St. Louis Fed: generative AI use is widespread across occupations — and still shallow on most tasks.
What happened: The Federal Reserve Bank of St. Louis published What Work Does Generative AI Do? in its On the Economy series (authors Bick, Blandin, Deming, and Schumacher), summarizing a 2026 working paper (Fed in Print item 103693). The piece uses the Real-Time Population Survey (RPS), a nationally representative sample of U.S. adults 18–64. Headline adoption locked: generative-AI use among adults rose from 45% to 62%, and on-the-job worker use rose from 33% to 45%, from August 2024 to May 2026. New occupation- and task-level indexes come from nearly 14,000 workers across four quarterly waves August 2025–May 2026; respondents see the 10 most important O*NET tasks in their occupation. Widespread-but-shallow pattern locked: in more than 80% of occupations, at least 1 in 5 workers uses AI on the job; more than 40% of tasks have adoption above 20%. Only 40% of occupations have adoption above 50%; 16% exceed 70%. On tasks, fewer than 3% have adoption above 50%, and none exceed 70%. Highest occupations locked: computer and information research scientists 87.3%; information security analysts 85.4%; network and computer systems administrators 82.4%; computer programmers, PR specialists, personal financial advisors, and chief executives near or above 80%. Highest tasks locked: reading documents for technical information 61.3%; preparing research reports 60.7%; analyzing data for trends 57.5%. Lowest occupations locked: animal caretakers 5.3%; receptionists and information clerks 7.6%; LPNs/LVNs 10.4%. Exposure scores explain roughly half of occupation/task adoption variation but miss cases — medical secretaries/admin assistants 16.8% versus a common exposure prediction near 61%; computer/office-machine repairers 75.7%, special-education teachers 70.9%, and laundry/dry-cleaning workers 49.0% run about double common predictions. Demographics explain little; longer familiarity correlates with deeper task use. Explicit frame on the page: employment and wage effects depend on what work AI actually does, not only how many people use it. This is a regional Fed survey blog plus working-paper summary of nationally representative occupation/task adoption — not yesterday’s Anthropic Claude conversation-sample productivity note (80% / conditional 1.8 percent), not the Dallas Fed Texas Lightcast/Claude-usage posting note already covered on 2 September, not Chicago Fed AI-applicability working paper, not Kiel profiles-not-headcount, not ILO limited-displacement synthesis, not Richmond Fed EB 26-27, not Atlanta Fed WP 2026-4, not Census CES-WP-26-25/27, not NY Fed firm-use shares 61%/51%, and not a 2026 layoff census.
Why it matters: The near-term jobs print is diffusion depth — AI already shows up in most occupations and almost no majority-use tasks — not headcount destruction and not vendor task-time savings.
IEA Electricity 2026 Grids: more than 2,500 GW stalled in connection queues — data centres arrive in 1–3 years; wires take 5–15.
What happened: The International Energy Agency’s Electricity 2026Grids chapter frames capacity, congestion, and interconnection — not a 2030 data-centre TWh census — as the bottleneck for new generation, storage, and demand. Locked queue print: over 2,500 GW of renewable, large-load, and storage projects are stalled in grid connection queues worldwide. Timing mismatch locked: planning, permitting, and completing new grid infrastructure typically takes 5 to 15 years; data centres 1–3 years; solar/wind 1–5 years; EV charging 1–2 years. Investment path locked: meeting demand through 2030 requires annual grid investment to rise by about 50% by 2030 from today’s about USD 400 billion, plus supply-chain and workforce scale-up; key grid-component prices have nearly doubled over the past five years. Near-term hosting capacity locked: complementary grid-enhancing technologies (GETs) and regulatory adjustments could free enough capacity to connect 1,200–1,600 GW of advanced-stage queued projects — about 750–900 GW via conditional non-firm connection agreements, with the remainder via dynamic line rating, advanced power-flow control, reconductoring, voltage uprating, and related options. GET rollout alone (holding other factors fixed): 450–700 GW. Benefits cannot be stacked additively — several solutions address the same thermal, voltage, or congestion constraint. Non-firm connection is faster access with the condition that output or consumption may be limited at certain times. This is an IEA medium-term grids / queue / hosting-capacity chapter — not LBNL’s national-lab decade-ahead TWh path / double-digit national electricity share, not IEA Demand’s U.S. data-centre growth-share print, notEnergy and AI415→945 bounds, not IEA mid-year global demand-growth update, not EIA September STEO generation-record monthly, not ERCOT 74.5 GW or Batch Zero audit timeline, not yesterday’s DOE draft National Transmission Needs Study congestion in 5% of hours, and not a metered 2026 campus kWh census. The 2,500 GW figure mixes renewables, large loads, and storage — not a data-centre-only queue.
