OECD’s Capability Gap map: office work sits closest to today’s AI; care, law, and teaching sit farthest.
What happened: OECD Artificial Intelligence Papers No. 59, The OECD AI exposure measure: Mapping the OECD AI Capability Indicators to occupations (approved 19 May 2026; July 2026 revision), builds an occupational AI Capability Gap — demand for each capability minus the current AI frontier. Nine capability domains are mapped to O*NET tasks. Lower gap means AI already matches more of what the job requires, so potential exposure is higher. Current AI is placed roughly at levels 2–3 on a five-point scale; occupational demands can reach 5. With today’s AI levels, the effective maximum of the total index is about 23 (theoretical upper bound 36). Major-group totals locked from the July executive summary: office and administrative support 0.8 (lowest gap / highest potential exposure); production 1.8; food preparation and serving 2.4; sales and related 2.6. Highest gaps (lowest potential exposure): community and social service 6.4; legal 5.8; educational instruction 5.8; healthcare practitioners and technical 5.6; management 5.5; protection 5.2. Domain averages on the page: social interaction, problem solving, and metacognition each about 1.1; vision 0.2 after the July correction; language 0.4. The paper is explicit: this is potential exposure to AI capabilities, not measured adoption, not realised job loss, and not a wage series. Keep it separate from BLS's newest official projection categories and from ECB's observed U.S. reallocation wedge from the prior edition.
Why it matters: After a week of projections and observed reallocation, this is a different instrument — how close current AI is to what each occupation actually demands. Office support is closest; care, law, and teaching still have the widest gaps.
PJM’s 20-year peak path still races — while near-term data-center requests get derated unless firm.
What happened: PJM’s Inside Lines recap of the 14 January 2026Long-Term Load Forecast Report is an RTO planning product for the Mid-Atlantic footprint — not yesterday’s EIA STEO stress test and not a global IEA TWh path. Summer peak growth is projected at 3.6 percent a year over 10 years and 2.4 percent a year over 20 years. The 2021 Long-Term Load Forecast’s 10-year summer rate was 0.3 percent a year — “less than one-tenth” the current path. Summer peak is forecast at about 222,000 MW in 2036 (a 10-year rise of nearly 66,000 MW) and surpasses 253,000 MW in 2046 (about +97,000 MW over 20 years). Over 15 years, summer peak climbs about 85,000 MW to more than 241,000 MW. Winter peaks grow even faster: 4.0 percent a year over 10 years and 2.7 percent a year over 20 years. Near-term years through 2032 are lower than the 2025 report after EV and economic updates and tighter vetting of large-load requests; summer 2026 large-load attribution is down 0.7 percent vs the prior report, with economic activity down 0.5 percent and a slight EV markdown. Near-term additions need firm commitments (ESO/CC); longer-term projects are treated as non-firm and derated. Current generating capacity is about 182,000 MW; the record summer peak remains 166,929 MW (2006). Named large-load zones include AEP, ATSI, APS, BGE, COMED, Dayton, DQE, JCPL, METED, PECO, PEPCO, PL, plus DOM and PS with additional programs.
Why it matters: Who pays for Mid-Atlantic peaks is a firmness problem as much as a growth problem. Derating speculative halls is forecast hygiene — not proof the long path will print 222 GW.
NIST’s public-facing AI documentation “zero draft” opens comments through 16 September.
What happened: NIST’s AI Standards Zero Drafts pilot has released an initial public draft of Guidance and Templates for Public-Facing AI Documentation — NIST AI 300-1 ipd (publication page dated 30 July 2026; Zero Drafts hub updated 14 August 2026). The draft targets model and dataset documentation for public consumption, including fields that should or could appear in a documentation artifact. It is not documentation of entire AI systems (left for later), not EU Article 50 labelling, and not draft high-risk classification. NIST will consider comments received by 16 September 2026 for a subsequent revision. The pilot’s stated aim is to broaden participation and accelerate voluntary consensus standards by handing thorough, stakeholder-driven “zero drafts” into private-sector standards developing organizations. “Shall” language in the draft is conformity criteria for a voluntary proposal — not U.S. statute. On the calendar only: this is day 28 of live EU transparency enforcement; today’s U.S. product is documentation process, not a restamp of Article 50 duties.
Why it matters: Public model/dataset cards are becoming a standards race. A comment window is not a transparency-outcome metric — but it is the U.S. process surface to watch beside Europe’s live labelling clock.
Adaptive platform trials for clinical AI: a master protocol when single-tool RCTs cannot keep up.
