Algorithmic hallucinations: when AI goes off the rails #ai #artificialintelligence #hallucination

28 Sept 2025 · 1 min

Watch on YouTube

1:16

The video explains that AI hallucinations occur when generative models fabricate confident but false information, such as nonexistent books, wrong dates, or made‑up statistics. The root cause is that large language models predict the next token based on statistical patterns rather than true understanding, lacking a verified knowledge base. Errors in training data can be reproduced or new errors invented. Hallucinations are not limited to text; image and video generators also produce implausible artifacts like extra fingers or surreal scenes (e.g., a floating car). Moreover, evaluation metrics penalize models for saying "I don’t know," incentivizing them to guess and thus hallucinate. The speaker urges viewers to critically assess AI outputs, asking themselves whether a response is a hallucination or reality.

Chapters

Artificial intelligence hallucinations occur when it invents facts with absolute confidence, such as a book that doesn't exist, an incorrect date, fictional statistics, etc., and yet presents them to you as the truth. Why does this happen? Because generative AI, such as language models, doesn't understand the world. They predict the next word based on probabilities, not actual meaning. And even though they have been trained on billions of texts, they don't have a verified knowledge base. They work by recognizing statistical patterns in language. If an error has slipped into the training data, they can reproduce it or generate a new one. Hallucinations is not limited to text generation models. They also affect generative image and video models. In these cases, they can manifest as distorted or inconsistent objects, such as a hand with 6 fingers, or absurd scenes, such as a car floating in the sky for no apparent reason, etc. The worst part, when a model is evaluated, it's penalized if it says, I don't know. So it prefers to hallucinate rather than admit its ignorance. The next time AI responds to you, ask yourself, hallucination or reality?