BLS’s new 2025–35 map: slower total job growth, AI-linked power and compute hiring, and office support as the steepest decline.
What happened: The Bureau of Labor Statistics employment-projections news release for 2025–35 (USDL-26-1422, for release 10:00 a.m. ET Thursday, 27 August 2026) is a new projection cycle — not the prior-cycle TED/MLR package. Total employment is projected to rise from about 170 million to about 176 million — +3.5%, or about 5.9 million jobs — slower than the prior decade’s double-digit gain. Utilities is the fastest major sector (+9.8% / +58,800), with nearly all of that gain in electric power generation, transmission, and distribution, including AI power demands. Private healthcare and social assistance adds the most jobs (+9.5% / more than 2 million, about 37% of all new jobs). Professional, scientific, and technical services grow +8.6% / +926,700 on demand for AI-based systems, R&D, and consulting. Computing infrastructure / data processing / web hosting grows +25.1% / +120,400 on accelerated AI adoption. Computer and mathematical occupations are the fifth-fastest major group (+7.3%), with data scientists +34.6% and computer and information research scientists among the fastest detailed roles. Office and administrative support is the fastest-declining major group (−4.0% / −752,100), with AI and automation tools named among demand dampeners; sales and related occupations also edge lower, with e-commerce and AI tools named in the sales process. These are long-horizon projections with embedded judgment, not a 2026 layoff census.
Why it matters: The measurable record is whether hiring and openings track the power/compute expansion and the office-support decline — not slogans that “half of all jobs” vanish. Readers should not mash these 2025–35 levels with the prior-cycle table.
BLS’s first official AI-exposure categories: relative ranks, not a job-loss forecast.
What happened: Alongside the 2025–35 projections, BLS published a new supplement that sorts detailed occupations into four relative AI-exposure categories — Low, Moderate, High, and Very high — for the 831 occupations in the projections matrix. The product combines five external sources: three theoretical measures (Felten et al.; Eloundou et al.; Eisfeldt et al.) and two observed-usage measures (Anthropic Claude traffic; Microsoft Copilot applicability). Exposure means AI could assist or complete some of the work, or that observed AI interactions map onto occupational tasks. The agency is explicit: exposure does not imply job loss, productivity gains, automation probability, wage effects, or replacement. “High/Very high” is not a decline forecast; “Low” is not a safety guarantee. Categories are relative, do not split automation from augmentation, and the theoretical sources conceptualize AI capabilities no later than mid-2023. An accompanying Excel table is the occupation-level product.
Why it matters: Career planners finally have an official relative ranking — and a hard disclaimer against treating the ranking as a displacement model. The measurable follow-through is whether later BLS products connect exposure ranks to actual employment change.
IEA 4E maps liquid cooling’s efficiency potential — and why PUE understates the gain.
What happened: An IEA 4E EDNA publication dated 22 June 2026, Liquid Cooling in Data Centres, maps liquid-cooling technologies and the barriers to wider use. The HTML summary puts energy-savings potential around ~8% at servers, 30–40% at facility level, and overall on the order of 10–21%. Current use is low because of missing standardisation, high upfront cost, and long-term reliability concerns. A further barrier: Power Usage Effectiveness (PUE) systematically understates efficiency gains from liquid cooling. The report argues for policy intervention and for retrofit solutions suited to existing multistorey halls. This is a 4E/EDNA efficiency product — not the IEA Energy and AI demand path, not Key Questions, and not a global TWh forecast.
Why it matters: After a week of demand-path headlines, the efficiency layer is the missing half of the ledger. Potential is not booked savings; the measurable record is deployed liquid-cooled capacity, better metrics than PUE alone, and whether retrofits reach legacy sites.
Commission draft high-risk classification guidelines: two Article 6 routes, practical examples, still non-binding.
What happened: On day 26 of live AI Act transparency enforcement, the European Commission’s draft guidelines on classifying high-risk AI systems under Article 6 are the fresh product — not the enforcement-architecture map. The draft is meant to help providers, deployers, and market-surveillance authorities decide whether a system is high-risk and to support uniform application of Article 6. Two routes are locked on the page: Article 6(1) — the system is a safety component of (or is itself) a product covered by Annex I harmonisation law and required third-party conformity assessment; and Article 6(2) — Annex III use cases. The guidelines include practical examples of systems that should or should not be high-risk; the examples are not exhaustive and may be updated. This is Commission interpretation in draft form, not adopted final guidelines and not a measured change in deepfake volume or public trust. A targeted consultation ran earlier; final text is still pending.
Why it matters: Classification is the gate before most high-risk duties bite. Getting Annex I product routes confused with Annex III use-case routes is how compliance calendars go wrong.
WHO/Europe’s first EU-27 AI-in-health snapshot: 74% report diagnostic AI, 63% use patient chatbots.