Why it matters: After a summer of TWh paths and a U.S. congestion-hours draft, the IEA names the missing global layer: AI and other large loads collide with multi-year wire timelines and record connection queues.
FTC extends personalized-pricing comment period to 25 September — personal data used to set willingness-to-spend prices.
What happened: The U.S. Federal Trade Commission published a 3 September 2026 press release extending public comment on a proposed policy statement regarding personalized pricing. Comment period extended by seven days; new deadline 25 September 2026 (was 18 September 2026). Locked definition on the release: personalized pricing is the use of personal data to set prices according to the amount a company believes an individual consumer is willing to spend. Original invitation dated 19 August 2026. Electronic comments via regulations.gov docket FTC-2026-1057. This is a comment-clock extension on a proposed enforcement policy statement — not a final rule, not a measured drop in misinformation, not the FTC’s accuracy/objectivity policy statement already covered on 4 September, not NIST NCCoE agent-identity work covered yesterday, not NIST AI 300-1 documentation comments (still due 16 September, 1 day), not NIST SP 800-239, not FDA’s GenAI-device discussion paper, not California SB 813 / AB 1405, not live EU Article 50 chatbot/mark duties (day 44 calendar only), and not an AI model-card statute. The linked PDF did not extract as full text for this edition — lock the news-release page only. Do not invent a ban on personalized pricing from the extension alone.
Why it matters: The policy beat is pricing trust and personal-data willingness-to-pay — a consumer-protection clock sitting beside live transparency duties, not another model-documentation template.
Nature Reviews Bioengineering: a four-stage framework for AI-enabled clinical trials — not a patient-outcome RCT.
What happened:Nature Reviews Bioengineering published AI-enabled clinical trials (DOI 10.1038/s44222-026-00487-7; online 10 September 2026). The Review proposes a four-stage AI-enabled trial-engineering framework: assemble and harmonize multimodal data and validate tools; match fit-for-purpose tools to the question and outcome; AI-supported conduct (patient-to-trial matching, surrogate end points, externally matched comparator arms, digital twins, automated collection/curation); and faster go/no-go that flags non-promising drugs early. Three principles run throughout: fit-for-purpose validation, continuous regulatory engagement, and human oversight. Locked trial-system context on the page (not an AI treatment effect): about 90% of drug candidates entering trials fail to reach approval; about 30,000 RCTs are published per year at roughly US$300 billion; less than half of trials meet pre-specified enrolment at completion. Named case illustrations on the page include iBox surrogate (transplantation); annualized relapse rate (multiple sclerosis); heart digital twins for ventricular tachycardia; AIM-MASH AI-assisted histology (hepatology); and an LLM for patient-to-trial matching. The Review maps EMA, FDA, and the EU AI Act as the evolving regulatory landscape — not a new FDA guidance. Explicit limits: tools must be demonstrated fit-for-purpose; shortening follow-up via surrogates needs post-trial surveillance. Ambition is faster identification of effective therapies and earlier elimination of futile or harmful ones — not a claimed mortality reduction from this Review. This is a methods/framework Review — not yesterday’s Nature Medicine I3LUNG retrospective (2,396 patients; TEST AUC up to 0.77), not LungIMPACT’s pathway-timing null, not breast-triage workload/CDR numbers, and not a device clearance.
