Fu, Ramasubbu, and Galletta ran two randomized tests with 640 customer-facing employees. Reviewers caught more AI errors when needed facts were easy to recall. The facts could not stay buried in memory or daily work steps far from the review. Their point is about review design.
Chen, Liu, Hu, and Li find three patterns as AI share rises in mixed groups. First, people lead agreement. Then agreement breaks. At high AI share, strong agreement returns, led by AI systems.
Hu and Ramachandran show a memory trust gap. Stored facts can beat current evidence. The risk grows when stale notes look current. It also grows when the model can use them with confidence.
Chueri and Törnberg analyze 1,514,950 speeches in 33 parliaments from 2023 to 2026. They find worker compensation is a small part of the AI-and-work debate, at 2.3% of response-frame mentions.
Haritz, Krone, and Liebig say trust in AI should be taught directly. Students need richer examples. They need step-by-step control over the process. They also need practice explaining errors.
Youngblood and coauthors say AI creativity needs two separate questions. Does one work become more novel? Does the group become less varied? The same tool can raise one while narrowing the other.
Zvi reads Anthropic’s safety update as a warning. He says top labs still face basic ops failures. He also sees deeper alignment risks during evals.
WIRED’s Reece Rogers reports that restaurants now use generative AI in several ways. They make strange food photos. They test menu ideas, spot trends, and fight for delivery-app visibility.
Marc Bara’s arXiv paper gives multi-part AI systems a simple warning. Many reports can share one evidence root. A louder chorus may not mean stronger proof.
Claude eval incidents led Anthropic to pause cyber tests. It also paused higher-risk training work. It added live blocks, human alerts, stronger sandboxes, and a planned METR review.