ECB staff on observed U.S. reallocation: high AI-substitution jobs fell while low-risk jobs grew.
What happened: Isabella Moder and Til Pommer, writing in the ECB Economic Bulletin Issue 4/2026, analyse U.S. employment growth by AI substitution risk using the Pizzinelli et al. (2023) index. Occupations are grouped low / medium / high. High-risk jobs (examples on the page: economists, graphic designers) saw average employment decline by more than 4% between 2019 and 2025. Low-risk jobs (examples: electricians, high school teachers) rose +13% over the same window. Composition shifted: low-risk share of U.S. employment 23% → 25%; high-risk 35% → 33%. A difference-in-differences design with three-digit NAICS sector fixed effects finds that, all else equal, high-risk jobs grew around 15 percentage points less than low-risk jobs over 2019–25, with the wedge widening after late-2022 ChatGPT. The same method on median hourly wages finds no significant wage-growth difference by substitution risk since 2019. Aggregate U.S. employment effects are described as muted / inconclusive so far — reallocation, not a national job-count collapse. This is observed 2019–25 outcomes on BLS employment with an IMF 2023 risk index — not yesterday’s new BLS projection cycle and not BLS’s official AI-exposure categories.
Why it matters: After a projection-heavy week, the measurable ledger is relative growth, not slogans that “half of all jobs” vanish. Keep the ~15 pp wedge separate from any long-horizon occupational table.
EIA’s near-term U.S. map: data-center load, gas stress, and an ERCOT price spike in the high-demand case.
What happened: EIA’s Today in Energy analysis dated 12 March 2026, completed with the February 2026 Short-Term Energy Outlook, treats faster data-center load as a U.S. grid and wholesale-price problem — not a global decade-ahead TWh path. U.S. electricity demand (net energy for load) grew about 1.7%/yr from 2020–25 versus 0.1%/yr in 2005–19; data centers are named as driving the growth, with industrial electrification as a co-driver. February STEO baseline: U.S. load +1.9% in 2026 and +2.5% in 2027, with the fastest regional averages in ERCOT ~10%/yr and PJM ~3%/yr over 2025–27. A high-demand scenario lifts 2026–27 growth rates 50% higher than baseline in data-center-heavy regions and +1 percentage point elsewhere, holding generating capacity fixed. Incremental power is mainly natural gas, with modeled delivered gas to generators about +$0.50/MMBtu. Baseline 2025–27 U.S. gas generation +1.7% / +29 BkWh; high-demand +123 BkWh (about +7 percent) (largest ERCOT gas increment +105 BkWh vs baseline +68). Baseline coal −9.3% / −68 BkWh; high-demand coal only −5.0% / −37 BkWh. Modeled 2027 wholesale prices: ERCOT +$37/MWh (79%) vs February STEO because the grid is isolated; ex-ERCOT major hubs about +$2.10/MWh vs a $48/MWh February STEO average; PJM +$2.60/MWh (4%).
Why it matters: This is a near-term STEO stress test with fixed capacity — not a prediction that ERCOT will print +$37/MWh, and not the IEA demand, supply, or liquid-cooling products from earlier this week.
Article 50 transparency guidelines: live chatbot notice, machine-readable marks, deepfake and biometric disclosure.
What happened: On day 27 of live EU AI Act transparency enforcement, the European Commission’s guidelines on transparency obligations under Article 50 are the product to read — not draft high-risk classification under Article 6. Article 50 applies from 2 August 2026. Providers must design systems so people are explicitly informed when they interact directly with an AI system, and must add machine-readable marks so AI-generated or manipulated content can be detected. Deployers must inform people when exposed to emotion recognition or biometric categorisation, deepfakes, or public-interest text published without human review or editorial control. The guidelines clarify provider vs deployer roles, define directly interactive systems / synthetic content / deepfakes / public-interest AI text, and give in/out-of-scope examples (including standard editing as an exception). Enforcement is named as national market-surveillance authorities, the AI Office for systems under its supervision, and the EDPS when EU institutions are providers or deployers. This is Commission interpretation of live duties — not a measured drop in misinformation volume and not Official Journal text.
Why it matters: Labeling and notice are the first public-facing clock of the Act. Product teams that still treat transparency as “coming later” are already late.
PULSE: longitudinal patient history reconstructs missing molecular profiles better than static multimodal baselines.
