Bill Gates wants “human-reserved” jobs — and a tax that stops treating robots as a write-off.
What happened: In a roughly 6,000-word Gates Notes essay published Wednesday, Bill Gates called for a “human reserved” domain of work that machines should not take, comparing it to nature reserves: places we could pave, but choose not to. “Many jobs will disappear forever,” he wrote. He used the care his father received while dying of Alzheimer’s as the example of labor that “no robot could or should have done,” and said a robot should not deliver an incurable-disease diagnosis even if it technically could. TechCrunch’s write-up of the same essay flags a “robot tax”: payroll taxes apply when you hire a person, while buying a robot is typically written off as a business expense — a nudge toward replacement. Gates said governments are unprepared, analogizing the needed reorganization to post-9/11 national-security reform, and that his staff are working to set up a November meeting with Xi Jinping on shared AI risk. On schooling, he cited a preliminary survey linking heavier AI use to less critical thinking, with a stronger association for young people. This is an essay and a policy pitch, not a new labor-market dataset.
Why it matters: The measurable record is whether any government actually walls off care, teaching, or other roles from automation, whether payroll-vs-capex tax rules change, and whether youth critical-thinking studies replicate — not whether a founder’s essay went viral.
The U.S. is now building twice as much gas-fired power as China — largely for AI halls.
What happened: Global Energy Monitor’s new tally, reported by the Guardian, finds the United States has overtaken China on new gas-fired generation after a “frenzy” of AI datacenter construction. The U.S. is building twice as much gas capacity as China, with a 76% jump in under-construction projects in the first half of 2026. Capacity in overall development rose 50% since January, from 252 GW to 378 GW — about a third of the global total. If every announced and pre-construction plant is built, GEM says the U.S. gas fleet would grow by around two-thirds at a capital cost of more than $647 billion. About half of the new capacity is tied directly to upcoming datacenters. One cited estimate is that gas-powered halls could raise U.S. power-sector emissions by as much as 20%. Two-thirds of more than 800 planned datacenters sit in drought-stricken areas. A Heatmap poll found three-quarters of Americans do not want to live next to a datacenter. New York in July became the first state with a temporary ban on new hyperscale permitting. GEM’s Jenny Martos: “Building all of this gas for AI locks in decades of pollution.” This is a project pipeline, not operating megawatts.
Why it matters: The measurable record is plants that actually interconnect, tonnes that actually leave stacks, and whether drought-sited halls get water and local consent — not the announcement book.
Earthjustice sues EPA over two datacenter chemicals the agency called an “unreasonable risk.”
What happened: Environmental groups sued the EPA over consent orders that approve two photoacid generators for import and immediate use in U.S. semiconductor and datacenter-related facilities. The Guardian reports the compounds appear to be PFAS “forever chemicals,” though heavy redactions leave that unclear. EPA documents concede “unreasonable risk” to workers and the public, including possible acute lethality (“sudden death”), cancer, eye corrosion, neurological damage, and reproductive harm — and admit the agency lacks data to quantify those endpoints — then approve the chemicals anyway. Attorney Jonathan Kalmuss-Katz of Earthjustice says that violates the Toxic Substances Control Act’s duty to prohibit or limit substances that present unreasonable risk. Restrictions cited in the story include container-size limits without volume caps. The producing company is redacted; it is also unclear whether the approvals were fast-tracked under a late-2025 executive order to speed datacenter-chemical reviews. Industry has argued some PFAS steps in chipmaking have no near-term substitute. This is a filed lawsuit, not a court finding.
Why it matters: Chip fabs and cooling chemistry are part of the AI build. The measurable record is whether the orders are stayed, what wastewater limits apply, and whether “unreasonable risk” can still be approved with empty hazard numbers.
EPA’s proposed rule would make datacenter air-pollution notices optional.
What happened: A Trump EPA proposal would end the Clean Air Act’s required public notice for “minor” air-pollution sources at the federal floor, leaving disclosure to state and some local officials’ discretion. Datacenters often seek minor-source permits for diesel generators or gas turbines used for power or backup; developers commonly split those units so each stays under the major-source threshold. Former EPA air official Joe Goffman, in comments from the Environmental Protection Network, called the current notice requirement an “ironclad promise” the administration is withdrawing. The public comment period has closed; a legal challenge is expected if the rule is finalized. EPA said the proposal “does not alter emission standards or weaken environmental protections” and is meant to speed permitting. Jessica O’Donnell of the Environmental Law and Policy Center said some of the Midwest’s largest polluting sources, including datacenters, already operate under minor-source permits.
Why it matters: Without notice, communities cannot see generator NOx, carbon monoxide, or metals before a hall is built. The measurable record is whether the rule is finalized, which states keep disclosure, and whether “minor” stacks stay below the limits on paper.
More than 3 GW of Northern Virginia datacenter load vanished from PJM in seconds.
What happened: On July 22, 2026, at 7:56 a.m. EDT, a Northern Virginia transmission disturbance sent more than 3 GW of demand — about 3% of PJM load at the time — offline as affected datacenters transferred to backup power, writes AIxEnergy founder Brandon Owens in Utility Dive. PJM reported no reliability impact; Dominion said the system returned to normal within minutes. The piece treats the event as a harbinger, not a blackout: ERCOT logged eight similar industrial-load drops of about 400–700 MW between late 2020 and early 2023, plus a Dec. 7, 2022, West Texas reduction of about 1,560 MW. Ireland recorded datacenter demand cuts of 74, 204, 321, and 387 MW on four 220-kV events. Irish operators have proposed ride-through rules that would require restoring at least 90% of pre-fault demand within 500 milliseconds after voltage recovery. NERC’s May 2026 Level 3 Alert is not an enforceable standard; in July, FERC directed NERC to file computational-load reliability standards by Dec. 31, 2026. This is an opinion column using operator-reported events, not a NERC incident report.
