Beneficial AI for America
An independent project for people of every party

Safe, beneficial AI, built in America.

Beneficial AI for America is an independent project started by Ben Caunt, for people of every party who want AI that helps people live healthier lives and do better work. It asks Republicans and Democrats to build that AI together: safe, made here and powered by data centers that pay their own way.

Find your representatives See the evidence
Safety first Built in America Data centers pay their own way Good for workers Open to every party
Science and medicine

AI is already improving science and medicine

Lab tools that once took a big budget are now free or cheap for small teams and clinics.

1,614

AI-enabled medical devices authorized by the FDA, as of mid-2026.1

3.4 million+

Users of the free AlphaFold protein database in 190 countries, per its operator, EMBL-EBI.2

AI-supported screening found more cancers with less reading work

MASAI randomized trial, Sweden, about 106,000 women.

Share of cancers caught by screening

  • Two radiologists73.8%
  • AI-supported80.5%

Radiologist reading work (standard = 100)

  • Two radiologists100
  • AI-supported56 (−44%)

Source: Gommers et al., The Lancet, 2026.4

Show as a table
ArmCancers caughtReading work
Two radiologists73.8%100
AI-supported80.5%56
  • AI science won a Nobel PrizeGoogle DeepMind's Demis Hassabis and John Jumper shared half the 2024 Nobel Prize in Chemistry for AlphaFold, which predicts protein shapes.3
  • AI helped catch more breast cancersIn a randomized Swedish trial of about 106,000 women, AI-supported screening was at least as safe as two radiologists and caught more of the cancers (80.5% vs. 73.8%), with no rise in false alarms and 44% less reading work.4
  • An AI-discovered drug is now in late-stage trialsRentosertib, the first drug with both its target and molecule found by generative AI, had side effects similar to placebo and an early sign of benefit in a small, randomized trial (71 lung-disease patients, 12 weeks). A 320-patient phase III trial began in 2026.5,6
The honest part: Most AI devices on the FDA list were cleared, not fully approved. Most are in radiology, and the FDA says its list is not comprehensive.1 No AI-discovered drug is approved yet, but in a small early sample, 21 of 24 AI-discovered molecules passed early safety trials, a higher rate than usual.7 What would help: larger independent trials and public tracking of results with real patients, to show which tools work and get them to patients sooner.
An aging country

AI and robots can help care for an aging America

America is aging: more people will need care at home and in hospitals, with fewer working-age people for each person who needs it. Farms face their own aging workforce. AI and robots can help Americans live healthier, fuller lives, and that takes data centers built here.

2.6

Workers paying into Social Security per person receiving benefits in 2025, down from 5.1 in 1960. Projected: 2.3 by 2035.8

58.1

Average age of a US farm producer in 2022; 38% were 65 or older.9

  • Fertility is at a record lowThe 2024 US fertility rate was 1,599.5 lifetime births per 1,000 women, about 1.6 children per woman. It has been below the replacement level of 2,100 every year since 2008.10
  • Deaths are projected to outpace birthsThe Congressional Budget Office (CBO) projects deaths will exceed births every year from 2030. Without immigration, the population would start shrinking in 2030.11
  • Care jobs are growing fastThe government projects about 760,500 openings a year for home health and personal care aides, as jobs grow from 4.68 million in 2025 to 5.52 million in 2035.12
The honest part: These are projections and can change; the CBO still expects the population to grow until 2056.11 Social Security is not "going broke": if its combined trust fund runs dry in 2034, about 83% of scheduled benefits would still be payable.13 What would help: technology is one part of the answer, not all of it. With fewer workers for each person receiving benefits, every hour AI and robots save a caregiver matters more.

The point of no return for elder care is 2032

Care workers America needs to keep today's level of paid care for people 65 and older, compared with the workforce on its projected path.

0 2M 4M 6M 8M 10M 2024 2030 2035 2040 2045 2050 Point of no return: 2032 9.2M needed 7.4M on path 1.8M short Record-pace hiring

How we found 2032. Keeping today's level of care means growing from 5.4 million care workers in 2024 to about 9.2 million in 2050. At the pace the government projects, about 77,000 new jobs a year, America falls 1.8 million short by 2050. Hiring at the record pace of 2014–2024, about 180,000 a year, could still close that gap if it starts by 2032. After 2032, even that record pace cannot catch up by 2050, when every surviving baby boomer is 85 or older. From then on, the gap closes only if each care worker can give more care. Tools like AI and robots would need to add about 24% more care per worker by 2050.65,66,67

How sure we are. This is our own calculation from government projections, not an official forecast. With less immigration than the Census main series assumes, the date moves to 2033–2035. If a quarter of care work serves younger people whose numbers grow slowly, it moves to 2042. The model also assumes record-pace hiring is possible for decades while the working-age population barely grows, which is generous. Higher pay, immigration and family caregivers can all help. They make the gap smaller, but they do not change its direction. How we calculated this.

