Sunday, September 27, 2026 · Daily edition
G20 water reuse for data centers, LLM exposure in job ads, and FDA’s 1,600 AI devices
Today’s ledger moves from facility water inventories and firm interview packages to diplomatic water-reuse rules that name data centers, how LLM exposure shows up in U.S. job ads, a NIST manual for planning AI evaluations, the size of the U.S. authorized AI-device market, and a free multilingual teacher course on ethical AI use.
EPA: G20 members back a fit-for-purpose water-reuse deal that names data centers
What happened. An EPA news release from Houston, dated 15 September 2026, says Administrator Zeldin launched a G20 Water Reuse Initiative during the Energy Abundance Ministerial Session. Every G20 member assembled there agreed to the Water Reuse Outcome Document. The frame is fit-for-purpose reuse: treat water to the quality its next use actually needs, rather than one universal standard. The goal is less pressure on freshwater, better drought resilience, and more reliable critical infrastructure. The sectors named include energy, manufacturing, agriculture, municipal systems, and data centers. The package also points to technology-neutral standards, streamlined permitting, health and environment safeguards, voluntary information sharing and pilots, and capacity building for developing countries. EPA’s Water Reuse Action Plan 2.0, launched 16 April 2026, sits in the background. No gallon totals or national electricity path are locked here. This is diplomatic reuse governance, not another facility withdrawal inventory.
What to watch. After facility gallons and ISO megawatt clocks, the environment beat is whether data centers get treated as a named reuse sector in international water rules — architecture, not a new withdrawal range.
Read the EPA G20 water-reuse release →
Cleveland Fed: higher LLM exposure means more AI language in U.S. job ads
What happened. Cleveland Fed working paper WP 26-24 by Kevin Rinz tracks AI mentions in U.S. job advertisements. Mentions have risen substantially since early 2024, especially in occupations more exposed to large-language-model capabilities. One extra standard deviation of that exposure is associated with a 3.1 percentage-point higher AI-mention rate in job ads. Relative to less-exposed work, more-exposed occupations see postings stabilize after a period of relative decline, posted wages rise, and both hires and separations rise. New-hire wages move up with posted wages, while wages for continuously employed workers have declined — possibly because those labor markets became less tight. The paper is a working abstract, not a national layoff census, and it reflects the author’s views rather than the Federal Reserve System.
What to watch. After firm interview packages on hybrid AI workflows, the jobs beat is demand composition in U.S. postings: more AI language where models bite hardest, plus churn and wage splits — not a pink-slip count.
Read Cleveland Fed WP 26-24 →
NIST AI 200-3: plan ARIA evaluations with model tests, red teams, and user tests
What happened. NIST published Trustworthy and Responsible AI report AI 200-3, the ARIA Evaluation Planning Manual, dated 18 September 2026. ARIA stands for Assessing Risks and Impacts of AI. The manual treats a full evaluation as three pieces working together: model testing, red teaming, and user testing. It is a first-step planning guide — it lists the elements people need and points to resources — not a finished accuracy scoreboard or a binding directive. Authors include Jensen, Przybocki, Kodwani, Greene, Hall, Amironesei, and Greenberg. Standing comment clocks on other NIST drafts still run in the background; this package is the planning layer itself.
What to watch. After EU Board meeting recaps and draft-comment calendars, the policy beat is how NIST wants AI risk evaluations designed before anyone claims a score.
Read the NIST ARIA Evaluation Planning Manual →
FDA: more than 1,600 AI-enabled medical devices authorized as of September 2026
What happened. FDA’s Center for Devices and Radiological Health says it has authorized more than 1,600 AI-enabled medical devices for marketing in the United States as of September 2026, with a searchable public list updated over time. Page examples include skin-cancer imaging tools, deep-learning image sharpening, diabetic-retinopathy detection from retinal photos, sensors that estimate heart-attack probability, and algorithms that help automate insulin dosing from continuous glucose monitors. FDA is clear that it does not regulate “AI as such”; it regulates medical devices, including through predetermined change-control plans. This is a marketing-authorization count, not a clinical-outcomes census and not the separate generative-AI device discussion still open for comment.
What to watch. After multi-agent wet-lab headlines, the health beat is the size of the cleared U.S. AI-device market — clearance volume, not a validated therapeutic claim.
Read the FDA AI-enabled medical devices page →
UNESCO MOOC: Educating in the age of AI — 7 modules, 49 lessons, three languages
What happened. UNESCO Open Learning lists the instructor-led course Educating in the age of artificial intelligence: Digital citizenship at school (ED-004). The package has 7 modules, 49 lessons, and 5 assessments, with a certificate path. Pages run in English, Spanish, and Portuguese. The focus is ethical AI use, disinformation, digital coexistence, rights protection, democratic participation, and critical thinking for teachers, school teams, and social actors, with a Latin American context and Microsoft support. This is a course catalog entry, not an uptake census or a learning-outcomes trial.
What to watch. After system-level maps of AI education strategies, the education beat is free classroom capacity building: a multilingual teacher course, not another policy-coverage percentage.
Open the UNESCO ED-004 course page →
Also in today’s ledger
• Diplomatic reuse frameworks that name data centers are a different evidence class from facility gallon inventories and ISO interconnection clocks.
• Job-ad exposure elasticities are not firm interview packages — and an ARIA planning manual is not a Board-meeting recap.
• Marketing-authorization counts are not unreviewed wet-lab hunts — and a teacher MOOC is not a system AI-strategy coverage map.
|