Census CES-WP-26-25: national BTOS AI adoption is rising — firm-reported AI job cuts stay rare.
What happened: U.S. Census Bureau working paper CES-WP-26-25, The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks, reports results from the 2026 AI supplement to the Business Trends and Outlook Survey (BTOS) for the reference window November 2025–January 2026. Firm-function AI use reaches 18% of firms (32% employment-weighted), with expected use 22% within six months. Worker-task use is higher still at 23% of firms (41% employment-weighted), led by writing, document analysis, and information search; 65% of firms limit AI to three or fewer tasks. Among adopters, scope is still thin: 57% integrate AI in three or fewer business functions — Sales and Marketing 52%, Strategy and Business Development 45%, IT 41%. Very large firms in Information, Professional Services, and Finance show 50%–60% use (60%–70% employment-weighted). Most users (66%) rely on AI solely to augment tasks. AI-related employment decreases are rare: 2% of firms. A labor-outcome split matters: functional breadth and operational investment associate positively with employment decreases, while worker-task integration shows no significant link to headcount reduction once those controls are in. This is a national BTOS adoption-and-tasks working paper — not CES-WP-26-27’s early-career QWI channel, not PwC’s job-ad two-track barometer, not NY Fed’s 61%/51% regional-firm use shares, not SIEPR’s aggregate unemployment synthesis, and not a CES payroll layoff census.
Why it matters: National firm microdata now separate broad task diffusion from rare firm-reported AI headcount cuts — without collapsing into yesterday’s youth-hiring or job-ad packages.
NNSA picks Amentum to negotiate a Savannah River AI campus with dedicated on-site power.
What happened: The National Nuclear Security Administration announced selection of Amentum to enter negotiations for a phased lease for an AI data center plus dedicated on-site generation at the Savannah River Site in South Carolina. The project scale named on the page is a 1-gigawatt AI data center paired with approximately 2 gigawatts of on-site generation, with natural gas bridging to nuclear. DOE frames the design against a Ratepayer Protection Pledge: pair new AI load with dedicated on-site generation so electricity needs are not shifted onto existing utility customers. Context on the same page: an April 2025 DOE list of 16 federal sites, with Savannah River and three other locations advanced for private-sector development. The announcement states explicitly that selection for negotiations is not a final lease award; any agreement still needs negotiations, permitting, safety and security evaluation, and other federal approvals. No gallons, acres, or operating-load inventory appear on the page. This is a named federal campus + dedicated-generation negotiation — not DOE Paducah’s 1.8 / 2 / 2.6 GW campus plan, not Virginia commercial-sales or Dominion peak prints, not FERC large-load tariff show-cause process, not global electricity-path forecasts, and not a national data-center electricity census.
Why it matters: A second federal-land AI campus with dedicated generation is a ratepayer-isolation design claim — still a negotiation, not operating megawatts.
NIST IR 8615: next-gen secure hardware standards put AI hardware on the priority list.
What happened: NIST released IR 8615, Workshop Report on Rolling Next-Generation Secure Hardware into Standards, on 1 September 2026 (workshop held 26 January 2026). The report organizes five priority areas: unified standards and governance including AI hardware (with chiplets and post-quantum cryptography); provenance and traceability via cryptographic identities, SBOMs, and attestation; supply-chain security and procurement incentives; scalable verification including AI-assisted analysis; and workforce development. This is a hardware-security workshop report feeding standards processes — not a model-card mandate, not a transparency-labeling statute, and not a finished binding U.S. rule. Keep it separate from NIST AI 300-1 public-facing documentation templates (comments still due 16 September). This is a secure-hardware / AI-hardware standards workshop report — not EO 14409’s voluntary frontier-model access frame, not live EU Article 50 chatbot/mark/deepfake duties, not FTC’s proposed Section 5 accuracy statement, and not California’s live transparency act packaging.
Why it matters: The standards fight is moving into chips, provenance, and verifiable supply chains — not only into chatbot labels and model cards.
npj Digital Medicine: a five-phase ladder before medical AI claims bedside proof.
