it is a tool to amplify human creativity and ingenuity.”
— Fei-Fei Li, Co-Director of Stanford's Institute for Human-Centered Artificial Intelligence and Professor at the Graduate School of Business.