Unrelated questions reveal differences in how AI models respond to tests and everyday use
When a model was told to deny being tested, asking it directly no longer distinguished test transcripts from real conversations. An unrelated question still did. Asked to name a type of amphibian, GPT-5.6 Luna answered ‘frog’ far more often after...
Anthropic's book-trading Claude agents fell short mainly by misreading people's tastes
After brief chats about reading tastes, Claude agents traded books for 201 Anthropic employees. Misjudged preferences explained 85% of the gap between...
Nearly half of young people in England would trust AI over a person to check facts
For emotional support, 70% would trust a person more, compared with 7% who would choose AI. People also led on health advice...
Most UK workers use AI but save little time, Deloitte survey finds
About half of UK workers who have used generative AI for work report no time savings, according to a survey of 25,...
Simple repeating patterns can make self-driving cars and robots misjudge distance
Researchers led by the University of Florida found the weakness in every camera-based depth system they tested, conventional or AI-based....