As AI gets smarter, it appears to trust humanity less Researchers have adapted a decades-old psychological test that measures how people perceive human nature into a new metric for large language models. They tested models across the spectrum, including the entire GPT series, and found a surprising pattern: AI models tend to view humans as unpredictable, complex, and difficult to trust. The researchers reported a significant negative correlation between model intelligence and trust in humans. As the models became more capable, their assessments of human nature became more pessimistic. But the biggest surprise came when they checked how the models were trained. Basic models, trained mostly on web text, generally showed a more positive view of humanity. After additional training designed to make them more confident and better suited to human expectations, their estimates became significantly less reliable. The researchers suggest that teaching an AI to recognize harmful behavior, navigate moral dilemmas, and follow safety rules may also affect how it values ​​human nature. No one necessarily programmed these models to mistrust us. Instead, the training process may have encouraged them to associate human behavior with risk, unpredictability, and possible harm. source.

As AI gets smarter, it appears to trust humanity less Researchers have adapted a decades-old psychological test that measures how people perceive human nature into a new metric for large language models. They tested models across the spectrum, including the entire GPT series, and found a surprising pattern: AI models tend to view humans as unpredictable, complex, and difficult to trust. The researchers reported a significant negative correlation between model intelligence and trust in humans. As the models became more capable, their assessments of human nature became more pessimistic. But the biggest surprise came when they checked how the models were trained. Basic models, trained mostly on web text, generally showed a more positive view of humanity. After additional training designed to make them more confident and better suited to human expectations, their estimates became significantly less reliable. The researchers suggest that teaching an AI to recognize harmful behavior, navigate moral dilemmas, and follow safety rules may also affect how it values ​​human nature. No one necessarily programmed these models to mistrust us. Instead, the training process may have encouraged them to associate human behavior with risk, unpredictability, and possible harm. source.