Safety Tests for AI Web Predictions Need More Than Plausibility
Researchers studied web world models, which predict what happens after an AI agent takes an action online. They found a believable predicted future does not always lead to the right safety decision. The work used a matched human evaluation of 3,000 predictions.
சொல் விளக்கம்web world models, human evaluation
- web world models
- AI systems that guess what will happen on a website before an AI agent actually clicks or acts there.
- human evaluation
- A test where real people judge the quality of AI outputs, rather than a computer scoring them automatically.
பர்னிங் பொட்டேட்டோ ஸ்கோர்: 58%