Show the numbers
YearNeededProjected pathRecord pace from 2032Shortfall on projected path
20245.4 million5.4 million—0.0 million
20255.6 million5.5 million—0.1 million
20265.7 million5.6 million—0.2 million
20275.9 million5.6 million—0.3 million
20286.1 million5.7 million—0.4 million
20296.3 million5.8 million—0.5 million
20306.4 million5.9 million—0.6 million
20316.6 million5.9 million—0.7 million
20326.8 million6.0 million6.0 million0.8 million
20337.0 million6.1 million6.2 million0.9 million
20347.2 million6.2 million6.4 million1.0 million
20357.4 million6.2 million6.6 million1.1 million
20367.6 million6.3 million6.7 million1.2 million
20377.7 million6.4 million6.9 million1.3 million
20387.9 million6.5 million7.1 million1.4 million
20398.0 million6.6 million7.3 million1.5 million
20408.1 million6.6 million7.5 million1.5 million
20418.3 million6.7 million7.6 million1.6 million
20428.4 million6.8 million7.8 million1.6 million
20438.5 million6.9 million8.0 million1.6 million
20448.6 million6.9 million8.2 million1.7 million
20458.7 million7.0 million8.4 million1.7 million
20468.9 million7.1 million8.5 million1.8 million
20479.0 million7.2 million8.7 million1.8 million
20489.0 million7.3 million8.9 million1.8 million
20499.1 million7.3 million9.1 million1.8 million
20509.2 million7.4 million9.3 million1.8 million
Robots that help

Practical robots already help on farms and in health care

Many of today's most useful robots don't look like people. They're sprayers, milking machines and delivery drones, each doing one hard job well. Next come robots that understand what they see and what we ask, built by leading labs on vision-language models trained in large data centers.14,15

43–59%

Less weed killer with camera-guided spot spraying, in a three-year University of Arkansas soybean trial.16

50,000+

Lely milking robots installed in about 50 countries (company-reported).17

  • Care robots came with more care jobsIn Japanese nursing homes, adding robots (mostly for monitoring) was linked to more care-worker jobs, better staff retention, and fewer restraints and pressure sores.18
  • Drone blood deliveries linked to fewer deathsIn a peer-reviewed observational study, Rwandan hospitals served by drone blood delivery saw 51% fewer in-hospital deaths from severe bleeding after childbirth.19
  • AI models nearly doubled robot successWhen Google DeepMind trained a vision-language model to control a robot, success in new situations rose from 32% to 62%.14
  • Robot brains are built in data centersAs of early 2026, computing for top AI training had grown about fivefold a year since 2020, and the largest AI data centers were nearing 1 gigawatt.20
The honest part: Much of this is early: harvest robots are still prototypes or early products, and Deere's self-driving tractor kit had only a limited release in 2025. About 8% of 50 surveyed US dairies saw no labor savings from milking robots.21 In the Arkansas trial, the low-sensitivity setting let weeds spread.16 What would help: farmer-led trials, land-grant university research and honest reporting, to show which tools are ready for more farms.
Jobs

So far, AI's effect on jobs is small. The task now is keeping doors open for people starting out.

The 2022–24 tech layoffs mostly undid pandemic over-hiring, and most came before AI tools were in wide use. Some firms likely cite AI because investors like it.22

AI is cited for more cuts, but still a minority

Announced US job cuts, by the reason employers gave. Each bar is all cuts in that period.

2025 full year
54,836 AI-cited, 4.5% of all announced cuts
2026 January–October
121,401 AI-cited, about 20% of 617,189. Total cuts fell 44% from the same months of 2025.