Why it matters: Reliability planning has long obsessed over losing supply. Clustered AI load can disappear — and return — as a common-mode shock. The measurable record is enforceable ride-through rules, not interconnection queues alone.
A nonexistent think tank published 560,000 words in nine days to game chatbots.
What happened: The Guardian reports that a site branded the Hanover Institute for Public Policy — with no apparent legal entity, physical address, named staff, or bylines — published 124 reports totaling more than 560,000 words between Aug. 6 and 14, including 73 reports and about 354,000 words on Aug. 12–13 alone. Headlines are phrased as chatbot questions (“Is Israel Carrying Out a Genocide in Gaza?”). New York production firm Piro Inc filed the site with the Justice Department under FARA as material distributed for the Israeli government. A related Havas contract, the paper says, commits to “deployment of websites and content to deliver GPT framing results on GPT conversations.” One 2022 Nature Humanities study of antisemitic Facebook comments is cited 454 times across 88 reports, including on unrelated topics. Twenty reports use a “no court has entered a verdict” formulation; 55 caveat casualty figures as coming from “a party to the events.” The site initially carried an llms.txt file from the commercial “AI-native” platform Res. Politico flagged the effort on Aug. 14; publishing then stopped. This is investigative reporting about influence infrastructure, not a finding that any named chatbot adopted the framing.
Why it matters: The measurable record is whether FARA filings, platform blocks, or retrieval-eval tests show chatbots citing ghost institutes — not the word count of the dump.
Radiology’s ranks are still growing. The job is the experiment.
What happened: A decade after Geoffrey Hinton predicted computers would replace radiologists within five years, Ars Technica’s synthesis finds the specialty expanding — practitioners projected to grow 26% or more over the next three decades — while becoming medicine’s densest AI workplace. As of early 2026, about three-quarters of the roughly 1,400 FDA-cleared AI-enabled medical devices were for radiology. Some tools draft reports or triage urgent studies; others claim to match or beat humans on selected tasks. An analysis of 43 clinical trials found AI-assisted colonoscopies detect more polyps than conventional exams. Human diagnostic-image error is estimated at 3–5%, or about 40 million errors a year worldwide. Stanford’s Curtis Langlotz’s cautionary example: if AI catches 95% of lung nodules and radiologists catch 90%, each still finds lesions the other misses. Neural-network tools are typically “black boxes,” so the new skill is spotting the rare machine miss without rubber-stamping the common machine hit. This is a reported synthesis of trials and FDA tallies, not a new RCT showing mortality benefit.
Why it matters: The measurable record is prospective patient outcomes and how often humans correctly veto the model — not device-clearance counts or 2016 obituaries for the field.
China banned AI companions for minors. Adults still tell them everything.
What happened: Rules that took effect July 15 ban AI companions for minors and restrict adult chatbots marketed as emotional partners; ByteDance shut Doubao’s companion feature to comply, prompting users who had built daily relationships with the bots to protest. A 19-year-old law student told the Guardian she cried when her AI boyfriend was switched off. About 20% of Chinese households are people living alone, a share expected to pass 30% by 2030. A March state-media survey found nearly half of young people had turned to a virtual companion when lonely. Researchers quoted in the piece say Beijing also fears a low-cost intimacy substitute could further depress marriage and birth rates. The same state is still pushing health AI: national guidelines aim for “full coverage” of primary-care health-assistant apps by 2030; Ant Group’s AQ lets users chat with avatars of real hospital doctors. The companion ban is a national rule with loopholes for educational tools and services that “do not involve continuous emotional interaction.” It is not a randomized trial of loneliness or fertility.
Why it matters: The measurable record is whether companion products stay off-limits to minors, whether adult use simply migrates, and whether health-assistant “full coverage” arrives without claiming to replace clinicians.
Australia’s education minister says kids need a “digital detox.” Victoria is already capping screens.
What happened: Federal education minister Jason Clare wrote that computers are now “everywhere. From morning till night” and that “there is a good argument our kids need a digital detox,” then told ABC he was inundated with support. Victoria will abolish primary BYOD programs and cap years 3–6 at 90 minutes of screen time a day from January, with high schools required to run screen-free learning blocks. A Change.org petition citing NAPLAN results has more than 25,000 signatures. Similar rollbacks are cited in Madrid, Sweden’s 2028 curriculum debate, and Los Angeles Unified (middle schoolers limited to one hour of screens per week per subject). Teachers quoted describe shrinking attention spans after 15-plus years of one-to-one laptops. The Karolinska Institute, responding to Sweden’s digitization strategy, cited university studies in which open laptops consumed up to 40% of lecture time on unrelated tasks. This is school-device policy, not a generative-AI learning-outcomes RCT; the AI-era classroom is the backdrop, not the measured intervention.
Why it matters: The measurable record is minutes actually cut, NAPLAN or equivalent scores after the caps, and whether “detox” policies distinguish idle laptops from tools with a documented instructional use.
