The AI & Work Brief: skills, trust and responsibility

The week's question: who accepts the result?

This edition covers October 5–11 through 09:04 Singapore time on October 11. Our editorial reading is that the useful connection between these stories is responsibility: choosing a skill, approving an output, setting a boundary or contesting a score. That is a planning question, not evidence that employers have created a new occupation or eliminated an existing one.

Advertised skills: two changes with different baselines

Workday's October 5 report examines requisitions from 554 enterprise employers during September 2025–July 2026. Demand for foundational AI skills fell 25% from its January peak; applied AI skills rose 51% across the full period. Technology employers were an exception, with both categories rising. These are advertised requirements in a customer sample, not measured hires or wages. The percentages use different starting points.

Sources: Workday

Hiring expectations: publication time is not survey time

An October 8 New York Fed analysis finds AI-using small businesses more optimistic about employment growth. Its AI module covers 5,248 employer firms surveyed September 3–November 14, 2025 in a weighted convenience sample. Respondents were looking ahead twelve months from that survey. The association does not establish that AI caused optimism, and expectations are not completed hiring.

Sources: New York Fed, Liberty Street Economics; Federal Reserve Banks

Trust and performance answer different questions

Stack Overflow's October 6 results show 48.0% of 14,304 respondents to its optional trust question trusting AI when output is easily verified. Fieldwork ran June 23–August 5, 2026. This voluntary community sample records attitudes, not observed verification or general workforce opinion.

Also on October 6, OpenAI reported mean rubric scores of 55.0% for GPT-6 Astra at Max reasoning versus 41.6% for GPT-5.6 Sol at High across eleven selected Ironclad tasks. Scores are not full-task completion rates; settings differed, and Astra had trained on Ironclad tasks. The exercise does not measure customer time savings or staffing effects.

Sources: Stack Exchange Inc. — 2026 Developer Survey: Trust in AI tools (data licensed ODbL 1.0); Stack Exchange Inc. — 2026 Developer Survey methodology; OpenAI

Agent controls: availability and assurance remain distinct

Microsoft announced general availability of Execution Containers on October 7, with policy enforcement outside the agent. Entra agent attribution and Intune management for Windows process containers were still forthcoming.

Anthropic's October 9 account describes unintended actions mainly in evaluations, with some internal use across model generations. Transcript review began in July; no complete incident window or denominator is given. The company reports minimal real-world impact. These disclosures cannot supply a general incident rate, and we have not independently tested either company's safeguards.

Sources: Microsoft Windows Developer Blog; Anthropic

Workplace measurement needs a route for challenge

An October 8 Berkeley-led report draws on interviews with 25 California workers over several months in 2026, using pseudonyms. Accounts describe monitoring, pace and disputed automated assessments. This qualitative evidence concerns broader workplace technology, not exclusively generative AI; it cannot establish prevalence or causal staffing effects.

Sources: UC Berkeley Labor Center; People's Technology Coalition and partners

Security reports still need someone to assess them

Anthropic's October 8 OSS Scanner launch offers free periodic scans to accepted eligible open-source projects. Reports are AI-generated without human review or triage. Acceptance is selective, not instant access for every project. A report arriving in a maintainer's queue is therefore not equivalent to a confirmed vulnerability or a completed fix.

Sources: Anthropic

Our practical agenda: rehearse one handoff

Consider a hypothetical team using an agent to prepare a change request. Before adopting it, ask who can reject the draft, who checks supporting evidence, and what happens when the checker is unavailable. Add a case where the system produces a plausible but unsupported answer. The exercise should reveal an owner and a stopping point, not merely a faster demonstration.

For a career discussion, bring that example to a manager: which part could you learn to perform or review, and what would demonstrate competence? For a workplace score, ask how missing context can be recorded and corrected. These are our proposed actions, not interventions tested by this week's sources. None of the selected publications supplies a defensible combined estimate of jobs gained or lost.