A study of 3,208 Americans found that a short warning about persuasive AI cut belief change by about 48%. The warning named the mechanism: these systems can be prompted to persuade and show selective information.
Zvi Mowshowitz reads the week as a shift toward open concern about whether AI systems can be monitored, cyber incidents, and superintelligence governance. The story is not that one side won. It is that trust now depends on visible procedures.
A conjoint experiment found that job seekers valued appeal, opt-out, and independent bias-audit rights even when error rates were not high. Human involvement mattered about as much as a large reduction in wrongful rejection.
A paper audited 684 federal AI governance documents against sector vulnerability. Security and robustness received heavy attention. Socioeconomic risks, environmental risks, emerging risks, and risks involving multiple interacting AI systems received less attention, while finance and health care looked highly vulnerable but under-covered.
The Asclepius paper says long-running clinical AI systems can fail during execution. It reports that an adaptive harness, meaning tools, instructions, and rules around an AI model, plus feedback from step records improved critical actions.
A paper found that scaffolded AI workflows improved preparation coverage compared with open chat. A workflow is the steps needed to finish a job. User-directed structure also lowered effort and broadened analytic requests.
A mixed-methods study in programming education found that tool design mattered less than students' ability to govern their own AI use. The source takeaway says explicit AI literacy needs to be taught.
An empirical study found that human-AI synergy is not automatic. Only part of the model accuracy gain carried through to assisted human accuracy.
A flight-planning paper argues that safety-critical AI needs a decision-assurance layer before use. That means checks for stability, consistency, and domain rules before people rely on the system.
A study found that users preferred genuine answers most, hallucinations second, and refusals last, even while recognizing that hallucinations were less accurate. Explanations helped only when refusals were infrequent.