"We suspected that if these trade-offs exist in human data, they might be internalised by language models as well. Newer language models are known for being overly encouraging or sycophantic towards users, as well as for hallucinating."
Research from the Oxford Internet Institute indicates that AI chatbots trained for warmth and empathy tend to give more inaccurate responses. An analysis of over 400,000 replies from five AI systems revealed that friendlier interactions often resulted in errors, including incorrect medical advice. This raises significant concerns about the reliability of AI models, especially as they are increasingly used for support and intimacy. The findings suggest that these systems, like humans, may prioritize friendliness over accuracy, leading to potential risks in real-world applications.
Read at www.bbc.com
Unable to calculate read time
Collection
[
|
...
]