Why it matters: The clinical print is trial-system engineering — how AI might change design, matching, surrogates, and go/no-go — with honest system failure rates, not a bedside outcome win.
World Bank EdTech Policy Academy: Seoul convening on AI education policy to implementation at scale.
What happened: The World Bank published the event page for the Global EdTech Policy Academy on AI, held in Seoul 31 August–4 September 2026 under the theme “AI and the Future of Education & Skills Development: From Policy to Implementation at Scale.” Convened by the World Bank Education and Skills Global Practice. Partners locked: Government of the Republic of Korea; UK FCDO; Mastercard Foundation. Audience locked: policymakers, World Bank teams, development partners, and industry — invitation only, intended for country teams enrolled in the AI in Education Policy Academy. Design locked: project-based mix of online prep, in-person convening, peer learning, field visits, and follow-up. Aims locked: AI strategies and policies in education; where AI adds value and what safeguards are needed; partnership models for scaling; country action plans / implementation blueprints. The academy also examines how AI is reshaping education, skills, and jobs. This is a completed invitation-only policy academy / convening — not yesterday’s UNESCO AI4EAC student challenge (~1,000 / 57 / 8), not UNESCO LAC Observatory, not Ghana TVET’s million-learner scale-up aim, not ICT Prize 2026 laureates, not HEPI’s UK undergraduate survey, not WDR 2026’s labor-split package, and not a multi-country learning-outcomes RCT. No enrolment, learning-gain, or safeguard census is locked on the page. Standing context only: Digital Learning Week ended 11 September; UNESCO global education-AI consultation comments still due 15 October. OECD PISA 2025 pages returned access errors this pass and are not used.
Why it matters: The education beat is government implementation capacity and safeguards for AI in schooling systems — a finished academy design, not proven learning gains.
Widespread adoption is not deep task use — and not a layoff print.
What happened: Keep St. Louis Fed RPS locked as adults 45% → 62% and workers 33% → 45% (Aug 2024–May 2026); nearly 14,000 workers across four waves; widespread >80% of occupations / >40% of tasks at ≥20% use; shallow 40% of occupations / <3% of tasks at ≥50%; named occupation/task highs and lows; medical-secretary 16.8% vs ~61% exposure miss. Keep separate from yesterday’s Anthropic conversation productivity note and from Dallas Fed Lightcast postings already used on 2 September.
Why it matters: Nationally representative occupation-and-task adoption is a different evidence class than vendor task-time savings, firm surveys, or online-ad declines.
What happened: Lock IEA Grids >2,500 GW stalled; data centres 1–3 years vs grids 5–15; investment +~50% from USD 400 billion by 2030; component prices ~2× in five years; unlock 1,200–1,600 GW (750–900 non-firm; GET-only 450–700); non-additive caveat. Keep separate from LBNL national-lab TWh path / double-digit national share, IEA Demand growth-share, ERCOT megawatt prints, and yesterday’s DOE Needs Study 5% of hours.
Why it matters: Grid policy for AI load is interconnection timing and hosting capacity — not only national electricity-share forecasts.
What happened: Lock FTC personalized-pricing comments due 25 September 2026 (was 18 Sep; +7 days); willingness-to-spend definition; proposed statement; docket FTC-2026-1057. Day-44 Article 50 and NIST AI 300-1 comments due 16 September remain standing calendar only — not re-fetched leads.
A finished policy academy is not a learning-outcomes RCT.
What happened: Lock Seoul 31 Aug–4 Sep 2026; Korea / UK FCDO / Mastercard Foundation partners; invitation-only policy-to-implementation design; safeguards and country action plans. Standing context only: Digital Learning Week ended 11 September; UNESCO consultation comments due 15 October.
Why it matters: Implementation capacity for education AI is the beat — measured learning or employment gains remain unclaimed.