What happened: WHO/Europe released a first snapshot of AI in health care across the 27 EU Member States, based on data collected June 2024 to March 2025 under a multi-year European Commission funding agreement and zooming in from a late-2025 regional report. All 27 countries name improved patient care as a driver; a majority already deploy AI in clinical settings. 74% report AI-assisted diagnostics (imaging, disease detection, clinical decision-making). 63% use chatbots for patient engagement. Nearly half have dedicated AI/data-science roles in health. 81% involve stakeholders in AI governance — above the broader WHO European Region average. Priorities named include workforce readiness (literacy, ethics, data governance), inclusive public and patient engagement, and centres of excellence for testing and standards. The release warns that systems built without meaningful public input may face rejection and could exacerbate inequities. This is country-reported adoption, not a patient-outcome RCT and not a 2026 clinical-performance paper.
Why it matters: Adoption shares are infrastructure facts. They do not prove better mortality, safety, or equity — and the public-engagement warning is part of the same record.
OECD’s Schleicher: general-purpose GenAI can polish homework and still hurt exam learning.
What happened: Andreas Schleicher, OECD Director for Education and Skills, published an Education and Skills Today commentary on using generative AI in education, drawing on the OECD Digital Education Outlook 2026. Unlike earlier edtech waves, much GenAI is free, intuitive, and requires no coding. Tools can support learning when guided by clear teaching goals or designed for education. When AI removes productive struggle, students may finish faster with better immediate outputs while consolidating less — risking “metacognitive laziness.” Studies cited on the page find students using general-purpose GenAI improved response quality versus no-access peers, but exam performance did not reflect that improvement and got worse. Specialised, purpose-built pedagogical tools “show more promise.” Early trials: GenAI tutoring assistants can raise less-experienced tutors’ strategy quality and math mastery; a simulated-student teacher-training chat improved preparedness and confidence. The page says future research is still needed. This is director commentary, not the full Outlook tables and not a multi-country learning-outcomes RCT.
Why it matters: Assignment polish is not learning. The measurable classroom record is exam and mastery evidence under tools built for pedagogy — not chatbots treated as homework machines.
Where 2025–35 still adds heads: healthcare volume, AI systems work, and data-centre-adjacent compute.
What happened: Inside the same USDL-26-1422 release, healthcare support (+13.3%) and healthcare practitioners/technical (+8.0%) together account for almost one-third of new jobs; nurse practitioners lead detailed occupations at +41.0%. AI electricity demand helps put four of the ten fastest-growing detailed industries in the power stack, while computing infrastructure posts mid-20s percent growth. Read that expansion beside office/admin’s −752,100 projected loss. Still a projection, not a metro-by-metro hiring guarantee.
Why it matters: Official tables have two sides. Coverage that only quotes the decline column misstates the BLS package.
BLS ties AI electricity demand to utilities hiring — an employment layer on the power story.
What happened: The same 2025–35 news release names rising electricity demand “in large part” from AI adoption and related infrastructure such as data centres as a driver of utilities growth and of several fastest-growing detailed power-generation industries. That is a U.S. employment framing, not an IEA generation-mix path and not a substitute for the liquid-cooling efficiency numbers above.
Why it matters: Power demand shows up in job tables as well as TWh tables. Keep the products labeled separately.
Draft ≠ live high-risk duties. Classification guidance is the bridge product.
What happened: The Commission library page frames the guidelines as interpretation and examples under Article 6, downloadable in sections (general principles; Annex I; Annex III). Targeted consultation feedback windows closed earlier; the text remains draft. Do not collapse this product into chatbot-labeling duties already live since 2 August, or into Annex III high-risk application dates still scheduled later.
Why it matters: Product teams need the right document for the right clock. Draft classification examples are not final obligations.
Adoption is not outcomes. WHO’s EU-27 shares stop at country reports.
What happened: Read the 74% diagnostics / 63% chatbot / 81% governance figures as self-reported country deployment and process markers from a mid-decade survey window. The release’s own priorities — literacy, ethics, data governance, public engagement — are the next measurement layer. Do not treat the snapshot as proof that diagnostic AI improved mortality or that chatbots improved access equity.
Why it matters: Health AI coverage fails when adoption percentages are sold as clinical wins.
Purpose-built tutoring tools beat generic chatbots in the early trials Schleicher cites.
What happened: The OECD commentary contrasts general-purpose GenAI (assignment quality up, exams worse in cited studies) with specialised pedagogical tools and tutoring assistants that showed better strategy use, math mastery gains for assisted novice tutors, and higher teacher-prep confidence in a simulated-student chat. Those are early trials with an explicit “more research needed” caveat — not a national policy evaluation.
Why it matters: Procurement language should prefer tools designed for learning over free general chatbots dropped into homework.
