Richmond Fed EB 26-27: the job-finding drop is real — and less-exposed workers still account for about half of it.
What happened: Richmond Fed Economic BriefNo. 26-27 (August 2026), Borovičková and Macaluso, Aggregate or Structural? Diagnosing the Rise in Unemployment, decomposes the U.S. job-finding decline from mid-2023 through March 2026. The largest within-group drop sits with strongly attached workers in high-AI-exposure occupations — but that group is small, so it explains at most about one-third of the aggregate decline. Less-exposed workers (primary and secondary attachment) account for roughly half of the decline; the largest single contribution is secondary workers in low-AI-exposure occupations. Compositional shift across groups is under 3%. Primary workers are about 55% of the population, secondary about 14%, tertiary about 31%. The authors’ bottom line: AI exposure is important but not dominant; aggregate forces remain the dominant factor. Targeted retraining for highly exposed occupations may reach an important segment without fixing most of the slowdown. Views are the authors’ and do not necessarily reflect Richmond Fed or Board policy. This is an unemployment-outflow diagnostic — not Atlanta Fed WP 2026-4’s executive survey (n≈750), not Brookings Metro’s adaptive-capacity residual, not Census CES-WP-26-25/27 adoption or early-career QWI packages, not NY Fed’s 61%/51% firm-use shares, not PwC’s job-ad barometer, and not a 2026 layoff census.
Why it matters: The sharpest AI-linked job-finding hit is concentrated — but most of the aggregate slowdown still lands on workers with little direct AI overlap.
EIA September STEO: record U.S. electricity generation in 2026 — data centers and manufacturing still drive sales growth.
What happened: EIA’s 9 September 2026 press release on the September Short-Term Energy Outlook projects U.S. electricity generation up 2.2% to a record 4,368 billion kWh in 2026, then up another 1.7 percent in 2027. Electricity sales rise to 4,135 BkWh in 2026 (almost 2% vs 2025) and 4,211 BkWh in 2027 (+2%). Drivers named on the page: data-center development and manufacturing in commercial and industrial sectors. Despite a pause in new data-center projects in Texas, West South Central remains the largest regional contributor to sales growth. STEO landing detail for the same September release: WSC sales 761 BkWh (2026) / 790 BkWh (2027) vs the August STEO path of 765 / 829 (−0.5% / −4.7%). Model inputs were finalized 3 September 2026. Generation-share context on the press page (rounded): natural gas about 40% through 2025, 2026, and 2027; coal 17% / 16% / 14%; nuclear 18%; wind 11% / 11% / 12%; solar 7% / 8% / 9%. Companion CO2 path: about 4.9 / 4.8 / 4.8 billion metric tons across 2025, 2026, and 2027. This is a national short-term sales and generation monthly — not IEA Electricity 2026 Demand’s U.S. data-centre share of national load-growth outlook through the decade, not the 4 September Texas load-growth 6% vs 14% card restamped, not AEO2026 server shares, not Gartner’s global data-centre TWh path, and not a campus megawatt announcement.
Why it matters: Even after cutting the WSC path and after the Texas pause, EIA still prints a national electricity-sales record with data centers among the named drivers — a monthly forecast, not another decade-ahead data-centre power path.
UK MHRA National Commission: staged “L-plate” authorisations and lifetime monitoring for medical AI.
What happened: On 10 September 2026, GOV.UK published the Independent Commission’s blueprint on regulating AI in healthcare, established by the Medicines and Healthcare products Regulatory Agency in September 2025 and led by NHS doctors. Evidence came from more than 12,000 people over a year — patients, public, clinicians, healthcare leaders, and industry. Government and MHRA will consider the advice; a formal response comes later — this is not adopted law. Locked recommendations from the news page: staged authorisations for new AI models, framed as “L-plates” (close supervision first, then fuller authorisation); continuous real-world monitoring across a device’s working life rather than a single point-in-time approval; a public search for safety information on specific AI-enabled medical devices, including adverse incidents; and stronger MHRA enforcement powers. Engagement finding locked on the same page: broad public support is conditional on strong safety standards, meaningful human oversight, and transparency about when and how AI is used in someone’s care. Notes-to-editors on the GOV.UK page cite Health Foundation deliberative workshops (March–April 2026) with 78 members of the public in Cardiff, Milton Keynes, and York. This is an advisory UK clinical-AI regulatory blueprint — not live EU Article 50 chatbot/mark/deepfake duties, not the EU GPAI Code’s Articles 53/55 model-provider chapters, not California’s transparency act, not NIST AI 300-1 documentation templates, and not Canada’s transparency consultation.
Why it matters: The UK is proposing learner-plate staged entry and lifetime safety monitoring for medical AI — with a public safety page — before any statute is written.
Nature correspondence: put patients at the centre of medical AI governance — rights still lack a systematic slot.
