_1777294237270_124456903.png&w=64&q=75){kind=small}
Predicting AI hallucination onset
Researchers at George Washington University have developed a mathematical equation that predicts when artificial intelligence will hallucinate—that is, when its responses shift from being reliable to unreliable or harmful. The equation is based on a mathematical model of how answer selection occurs within neural networks, where the proximity of the model to a tipping point is determined by the direction of its attention during processing. Tests conducted on seven AI models from three different companies showed that their predictions matched actual behavior in 18 out of 19 cases, helping to improve system safety and reduce potential errors. The researchers noted that increasing pressure on the attention mechanisms may lead to the emergence of inaccurate responses or hallucinations, while using clear language linked to the question is believed to reduce distraction.
Notice: This Is an AI-Generated Summary
Comments (0)