Researchers have investigated the phenomenon of 'hallucinations' in large language models, where the models generate outputs that are not based on the input data. The study found that higher temperatures in the model's sampling process lead to more plausible but incorrect answers, while lower temperatures result in factually accurate ones. The authors argue that this is due to exposure to subjective and socially diverse training data, rather than any cognitive ability. They also claim that a model's self-reports of emotion or sentience should be considered hallucinations, making it difficult to distinguish between machine consciousness and advanced hallucinations. --- Why it matters: This study matters because it challenges the idea that AI can truly understand and report its own emotions or thoughts. It highlights the limitations of current language models and raises questions about the nature of machine consciousness. Source: https://arxiv.org/abs/2608.18816