Saturday, September 12, 2026 · Daily edition
Today’s ledger follows European microdata on how AI reshapes occupations without collapsing employment totals, Texas grid operators racing a December 10 data-center interconnection audit before Batch Zero can move, California’s new independent AI verification and auditor-registry laws, a methods review showing most medication-adherence AI models fail quality and bias checks, and Ghana’s TVET push to scale AI skills toward about 140 institutions and up to one million learners by end-2027.
Kiel Institute: AI transforms job profiles — not overall employment — in DK, PT, and SE
What happened. The Kiel Institute published a recap of Kiel Policy Brief 198, AI is transforming job profiles – not employment, using employer–employee data from Denmark, Portugal, and Sweden covering 2010–2023. Locked claims: overall AI progress over the last decade has had little to no impact on overall employment levels; exposure is not the same as jobs at risk. Language-modelling and reading-comprehension AI are associated with positive employment effects across occupations, especially production workers and managers, with hiring expanding and required skills rising. Translation and text-editing AI show negative effects for middle-skill (and to a lesser extent high-skill) administrative and office occupations — clerical, assistance, and call centers. Image-recognition AI is also negative for clerical and administrative roles. Most exposed occupations named: data analysts, software developers, and translators. Least exposed: construction workers and nursing staff. Görg’s locked line: higher AI exposure has no measurable effect on overall employment but systematically increases demand for higher qualifications.
What to watch. The print is reallocation inside the job mix — language AI lifting demand and skills, translation and image AI pressuring clerical profiles — without a measurable aggregate employment collapse. Keep this European microdata brief separate from yesterday’s ILO GenAI synthesis, Richmond Fed EB 26-27, Atlanta Fed WP 2026-4, Brookings Metro adaptive-capacity residual, Census CES packages, NY Fed firm-use shares, and any 2026 U.S. layoff census.
ERCOT targets December 10 for Texas’s data-center interconnection audit
What happened. Utility Dive reported that the Electric Reliability Council of Texas told the Public Utility Commission of Texas it intends to audit hundreds of data-center proposals by December 10 — a required step before the state’s Batch Zero large-load study can continue and before new large-load interconnections resume. The pause follows Gov. Greg Abbott’s August 3 call for a moratorium on new data-center interconnections until questions on energy, water, and public funds are answered. Locked figures: about 300 data centers of 75 MW or larger are navigating Batch Zero; ERCOT will also run a community-impact review on data centers and crypto facilities of 25 MW and above. Abbott’s letter cited interconnection-queue requests totaling about 474 GW, with approximately 90 percent data centers — more than five times ERCOT’s record peak demand. Medium-load snapshot: about 157 medium data-center and crypto facilities representing almost 8,800 MW. Separately, 17 large loads (mostly data centers) have cleared stability assessment toward energization later this year with total peak demand about 6.6 GW ramping over about five years — still subject to verification and community-impact RFIs.
What to watch. Texas is not just forecasting AI load — it is freezing new large-load gates until an audit can separate real projects from a multi-hundred-GW speculative queue. Keep this grid-operator audit and queue-governance story separate from U.S. national generation-record monthlies, Texas near-term load-growth forecast cuts, global electricity mid-year demand paths, and campus megawatt groundbreakings.
California signs SB 813 and AB 1405: independent AI verifiers plus a state auditor registry
What happened. On 9 September 2026, Governor Newsom signed a pair of AI-safeguard bills framed as first-in-the-nation third-party audit infrastructure. SB 813 (Sen. Jerry McNerney) creates a framework for independent verification organizations that can assess AI systems and models for compliance with state law. AB 1405 (Asm. Rebecca Bauer-Kahan) creates a state registry for AI auditors and sets standards for independence, transparency, and integrity — locked governor-page line: industry should not “grade its own homework.” Together the bills push independent third-party evaluation as AI embeds in critical sectors; the Governor also calls for federal national regulation. Background on the same page (not new signing-day duties): 2025 SB 53 frontier transparency, child AI-companion protections, ADMT privacy rules, 2026 procurement EO, Tech Fraud Task Force, and the August AI Cyber Defense Program.
What to watch. California is building who may audit AI — independent verifiers and a public auditor registry — a different compliance question from transparency marks already live under CAITA and EU Article 50. Keep this separate from live CAITA provenance/detection duties, Article 50 chatbot/mark/deepfake labels, AI Omnibus Annex III / Annex I clocks, GPAI Code Articles 53/55, and NIST AI 300-1 documentation comments due 16 September.
npj Digital Medicine: most medication-adherence AI models fail quality and bias checks
What happened. Cowdy, McPherson, da Silva, and Luo published AI models for medication adherence prediction: closing the gap to clinical readiness in npj Digital Medicine (published 2 September 2026; DOI 10.1038/s41746-026-03018-1). The systematic review covers 41 adherence-prediction modelling studies assessed with PROBAST + AI across participants/data, predictors, outcomes, and analyses. Locked findings: 71% of models showed great concern for development quality; 80% had high risk of bias in evaluation — commonly poorly defined adherence outcomes, inadequate missing-data handling, and limited validation. Reported discrimination did not consistently improve with more complex algorithms; methodological rigour, rather than model type, is the key barrier to translation.
What to watch. Before adherence models reach the bedside, the binding constraint is study design and bias control — not algorithm complexity. Keep this methods and bias review separate from patient-outcome RCTs, new AUROC claims, Lång’s patient-outcomes comment, the five-phase bedside ladder, voice-biomarker Perspectives, and yesterday’s WeatherNext Cyclones weather evaluation.
UNESCO: Ghana scales AI skills across TVET toward ~140 institutions and 1M learners
What happened. UNESCO reported that Ghana is scaling AI EmpowerED (UNESCO Global Skills Academy) with the Ghana TVET Service and GETFund, building on a 2025 AAMUSTED pilot. Locked program figures: more than 30 master trainers and around 3,300 learners have completed journeys and earned Microsoft certifications in AI and digital skills. Phase two targets around 140 TVET institutions across four consecutive three-month regional cycles, aiming to train and certify up to one million TVET educators and learners by end of 2027. Train-the-trainer path: master trainers complete an intensive bootcamp and Microsoft-accredited pathways, then share with colleagues and learners; ethical use is named on the page. Equity line locked: digital transformation must reach deprived and underserved communities where connectivity remains a challenge, including every TVET teacher regardless of location. Other scale-ups named: India, Kenya, Malaysia, Uganda, and the United Republic of Tanzania.
What to watch. An African TVET system is moving from a few thousand certified learners toward a multi-campus, million-learner certification aim while naming connectivity as the equity constraint. Keep this program self-report plus 2027 aim separate from UNESCO ICT Prize 2026 laureate reach, Estonia’s AI Leap expansion package, HEPI’s UK undergraduate survey, NYC K–8 bans, and multi-country learning-outcomes RCTs.
Also in today’s ledger
• NIST AI 300-1 documentation comments still due 16 September (4 days).
• EU Article 50 transparency duties remain live calendar context (day 41) — not a new lead.
• Texas Batch Zero timeline now tracks verification capacity, not a single AI electricity path number.