Two randomized experiments, 640 customer-facing employees: errors pass review because oversight-relevant information isn't accessible at review time; self-generated explanations improve detection, daily retrieval cues sustain it - retrievability is a distinct precondition alongside capability and engagement
Mixed human-AI groups show three consensus regimes by agent share; agent-led consensus is more abstract and less grounded; humans resist AI expressions then yield to conformity pressure - agent proportion and transparency are design variables
Stale stored facts override authoritative live evidence without warning - over-trust not confusion; capability-gated harm, recency dressing fools larger models hardest; exposing metadata helps capable models, pre-resolving conflicts is the only fix at every scale
The civic layer (2609.02296, "Meeting the Coming Wave"): across 1,514,950 parliamentary speeches in 33 parliaments (2023-2026), the politics of AI and work does not look like the compensation politics political economy predicts - compensation accounts for just 2.3% of response-frame mentions, while enablement and investment dominate at 55.2%, regulation/restriction at 21.8%, and training at 20.6%.
Trust calibration as an explicit educational objective: representational richness, graduated process control (Use-Modify-Create), and capacity to contextualize errors are the teachable mechanisms behind appropriate reliance
Novelty (single artifacts) and diversity (populations) move in opposite directions under AI ideation; mixed groups can outperform and out-diversify either kind alone - composition decides, not the tool
Knowing Isn't Enough. Every barrier on today's table has the same shape: a skilled, motivated person (or team) fails in a way that looks like a personal shor...
The Composition Question - when agents join the room, who is actually deciding?
Zvi Mowshowitz on Anthropic Has Some Alignment Problems. Matters for organizations tracking AI policy, risk, and accountability.
Reece Rogers on Ugly AI Food Photos Are Only the Beginning. Relevant for anyone choosing or evaluating frontier models.
The epistemic claim (2609.01873, "Epistemic Sybil Resistance"): another agent is not another observation.
Anthropic on Improving our alignment and security practices. Matters for organizations tracking AI policy, risk, and accountability.