Do AI language models ‘understand’ the real world? On a basic level, they do, a new study finds
A new study by Brown University researchers found that AI language models have some understanding of the real world, encoding causal constraints similar to human judgments. The study analyzed the internal workings of several AI models to determine their ability to distinguish between plausible and implausible events.
Researchers at Brown University conducted a study to determine if AI language models have an understanding of the real world. They analyzed the internal workings of several AI models, including Open AI's GPT 2 and Meta's Llama 3.2, to see if they could distinguish between commonplace, improbable, impossible, and nonsensical events. The study found that models of sufficient size developed distinct mathematical patterns that correlated with plausibility categories, achieving around 85% accuracy. The researchers also found that the AI models captured human uncertainty about ambiguous statements. The study's findings were presented at the International Conference on Learning Representations in Rio de Janeiro, Brazil.
This content was automatically generated and/or translated by AI. It may contain inaccuracies. Please refer to the original sources for verification.