Research Reveals Larger AI Models Are More Prone to Inaccuracies

As AI language models grow in size and complexity, they increasingly provide incorrect answers rather than admitting uncertainty.

A study published in Nature examined the accuracy of leading AI language models like GPT, LLaMA, and BLOOM.
Researchers found that while larger models are better at handling complex tasks, they are more likely to give incorrect answers on simpler questions.
The tendency to provide confident but inaccurate responses is described as 'ultra-crepidarian' behavior.
Human evaluators often struggled to identify incorrect answers, raising concerns about over-reliance on AI outputs.
Experts suggest that programming models to admit uncertainty could improve reliability, but this may expose the technology's limitations.