What happened: Wu et al. published Longitudinal alignments and syntheses of multimodal clinical data for personalized medicine with the PULSE framework in Nature Computational Science (research article; companion News & Views dated 26 August 2026). PULSE (Patient Unified Longitudinal Signal Engine) treats clinical modalities as temporally covarying readouts of an evolving patient state, carrying a visit-level Past-State into the next visit for cross-modal imputation. In UK Biobank benchmarking with about 10,000 participants who had paired baseline and follow-up visits across 61 routine lab tests and 251 metabolomics biomarkers (patient-level 70/30 split; test n = 3,000), PULSE reconstructed follow-up metabolomics from follow-up labs plus baseline multimodal context and outperformed adapted static multimodal baselines (MIDAS, scVAEIT, StabMap) on Pearson’s r, MAE, and MSE across the full 251-metabolite panel (Wilcoxon signed-rank, P < 0.001). For free cholesterol-to-total lipid ratio in large LDL particles, predicted vs measured longitudinal change reached Pearson’s r = 0.850 under full PULSE, ahead of the static frameworks. Ablations show removing historical-state input substantially hurts reconstruction. The paper also reports ICU 30-day readmission modeling across U.S. and Chinese cohorts and retinal–blood alignment for cross-body disease prediction — those endpoints are separate tasks, not a single patient-outcome RCT. This is research-benchmark evidence, not a clinic-deployed mortality win.
Why it matters: Sparse routine labs are the real-world constraint. Methods that use a patient’s own prior visits — not population averages alone — are the measurement layer to watch next.
Commission + OECD AILit: four dimensions and nineteen competences as a shared school reference.
What happened: A European Commission Digital Skills and Jobs Platform recap covers the 18 June launch of the AI Literacy (AILit) Framework for primary and secondary education, joint work by the European Commission and the OECD with Code.org/CodeAI experts, unveiled in Brussels at Collaborate for Impact. The page frames AI literacy as more than tool use: understand how AI works and evaluate its role. Named barriers include no shared definition, inconsistent implementation, misconceptions, and limited effective AI-supported pedagogy. Structure locked from the recap: four dimensions (engage with / create with / manage / shape AI) and nineteen competences organised as knowledge, skills, and attitudes, plus learner expectations, scenarios, and classroom examples. Audiences named: teachers, education leaders, policymakers, learning designers. Complements named on the page include PISA 2029 Media and AI Literacy (MAIL), updated ethical guidelines for educators, AI Act literacy/skills policy, and DigComp 3.0. An “Education Package” is “planned for later this year” with no date locked. This is an institutional recap of a competence map — not a multi-country learning-outcomes RCT and not a teacher-survey percent table.
Why it matters: Shared competence language is infrastructure for curriculum and procurement. It does not by itself prove better exams or narrower equity gaps.
Muted aggregates, real occupational wedge — and wages not moving yet.
What happened: The same ECB box stresses that firm-level AI adoption can raise employment in some settings while occupational subgroups still lose relative ground. Junior workers in highly exposed occupations are named as a negatively affected sub-group in the cited literature. The DiD wage series shows no significant high-vs-low substitution-risk difference on median hourly wage growth since 2019. Read that beside the employment wedge, not as proof wages are permanently insulated.
Why it matters: Reallocation without wage signal is still a labor-market fact. Coverage that only quotes national job totals misses the occupational split.
High-demand case holds capacity fixed — more gas, slower coal decline, regional price splits.
What happened: Inside the same EIA TIE, 2025 generation shares on the page are gas 40%, coal 17%, wind+solar 18%. In Mid-Atlantic (PJM), Midwest (MISO), and Southeast (SERC), coal supplies more than half of the additional generation in the high-demand case because spare coal capacity exists. New England/NYISO see about +$3.00/MWh (5%) in 2027; California/Southwest about +$1.30/MWh (4%). MISO/SPP can host high data-center demand while household growth stays low enough that regional averages stay below the U.S. mean; Southwest data-center growth continues, with most load after 2027.
Why it matters: Who pays depends on grid topology. ERCOT’s isolation is why the modeled price response is acute; interconnected eastern hubs absorb more.
Library landing + policy page: interpretation of live duties, not a new Act amendment.
What happened: The Commission library landing for the same guidelines package is the downloadable companion to the policy page. Compliance for marking/labelling may be shown via the Code of Practice on transparency of AI-generated content; non-adherents must use equivalently adequate means. Other Article 50 duties leave operators to choose adequate measures while taking the guidelines into account. Do not collapse this product into draft Article 6 high-risk classification routes from yesterday, or into Annex III application dates still scheduled later.
Why it matters: Right document, right clock. Live transparency notice is not high-risk conformity assessment.
Imputation fidelity is not bedside readiness — read the UKB task as research infrastructure.
What happened: The PULSE paper’s headline UKB task is reconstructing a 251-biomarker metabolomics panel from routine labs under patient-level holdout, with historical context from the prior visit. Top reconstructed features skew toward lipid-ratio markers with higher intra-individual stability. Removing the Past-State branch substantially reduces reconstruction performance. That is strong evidence for longitudinal modeling — and still one step removed from prospective clinical decision impact.
Why it matters: Health AI coverage fails when benchmark r values are sold as outcome wins. Keep the task label on the page.
What happened: AILit is framed as a common reference schools can adapt — not a single mandated curriculum. The recap does not lock national adoption shares or exam deltas. Pair it with the live Article 50 labeling clock outside the classroom: students will meet disclosed AI systems as citizens as well as learners.
Why it matters: Procurement and syllabus language need shared terms. Outcome evaluation is a separate measurement layer.