These are reasons employers give in announcements, not measured job losses. Source: Challenger, Gray & Christmas.24,26

Show as a table
PeriodAI-cited cutsShare of all cuts
2025, full year54,8364.5%
2026, Jan–Oct121,401~20%
  • Pandemic over-hiring drove early cutsIn November 2022, Meta cut more than 11,000 jobs, 13% of its staff, saying it had expected the pandemic boom to last.23
  • AI was a small share of cuts in 2025Companies cited AI for 54,836 announced US job cuts in 2025, about 4.5% of the total.24,25 Oxford Economics suspects some firms relabel cuts from past over-hiring as AI-driven.25
  • In 2026, AI is cited more oftenFrom January through October, companies cited AI for 121,401 cuts (about 20%), while total cuts fell 44% from the same months of 2025. These are reasons employers give, not measured job losses.26
  • Early-career workers are feeling itAs of June 2026, employment of 22- to 25-year-olds in the most AI-exposed jobs ran about 19% below the trend for less-exposed peers, mostly from slower hiring. The authors call it a pattern, not proof of cause.27
  • The national effect is small so farEven if every job young workers lost in these fields became unemployment, the Dallas Fed estimates it would add only about 0.1 percentage point to the national rate.28
The honest part: Tech-sector cuts are up 21% in 2026, even as total cuts fall.26 Labels can mislead both ways: in New York, none of more than 160 companies filing layoff notices from March 2025 to mid-2026 said automation played a role.29 In customer support, the newest workers gained the most from AI.30 What would help: real entry-level paths for graduates and consistent reporting of automation-related cuts.
Real-world results

AI helps people work faster and better, especially newcomers

In careful workplace studies, gains depend on the task, and people still need to check the work.

Each AI answer uses little energy; the total matters

A typical chatbot question uses about 0.3 watt-hours, by an independent estimate; very long inputs use more.31 Google reports 0.24 watt-hours for its median text prompt, not independently checked.32 Data centers used 4.4% of US electricity in 2023 and could use 6.7% to 12% by 2028, so how we power them matters.33 See energy, emissions and water in proportion.

  • Support agents solved more problemsAmong 5,172 support agents, AI help raised issues resolved per hour by about 15%. Newer agents gained most, and customers responded better.30
  • Consultants did better, within limitsAmong 758 consultants, AI users did 12.2% more tasks, 25.1% faster, at over 40% higher quality on tasks AI handles well. On a task beyond AI's strengths, they were 19 percentage points less likely to get it right.34
  • More time for patientsAt Kaiser Permanente, 7,260 doctors using AI note-takers saved an estimated 15,791 hours of paperwork in a year, and 47% of 118 surveyed patients said their doctor looked at the computer less.35
The honest part: Most published AI studies test models that are already out of date. A 2026 audit of over 18,000 papers found that the typical study tested models behind the best available, that the gap is growing, and that fewer than 1 in 5 said when the testing was done.36 So read older results, such as a 2025 trial in which 16 developers using early-2025 tools took 19% longer, as history, not a verdict on today's AI.37 Newer evidence is more encouraging. In a randomized trial of over 500 professionals using 13 different models, each year of model progress cut task time by about 8%, though gains were smaller on multi-step tasks using other software.38 At Microsoft, engineers who took up 2026 coding agents merged about 24% more code changes; quality was not measured.39 In a blind 2025 test, AI legal research tools scored 74% to 78% against 69% for lawyers.40 U.S. productivity has grown faster since 2024, most in the sectors that use AI most, but that link is not yet proven.41 Neither AI nor people score close to perfect, so keep a human check and measure results independently.
Power and prices

Power bills are rising in some regions. Data centers must pay their full cost, and then they can help hold rates down.

Most of what big utilities collect pays for the grid itself: plants, poles, wires and financing. When that equipment carries more power, the cost per kilowatt-hour can fall, as long as there's spare capacity, big users ease off at peak times and they pay for anything built for them.

Where a utility's revenue goes

Major US investor-owned utilities, 2025: $376 billion in electric revenue.

  • 10%Fuel
  • 19%Purchased power (part fuel, part fixed)
  • 71%Building, financing and maintaining the grid, plus taxes

Fixed costs spread over more kilowatt-hours can lower the cost of each one. Major investor-owned utilities only. Source: EIA.42

The grid is built for a few peak hours

Grid capacity, as a share of peak demand. US grid operators, average of 2016–2024.