What happened:npj Digital Medicine published A five-phase evaluation framework for diagnostic and predictive medical artificial intelligence (DOI 10.1038/s41746-026-03155-7). Phase 1 calls for multi-center retrospective external validation on a frozen model — typically at least three centers — using STARD-AI for standalone diagnostic accuracy and TRIPOD+AI for prediction models, without post-hoc threshold tuning. Phase 2 is shadow-mode / silent trial on live hospital streams before outputs reach clinicians, tracking failure rate, latency, throughput, and calibration drift. Phase 3 is controlled human–AI interaction under DECIDE-AI (and STARD-AI if the endpoint is team diagnostic accuracy), including standalone benchmarking against seniority tiers and exploratory paired designs. Later clinical-evidence and post-market phases are part of the ladder but were not fully extracted in this pass — no invented Phase 4–5 sample sizes. This is a methods and reporting framework — not a patient-outcome RCT, not MoChiAgent’s obstetric AUROCs, not ECG-CLIP, not Retina4IRD’s genotype accuracy trial, not breast-triage workload/CDR results, and not LungIMPACT’s null pathway study.
Why it matters: Strong retrospective AUROCs are only the first rung; shadow-mode and human–AI interaction sit between paper performance and bedside claims.
Nature / HEPI: ~94% of surveyed UK undergrads used genAI on assessed work — 12% inserted AI text.
What happened: A Nature careers feature on how professors are redesigning assessment under generative AI cites the Higher Education Policy Institute (HEPI) 2026 survey of 1,054 UK undergraduates: roughly 94% used generative AI to help with assessed work, and 12% directly inserted AI-generated text into coursework. A separate May study from survey data of more than 95,000 students at 20 U.S. universities estimated 9% used AI on coursework despite knowing it broke the rules. Practice examples named in the piece include critique-the-model assignments, in-class writing and orals, revision-history submission, and pass/fail take-homes paired with offline exams — no single institutional policy is crowned “correct.” This is a student-use survey package inside journalism — not a learning-outcomes RCT, not UNESCO’s U18 China MIL event, not the ICT in Education Prize ceremony on 9 September, not Egypt’s teacher competency framework, not NYC’s K–8 student ban, and not a global student census.
Why it matters: Near-universal assessed-work use forces assessment redesign now — one day before UNESCO Digital Learning Week and two days before the ICT Prize ceremony.
Augment-first diffusion — do not convert 2% of firms into a national layoff rate.
What happened: Keep the BTOS split labeled: 66% augment-only among users; AI-related employment decreases in 2% of firms; functional investment associates with decreases while task use does not once controls are in. Distinct paper from CES-WP-26-27’s early-career hiring channel and from SIEPR’s unemployment synthesis.
Why it matters: Task diffusion can look broad while firm-reported AI separations stay rare — and the regression split warns against flattening those channels.
Negotiations, not a lease — Savannah River’s 1 GW / ~2 GW pair is design intent.
What happened: Keep Amentum’s selection labeled as entry into negotiations under the Ratepayer Protection Pledge framing. Do not mash 1 GW / ~2 GW with Paducah’s 1.8 / 2 / 2.6 GW file, Virginia commercial MWh, or global electricity-path forecasts. No acres or gallons on the NNSA page.
Why it matters: Dedicated on-site generation is the policy claim; operating load and signed lease status are still unproven.
Hardware standards vs documentation templates — IR 8615 is not AI 300-1.
What happened: Pair IR 8615’s AI-hardware / provenance / AI-assisted verification priorities with the still-open AI 300-1 comment clock (16 September) without collapsing them. Workshop consensus is not a binding standard and not Article 50 transparency law.
Why it matters: Readers confuse chip-security standards work with either model-card templates or live transparency duties.
Standing clock: WHO/ITU/WIPO GI-AI4H meets in Hangzhou 16–18 September.
What happened: The third Global Initiative on AI for Health meeting (Hangzhou, 16–18 September 2026) emphasizes practical implementation and a privacy-preserving Benchmarking Challenge on correctness, non-disclosure, and cost without exchanging data or weights. Results are not regulatory approval. Keep distinct from the five-phase methods paper and from WHO/Europe’s governance dialogue.
Why it matters: Benchmarking without weight-sharing is a governance design choice nine days out — not a bedside outcome claim.
Two surveys, two claims — UK 94%/12% is not the U.S. 9% “knew it broke rules” estimate.
What happened: HEPI’s 1,054 UK undergraduates (~94% / 12%) and the U.S. >95,000 / 20-university 9% figure answer different questions. Standing calendars: UNESCO Digital Learning Week 8–11 September; ICT Prize ceremony 9 September — not restamped as today’s lead.
Why it matters: Assessment redesign and knowing-rule-breaking prevalence are related but not interchangeable stats.
