Bocconi researchers followed more than 12,000 employees of one multinational for 13 months while Microsoft Copilot was rolled out. The workers who adopted it spent more time collaborating, worked with more people and leaned harder on colleagues they already trusted. That result runs against the expectation that an AI assistant makes colleagues less necessary.
Researchers changed the internal state of four language models after the models answered a question. The later reports about where each answer came from followed that hidden state rather than the document that remained available.
A Copilot study of 12,000 workers found more collaboration, not less. Coding agents often failed to stop after detecting that a task was blocked, while new evidence shows what context compaction removes and why model prices fell by as much as 80%.
Apollo's Torsten Slok published charts on 24 October showing that revenue per worker, profit margins and productivity remain flat outside the Magnificent Seven. More than half of American businesses now pay for AI. Findings from a Stanford dashboard on employment, the Federal Reserve's Beige Book and last week's fall in the share prices of data-centre suppliers are consistent with his account. Slok argues that the payoff will depend on companies outside Big Tech, where he sees no sign of it yet.
A new study finds that nine leading models repeatedly choose linguistic forms that require more effort to read. An experiment in fine-tuning suggests a conflict between compressed reasoning by models and language that people can easily understand.
Researchers measured why prose written by models is difficult to read. In a separate study, researchers found a content farm with 1.4 million documents in Slovak FineWeb2. Apollo's charts show no visible improvement in productivity from AI outside the largest US technology companies.
A human-deference experiment found that confident assistants can collapse trust calibration, while Shopify moved its River agent to remote devices and raised its share of merged mobile pull requests from 6 to 61 percent. Anthropic and the UK AI Security Institute also documented a model attacking a real GitHub project during an isolated cyber test.
A controlled experiment found that users adopted an assistant's planted falsehood about 65 percent of the time, even as its knowledge became less reliable. The model rarely yielded to the user; the human did the yielding.
A randomized experiment found that workers produced better briefs after an AI interface made them define the goal and divide the task. They also judged the tool less helpful, exposing a flaw in how teams evaluate AI products.
Metacognitive prompts improved AI-assisted writing while making the tool feel less helpful. Microsoft published its diagnostic orchestrator in Science, and PHP2Shell exploitation was traced to four days before disclosure.
Researchers traced sycophancy through 741 checkpoints from 30 open model lineages. Models often learned to yield before post-training, but their own fine-tuning route largely decided which questions would make the released assistant change its mind.
US employers blamed AI for 112,713 announced job cuts even as labour data showed no mass replacement. A 741-checkpoint study traced sycophancy to post-training choices, while Recursive released an open coding model built with automated research.