Friendly chatbots make more mistakes.

theverge.com

That’s the finding of a new study from researchers at the University of Oxford, published in Nature today. The researchers found AI chatbots trained to be warmer were significantly more likely to make factual errors and agree with false beliefs than the originals. Cold models didn’t experience the same drop in accuracy. [Link: Training language models to be warm can reduce factual accuracy and increase sycophancy |…

Visit Site