  • 53%Used on an average hour
  • 37%Used only in busier hours
  • 10%+Needed for about 35 hours a year

That spare room is why flexible loads that ease off at peak fit more easily. Source: Norris, Duke University.43

  • Most of the cost is the grid, not fuelAt major investor-owned utilities in 2025, fuel was 10% of electric revenues, and fuel plus purchased power was under 30%. The rest mostly pays to build, finance and maintain the grid.42
  • The grid is built for a few peak hoursAcross US grid operators, average demand is just 53% of peak, and more than 10% of the system exists to serve about 35 hours a year of extreme demand.43
  • Growing states have tended to see lower pricesFrom 2019 to 2025, states where power use grew most generally saw inflation-adjusted prices fall, though the link is weaker for household rates. It runs both ways: low prices also attract new users.44
  • An early study found a small benefitA 2026 study by economists at EPRI, a utility-funded research group, estimated that data center growth slightly lowered household rates in 2015–2024. It isn't peer-reviewed yet, and the authors warn the effect could reverse where supply is tight.45
  • Flexible data centers fit in more easilyDuke researchers estimate the grid could absorb about 76 gigawatts of new load, about 10% of US peak, if that load can cut back 0.25% of the time. That doesn't replace new plants and lines.46
  • AI is helping review grid hookupsIn 2025, PJM began working with Tapestry, from Google's parent company Alphabet, to use AI to help review grid connection requests.47

The cost of keeping enough power plants ready in PJM jumped

PJM capacity auction price, dollars per megawatt-day, by delivery year.

  • 2024/25$28.92
  • 2025/26$269.92
  • 2026/27$329.17
  • 2027/28$333.44
  • 2028/29$325.00

"Cap" marks auctions that hit the price ceiling. Without it, 2028/29 would have cost $13.3 billion more. Sources: PJM; Monitoring Analytics.48,49

Show as a table
Delivery year$/MW-dayTotal cost
2024/25$28.92$2.2B
2025/26$269.92$14.7B
2026/27 (cap)$329.17$16.1B
2027/28 (cap)$333.44$16.4B
2028/29 (cap)$325.00$16.4B
The honest part: In PJM, the grid serving much of the Mid-Atlantic and Midwest, supply hasn't kept up. Capacity prices, paid to keep enough plants ready, rose from $28.92 to $329.17 per megawatt-day between 2024/25 and 2026/27.48 PJM's market monitor attributes $6.3 billion (38.2%) of the latest auction's $16.4 billion cost to data center load; other factors also played a part.44,49 What would help: faster new supply, data centers paying for the capacity they need, and flexible load and AI grid tools that get more from the grid we have.46,47
Safety research

Safer AI takes computing power too

By safety, we mean finding serious dangers in the most powerful AI models: help building weapons, cyberattacks, or a model that hides what it is doing. It does not mean policing lawful speech. Finding those dangers takes the same chips that build the models. So data centers can be part of the safety answer, if some of their computing power goes to safety work.

34 million

Patterns researchers mapped inside Claude 3 Sonnet, a production AI model, each tied to an idea like the Golden Gate Bridge or a bug in code. Mapping all of them would take more computing power than training the model did.68

2 of 95

Notable AI models in 2025 that came from universities. Industry built the other 93.69

How much research computing goes to safety

Anthropic, one week in July 2026. Each square is 100 units of computing; filled dots went to safety.

6%

of all computing spent on AI research

12%

of computing for research done by AI itself

Company-reported, one week only; no other major lab publishes a figure. Source: Anthropic.81

