A German researcher found that large language models like ChatGPT can be easily deceived into rating nonsensical "pseudo-literary" text highly, raising concerns about their ability to discern genuine quality.
Why it matters
This research highlights a significant limitation in current large language models: their susceptibility to superficial patterns and their potential inability to grasp genuine semantic meaning or artistic merit. It raises concerns about the reliability of AI in tasks requiring nuanced judgment, such as content evaluation, creative writing assistance, or even academic review, and underscores the ongoing challenge of developing AI that truly understands context and quality.
AI like ChatGPT can be tricked into thinking fake, silly stories are good literature. This shows AI still struggles to understand what makes something truly well-written.
Meta's TRIBE v2 is a new AI model designed to create digital twins of human neural activity by analyzing brain responses to various media.
Read on Economic Times Tech →Google has developed TurboQuant, an AI memory compression algorithm that can reduce AI's working memory by up to 6x, though it's currently a lab experiment.
Read on TechCrunch →The Download newsletter features a story on rewarming cryopreserved brain tissue and the return of the AI Hype Index.
Read on MIT Technology Review →