What happened: van de Sande, Bloemen, Sleijfer, Ikram, Hooft, Gommers, and van Genderen published Modernizing evidence generation for clinical artificial intelligence: the case for adaptive platform trials in npj Digital Medicine on 20 August 2026 (received 9 Jan 2026; accepted 11 Aug 2026; open access). The perspective argues that traditional diagnostic and prognostic trial designs cannot keep pace with rapidly evolving clinical AI — especially generative tools and continuously updating models. Adaptive platform trials are proposed as a scalable frame: multiple AI tools and updates tested head-to-head inside a master protocol that standardizes populations, endpoints, decision rules, and governance; arms can be added or removed dynamically; control groups can be reused; equity is named as a design focus. No patient n, AUC, mortality, or treatment-failure endpoint is locked on this page. This is methods architecture for continuous comparative evaluation — not a clinic-deployed outcome win and not a restamp of this week’s Kenya RCT, LiON, SHAKED, or WHO/Europe snapshots.
Why it matters: If models update weekly, evidence generation has to match the product cycle. Platform designs are the infrastructure bet; outcome claims still need results papers.
Egypt launches a national AI competency framework for teachers — among the first country adaptations.
What happened: Egypt’s Ministry of Education and Technical Education, with UNESCO Cairo, launched a national AI competency framework for teachers contextualized from UNESCO’s global teacher framework. The institutional recap positions Egypt among the first countries worldwide to develop and adopt such a national version, tailored to local priorities. The product aims to equip teachers with knowledge, skills, and ethical understanding to use AI responsibly in teaching and learning — framed as empowering teachers rather than replacing them. Development ran through the UNESCO–Huawei Technology-Enabled Open Schools for All project after a National Consultation and Validation Workshop in Cairo (October 2025) that brought together more than 120 representatives from government, universities, teacher-training bodies, civil society, development partners, and the private sector. Next steps named on the page include capacity-building programmes, AI master trainers, Arabic learning resources, and teacher-training hubs. This is a national policy-and-training instrument — not a multi-country learning-outcomes RCT, not yesterday’s AILit 4×19 school map, and not a restamp of the UNESCO Digital Learning Week ministerial advisory already covered earlier in the lookback window.
Why it matters: Global frameworks only bite when countries turn them into teacher training and curriculum language. Egypt’s launch is a concrete adoption signal, not an exam-score claim.
Exposure is multi-dimensional — language systems are not the whole story.
What happened: The same OECD paper stresses that AI exposure is not one-dimensional. Largest average domain gaps sit in social interaction, problem solving, and metacognition. Smaller average gaps appear in creativity, vision (July correction), and language. Least-exposed examples named in the summary include judges, chief executives, clinical/counselling psychologists, and several therapy occupations on social/problem-solving demand, and firefighters, surgical assistants, and certain specialist physicians on manipulation demand. Cost, organisational uptake, regulation, and social choice still sit between a low gap and a changed headcount.
Why it matters: A capability-gap table is not a layoff census. Read it beside observed reallocation and official projection tables — not as a substitute for either.
EIA still has no national data-center TWh census — only three voluntary pilots.
What happened: EIA’s 25 March 2026 press release launches three voluntary pilot field studies on data-center energy use: web surveys in Texas and Washington, and in-person interviews in Northern Virginia / Washington, DC. EIA identified 196 companies operating data centers across those regions; each is asked to report on at least one site. The questionnaire covers energy sources, electricity consumption, site characteristics, server metrics, and cooling. No national TWh total is locked. This is measurement-gap product design — not a load forecast and not a substitute for PJM’s peak-MW planning path.
Why it matters: Planning forecasts and metered inventories are different instruments. Until pilots scale, “data-center TWh” claims still outrun the official U.S. survey frame.
Zero drafts are proposals into SDO processes — not finished U.S. rules.
What happened: NIST’s hub frames Zero Drafts as thorough stakeholder-driven proposals meant to help standards developing organizations reach consensus faster while widening who gets heard. A companion TEVV zero-draft remains at outline stage on the same hub — no TEVV statistics locked here. Comment submissions become part of the public record. Do not collapse this voluntary documentation draft into live EU Article 50 notice duties or into Annex III high-risk timelines.
Why it matters: Right clock, right instrument. U.S. documentation templates and EU labelling obligations can move in parallel without being the same product.
Methods paper, not a results paper — keep the label on the page.
What happened: The npj Digital Medicine piece is a perspective on trial architecture for clinical AI. It does not report a multi-site patient-outcome trial of a named model. Head-to-head arms, shared controls, and governance rules are the proposal. Pair it with prior clinic results only when those papers are independently re-extracted — not as a merged “AI improved care” claim.
Why it matters: Health coverage fails when infrastructure arguments are sold as mortality wins. Platform design is necessary; it is not sufficient evidence of bedside benefit.
From global framework to national teacher training — implementation is the test.
What happened: Egypt’s recap ties the national framework to UNESCO’s human-centered line: AI should support teachers, not replace them. Named anchors include the Beijing Consensus, the Recommendation on the Ethics of AI, and the Santiago Consensus on Teachers. The page does not lock exam deltas, adoption shares, or classroom outcome percents. Competence language plus training hubs is the near-term deliverable.
Why it matters: Teacher frameworks only change practice when master trainers, materials, and hubs ship. Watch implementation, not the launch photo alone.
