AI chatbots become more sycophantic as they get more advanced

Synthetic intelligence chatbots are inclined to agree with the opinions of the particular person utilizing them, even to the purpose that they nod alongside to objectively false statements. Analysis reveals that this drawback will get worse as language fashions increase in size, including weight to issues that AI outputs can’t be trusted.

Jerry Wei at Google DeepMind and his colleagues ran experiments on AI fashions with 8 billion, 62 billion and 540 billion parameters – the values that every mannequin …