McKinsey’s trust survey frames the big shift from “AI that might say the wrong thing” to agentic systems that can do the wrong thing—trigger actions, misuse tools, or exceed intended scope. Useful if you’re trying to scale AI beyond pilots: it lays out where organizations tend to be weakest (governance, risk processes, and ownership), and what “trust maturity” looks like in practice.
OpenAI is expanding shopping in ChatGPT into a more structured product-discovery experience: richer browsing, comparisons, and clearer paths from “I’m exploring” to “I’m deciding.” The strategic signal is distribution—AI is competing with search and marketplaces by becoming the first place people go when they don’t yet know what they want.
Google’s Lyria 3 Pro is a music-generation model aimed at creators: longer tracks (up to ~3 minutes), more control over song structure (intros/verses/choruses), and availability across more Google surfaces. The non-technical takeaway is productization: “models” matter most when they show up as usable features inside tools people already use to create.
DeepMind summarizes new work on the risk of AI-enabled persuasion and releases measurement tools for evaluating “harmful manipulation” more rigorously. The value for non-technical readers is the shift toward testable safety claims: instead of vague assurances, this is an attempt to define what counts as manipulation and measure it systematically.
Brookings argues the near-term risk isn’t only job counts—it’s job quality, bargaining power, and who benefits from productivity gains. The actionable part is a concrete “people-first” agenda: build scalable training pathways, strengthen institutions that support workers, and shape deployment so AI augments work rather than silently degrading it.
#workforce#labor#policy
Going Deeper
Optional reads for those who want more. (Some may be behind a paywall)
Inside our approach to the Model SpecOpenAIA readable governance explanation of how OpenAI defines “model behavior” and updates it as capability and usage scale.
Where to look for generative AI risksMIT SloanA clean non-technical framework: “embedded” vs “enacted” risk, plus a practical inventory + ownership approach for reducing exposure.