What happened:Nature published an 8 September 2026 Correspondence by Shen and Yu, Put patients at the centre of medical AI governance (Nature657, 566; DOI 10.1038/d41586-026-02796-8). The locked lede: Lam et al. classify medical AI by autonomy, automation, and operational scope (Nature655, 1129–1132; 2026). They mention informed consent, yet the framework remains centred on allocating responsibility to clinicians, health-care organizations, device manufacturers, and regulators. Patient rights receive no systematic place in its design. This is a correspondence letter — not a patient-outcome RCT, not Lång’s Nature Medicine outcomes-of-human–AI-systems comment, not the npj voice-biomarker foundation-model Perspective, not the npj five-phase evaluation ladder, and not a new locked diagnostic accuracy print. Pair it with the same-day UK Commission’s staged-authorisation and lifecycle-monitoring advice without collapsing letter and advisory report into one claim.
Why it matters: Governance taxonomies that sort tools and institutions can still leave patients outside the design frame — a different gap from missing AUROCs or missing outcome trials.
UNESCO opens a global consultation on education in the age of AI — comments due 15 October 2026.
What happened: UNESCO’s consultation page invites feedback on the discussion paper Sustaining education as a common good in the age of AI: The case for deliberative governance, plus six background papers on digital/AI sovereignty; age-appropriateness; learner safety (“safe enough to scale”); procurement; total cost of ownership; and the right to an explanation. Comments are due 15 October 2026 (window September–November 2026) to aied@unesco.org; feedback is slated to feed policy briefs at Digital Learning Week 2027. Consultation questions name UNESCO principles of human agency, equity, inclusion, safety and ethics. A companion ministers article from Digital Learning Week frames the common-good statement on the 2023 generative-AI guidance and 2024 AI competency frameworks, notes seminars involving participants from more than 100 countries, and recalls the 2021 Recommendation on the Ethics of AI adopted by 194 Member States. This is a normative consultation and working-paper package — not yesterday’s ICT Prize ceremony for Finland and Brazil, not the UNESCO–UNICEF–ITU Charter seven principles restamped, not HEPI’s ~94%/12% UK undergraduate package, and not a multi-country learning-outcomes RCT. Standing context only: Digital Learning Week runs 8–11 September (day 3 of the window).
Why it matters: The education-AI agenda is shifting from prize ceremonies and usage polls toward open governance design — age, safety, inclusion, and cost of ownership on the table before next year’s policy briefs.
High-exposure pain is sharp; aggregate drag is broader.
What happened: Keep the Richmond package locked as less-exposed workers ~half of the job-finding decline; high-AI-exposure attached workers ≤~one-third; composition <3%; aggregate forces dominant. Do not convert the within-group high-exposure drop into a national layoff rate, and keep it separate from Atlanta Fed’s executive “little near-term aggregate loss” survey and Brookings’ high-exposure / low-capacity residual.
Why it matters: Retraining only the most AI-exposed occupations would miss most of the measured job-finding slowdown.
National sales-record monthly is not a decade-ahead data-centre power path.
What happened: Lock EIA’s 4,368 BkWh +2.2% generation print and sales 4,135 / 4,211 beside — not mashed with — IEA Demand’s U.S. data-centre growth-share card and the 4 September Texas 6% vs 14% load-growth cut. WSC 761/790 is a revised regional sales path, still the largest regional sales-growth contributor despite the Texas pause. Forecast inputs froze 3 September; this is not metered 2026 campus kWh.
Why it matters: Mixing short-term national sales with decade-ahead growth-shares invents a single “AI electricity number” neither product prints alone.
What happened: MHRA Commission: >12,000 engaged; staged authorisations; lifecycle monitoring; public safety search; stronger enforcement. Label as advice with a response pending. Keep separate from day-39 EU Article 50 calendar context, yesterday’s GPAI Code Articles 53/55, and NIST AI 300-1 comments still due 16 September.
Why it matters: Clinical-device staged entry and system-level transparency labels answer different compliance questions on different clocks.
Patient-rights letter beside staged-authorisation advice — still not a trial result.
What happened: Pair Shen/Yu’s patient-rights gap with the UK Commission’s L-plate and lifetime-monitoring recommendations without treating either as a bedside clearance claim. Standing clock only: WHO/ITU/WIPO GI-AI4H Hangzhou remains 16–18 September.
Why it matters: Who is centered in the governance frame is a different question from whether a model beat a baseline AUROC.
Consultation window, not a restamped prize ceremony.
What happened: Comments due 15 October 2026; six named governance papers; principles of human agency, equity, inclusion, safety, and ethics; policy briefs aimed at DLW 2027. Do not invent eight ministerial “priorities for action” that were not on the extracted ministers page, and do not restamp yesterday’s Finland/Brazil ICT Prize reach as today’s primary.
Why it matters: Open consultation on age, safety, inclusion, and total cost of ownership is a different education-AI story than ceremony-day creativity awards.
