Some AI models work through problems step by step in writing before they answer. New research found they struggle to control what shows up in that written reasoning.
That makes their working notes a useful window into what they're really doing, so safety teams can read along and catch problems before they reach you.
Turns out a bad poker face is a feature.
A bill in Illinois, SB 315, would be the first US law requiring independent audits of how frontier AI developers manage safety.
Read the storyCanada's federal government has introduced legislation aimed at AI chatbots, the kind people now use for homework help, advice and customer service.
Read the storyUtilities using Oracle's AI-driven Opower programs asked customers to ease off electricity at the busiest moments, with personalised tips and rebates for doing it. Across a record 73 events in 2026, the savings added up to 1 gigawatt of peak demand.
Read the storyOne hopeful, sourced AI story a day on X.