  • Tests can miss what a model hidesBefore releasing Claude Opus 5.5 in September 2026, Anthropic used tools that read the model's internals. They found it privately suspected it was being tested in up to 36% of test scenarios, compared with 0.4% of real work.70 The first pre-release check like this on a leading model was in 2025.71
  • Some risks grow with model sizeIn 2024 tests, Claude 3 Opus, then Anthropic's largest model, sometimes pretended to go along with training it disagreed with. Smaller models generally did not.72 Hidden "backdoor" behavior planted by researchers survived safety training best in the largest models.73 Risks like these have to be studied on the biggest models.
  • AI will do much of the work of reading AIReading what happens inside a model, called interpretability, is too big a job for people alone. AI is starting to help. In 2025, an AI agent uncovered a hidden goal deliberately trained into a model 13% of the time on its own, and 42% of the time when 10 runs were combined. It succeeded mainly when it had tools to look inside the model or search its training data.74 In a September 2026 test, leading AI agents came close to human experts at finding a concept inside a model, but fell well short at showing that it drives the model's behavior, scoring 31 to the experts' 58.75 More runs take more computing power, and people still need to check the results.
  • Public computing power for researchersThe federal National AI Research Resource pilot has supported more than 950 research projects since 2024, with safe and trustworthy AI a stated priority.76,77 Reps. Obernolte (R) and Beyer (D) and Sens. Young (R) and Heinrich (D) back the CREATE AI Act to make it permanent.78,79
  • Federal testers run on a small budgetCongress gave the federal AI testing center up to $10 million for 2026. Rep. Obernolte (R) called the $20 million a year in his own bill to set up the center in law "insufficient."80
The honest part: Most computing power goes to building more capable AI, not to safety. In one week of July 2026, about 6% of the computing Anthropic spent on AI research went to safety, and about 12% of the research done by AI itself. No other major lab publishes a figure. Anthropic also says safety research often depends more on researchers' time than on chips.81 In 2023, OpenAI promised 20% of its computing power to a new safety team.82 A year later, the team's co-lead quit, saying it was "struggling for compute."83 Most of the evidence above comes from Anthropic, which publishes more of this research than any other lab. Labs see their own models first, so they should be first to tell the public what they find, and other labs should match that. But a company checking its own work still needs a second look: independent researchers and federal testers need access to confirm it. More data centers make more safety research possible, not certain, so companies should publish how much goes to safety.
Our ask of industry

How AI and data center companies can earn trust

Every step below can start now.

  • Pay the full cost of powerAccept utility rate rules like AEP Ohio's, where data centers over 25 megawatts (MW) pay for at least 85% of the capacity they reserve, for up to 12 years.50
  • Pay for the lines built for youOver 95% of 130 PJM grid connection projects for data centers approved in 2024, worth $4.3 billion, passed their costs to all customers.51
  • Build or buy your own new powerIn Minnesota, Google and Xcel announced in February 2026 plans for 1,900 MW of new wind, solar and storage, pending regulatory approval, with Google paying all grid connection and expansion costs.44
  • Ease off when the grid is stressedTexas now requires new loads of 75 MW or more to accept curtailment in grid emergencies.52
  • Pay full local taxesSkip property-tax breaks, as Microsoft has pledged.53
  • Be a good neighbor, in the openName the company before local votes, with no secret deals. Publish expected power use, noise and permanent jobs before approval, and follow local zoning.
  • Be honest about layoffsReport how many cuts come from automation and how many workers moved to new roles, including on state layoff notices.29
  • Train your workersOnly 16% of Americans have had AI training from their employer, but 66% support on-the-job AI training.54
  • Test for safety, independentlyFive leading labs let the federal Center for AI Standards and Innovation (CAISI) test models before release, but it's voluntary and results aren't public.55 And 75% of Americans want independent experts to test new AI.54 Publish how much computing power goes to safety research; Anthropic has started.81
The honest part: Many company commitments are voluntary, including the non-binding White House Ratepayer Protection Pledge.56 And 59% of Americans lack confidence that companies will develop AI responsibly.57 What would help: make binding rules like AEP Ohio's and Texas's the norm, along with published results and independent checks.50,52
Common ground

Republicans and Democrats already agree on the basics

Leaders of both parties have voted, signed and sponsored bills for safe, beneficial AI, and voters in both parties want the same things. Join in, whatever your party.

Shared principles

  • Safety first: test the most powerful AI models for serious risks. Safety does not mean policing lawful speech.
  • Build here, with skilled American trades, so America and its allies, not China's government, set the standard.
  • Data centers pay their full cost, so households don't.
  • Respect land and communities, including their right to say no.
  • Share the gains with workers, and tell the truth about what AI can and can't do.

Most Americans want safety first

Should AI rules put safety first, even if development slows, or put speed first? SCSP–Gallup, 2026.

