WIRED reports that a flaw in ChatGPT’s Mac app could let a harmful script reach data through a trusted process. OpenAI says it fixed the flaw. The report does not show that users were breached.
Cameras inside driverless cabs may help keep riders safe. WIRED also asks who can see the video, how long it stays, and what else it might be used for. The report does not say how often staff watch live feeds.
Thore Graepel argues that an AI answer can sound like careful thought without showing what evidence changed its mind. His essay makes a case for checks we can inspect. It does not prove every AI system fails to reason.
Jérémie Lumbroso’s research preprint asks how a person can keep control when handing a choice to an AI system. Its key test is plain: what new fact would make the decision change? This is a proposal, not a field test.
In two simulated factory studies, people using AI advice felt more sure of their choices with less effort. Results after the advice went away were mixed. The studies do not show what would happen in a classroom.
In a writing study of 398 people, a step that asked users to engage before AI generated text shifted more effort toward their own writing. It did not add to total task time. One study cannot show a lasting gain in learning.