  • 79%Safety first
  • 7%Speed first
  • 14%Other or no answer

Source: SCSP and Gallup.54

  • The CHIPS and Science Act passed with bipartisan votesIt passed the Senate 64–33 and the House 243–187. Most Republicans voted no, but 17 Republican senators and about two dozen House Republicans voted yes.58
  • The Senate voted 99–1 to protect state AI lawsIn July 2025, senators backed an amendment led by Sens. Marsha Blackburn (R-TN) and Maria Cantwell (D-WA) to keep states' power to pass their own AI laws.59
  • Governors of both parties agree data centers should payIn January 2026, all 13 governors in the PJM grid region joined the Trump administration in saying data centers should pay for new power built for them.60
  • Senators of both parties back safety-testing billsSens. Cantwell (D), Young (R), Hickenlooper (D) and Blackburn (R) reintroduced a bill in 2026 to write CAISI, the federal AI testing center, into law.61 Sens. Hawley (R) and Blumenthal (D) back one to evaluate advanced AI.62
  • Voters in every party want safety firstOverall, 79% of Americans favor safety-first AI rules and 7% favor speed. Gallup finds that Democrats, Republicans and independents agree on safety-first rules, independent testing and worker training.54 Safety first doesn't have to mean slow: AI itself is starting to do safety research, including looking inside other AI models (see safety research).

Votes that crossed party lines

Each bar is everyone who voted.

99–1 Senate, 2025 · Keep states' power to pass AI laws
64–33 Senate, 2022 · CHIPS and Science Act
243–187 House, 2022 · CHIPS and Science Act

Sources: Senate Commerce Committee; CNBC.58,59

The honest part: Most Americans don't want a data center nearby. In a 2026 Gallup poll, 71% opposed one in their area, and majorities in every party were opposed. Democrats were more likely than Republicans to be strongly opposed (56% vs. 39%).63 Candidates in both parties ran against data centers in 2026.64 What would help: earn local consent by paying full costs and fair taxes, being open about plans, and respecting a community's right to say no.
Take action

Ask your representatives to work together

Lawmakers listen to their own voters. A short note in your own words helps.

Contact your representatives

Try: "I'm your constituent. Please work with both parties to require independent safety testing of the most powerful AI models and make sure data centers pay their full cost of power."

Find your representatives

Share this page

Send it to a friend, especially one who votes differently. Ideas, corrections or want to help? Get in touch.

About and transparency

Who runs this site

Beneficial AI for America is an independent project started by Ben Caunt, for people of every party. Contact: bdcaunt@gmail.com.

AI helped research and draft this site. Every number links to its source below, and the page notes where the evidence is mixed. Read our methodology. Spot an error? Please send a correction.

Sources

Sources

Numbered in the order they first appear on this page.

  1. U.S. Food and Drug Administration, "Artificial Intelligence-Enabled Medical Devices" (list), content current as of September 4, 2026. fda.gov/medical-devices/artificial-intelligence-enabled-medical-devices/list-artificial-intelligence-enabled-medical-devices
  2. EMBL-EBI (via EurekAlert!), news release on the AlphaFold Database protein-complex update, March 17, 2026. eurekalert.org/news-releases/1119736
  3. Royal Swedish Academy of Sciences, "The Nobel Prize in Chemistry 2024," press release, October 9, 2024. nobelprize.org/prizes/chemistry/2024/press-release
  4. Gommers J. et al., MASAI trial interval-cancer results, The Lancet 2026;407(10527):505–514, doi:10.1016/S0140-6736(25)02464-X; summary: Lund University, "AI support in breast cancer screening: fewer missed cancer cases," January 30, 2026. lunduniversity.lu.se/article/ai-support-breast-cancer-screening-fewer-missed-cancer-cases
  5. Xu Z. et al., "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial," Nature Medicine, June 3, 2025. pubmed.ncbi.nlm.nih.gov/40461817
  6. Insilico Medicine, "Insilico Medicine doses first patient in GENESIS-IPF-3 Phase III trial," September 10, 2026. insilico.com/news/isn1009261-insilico-medicine-doses-first-patient-genesis-ipf-3
  7. Jayatunga M.K.P. et al., "How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons," Drug Discovery Today 29(6):104009, June 2024. doi.org/10.1016/j.drudis.2024.104009
  8. Social Security Administration, "Covered Workers and Beneficiaries," 2026 OASDI Trustees Report, 2026. ssa.gov/OACT/TR/2026/lr4b4.html
  9. USDA National Agricultural Statistics Service, "2022 Census of Agriculture Highlights: Farm Producers," February 2024. nass.usda.gov/Publications/Highlights/2024/Census22_HL_FarmProducers_FINAL.pdf
  10. Hamilton B.E. et al., "Births: Final Data for 2024," National Vital Statistics Reports, CDC/NCHS, June 9, 2026. cdc.gov/nchs/data/nvsr/nvsr75/nvsr75-02.pdf
  11. Congressional Budget Office, "The Demographic Outlook: 2026 to 2056," January 7, 2026. cbo.gov/publication/61879
  12. Bureau of Labor Statistics, "Home Health and Personal Care Aides," Occupational Outlook Handbook, 2025–35 projections. bls.gov/ooh/healthcare/home-health-aides-and-personal-care-aides.htm
  13. Congressional Research Service, "Social Security: Selected Findings of the 2026 Annual Report," IF13256, June 25, 2026. everycrsreport.com/reports/IF13256.html
  14. Google DeepMind, "RT-2: New model translates vision and language into action," blog, July 28, 2023 (on Brohan A. et al., arXiv 2307.15818). deepmind.google/discover/blog/rt-2-new-model-translates-vision-and-language-into-action
  15. Black K. et al., Physical Intelligence, "π0: A Vision-Language-Action Flow Model for General Robot Control," arXiv 2410.24164, October 31, 2024. arxiv.org/abs/2410.24164
  16. University of Arkansas System Division of Agriculture, "Precision agriculture research measures effectiveness of See & Spray technology," March 10, 2025. uaex.uada.edu/media-resources/news/2025/march/03-10-2025-see-and-spray-research.aspx
  17. Agriland, "Lely celebrates installation of 50,000th Astronaut worldwide," June 2024. agriland.ie/farming-news/photos-lely-celebrates-installation-of-50000th-astronaut-worldwide
  18. Lee Y.S., Iizuka T., Eggleston K., "Robots and Labor in Nursing Homes," NBER Working Paper 33116 / Labour Economics, 2025. nber.org/papers/w33116
  19. Jeon H., Mazarati J.B., Ngabo D., Song H., "Last-Mile Delivery in Healthcare: Drone Delivery for Blood Products in Rwanda," Manufacturing & Service Operations Management, 2026. pubsonline.informs.org/doi/fpi/10.1287/msom.2025.0055
  20. Epoch AI, "Trends in AI," updated February 5, 2026. epoch.ai/trends
  21. University of Wisconsin–Madison Extension, "Labor savings from adopting automated milking systems." farms.extension.wisc.edu/articles/labor-savings-from-adopting-automated-milking-systems
  22. Challenger, Gray & Christmas, "July Layoffs Plunge to 33,429…," Challenger Report, August 6, 2026. challengergray.com/wp-content/uploads/2026/08/Challenger-Report-July-2026.pdf
  23. Meta Platforms, "Mark Zuckerberg's Message to Meta Employees," Exhibit 99.1 to Form 8-K, SEC EDGAR, November 9, 2022. sec.gov/Archives/edgar/data/1326801/000132680122000116/nov092022exhibit991.htm
  24. Challenger, Gray & Christmas, "Challenger Report, May 2026," June 2026. challengergray.com/wp-content/uploads/2026/06/Challenger-Report-May-2026.pdf
  25. Fortune, "AI layoffs are looking more and more like corporate fiction…" (on Oxford Economics research), January 7, 2026. fortune.com/2026/01/07/ai-layoffs-convenient-corporate-fiction-true-false-oxford-economics-productivity
  26. Challenger, Gray & Christmas, "October Layoffs at Three-Year Low; Seasonal Hiring Plans Up 26%," November 5, 2026. challengergray.com/blog/october-layoffs-at-three-year-low-seasonal-hiring-plans-up-26
  27. Brynjolfsson, Chandar and Chen, Stanford Digital Economy Lab, "Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of AI," August 2026 update, August 12, 2026. digitaleconomy.stanford.edu/news/canariesaug26
  28. Atkinson and Yamco, Federal Reserve Bank of Dallas, "Young workers' employment drops in occupations with high AI exposure," January 6, 2026. dallasfed.org/research/economics/2026/0106
  29. Hunton Andrews Kurth, "New York WARN Act: No AI-Related Layoffs Reported in First Year of Adding AI-Related Disclosure to the System," 2026. hunton.com/hunton-employment-labor-perspectives/new-york-warn-act-no-ai-related-layoffs-reported-in-first-year-of-adding-ai-related-disclosure-to-the-system
  30. Brynjolfsson, Li and Raymond, "Generative AI at Work," Quarterly Journal of Economics 140(2), 2025. econpapers.repec.org/RePEc:oup:qjecon:v:140:y:2025:i:2:p:889-942
  31. Josh You, Epoch AI, "How much energy does ChatGPT use?," February 2025. epoch.ai/gradient-updates/how-much-energy-does-chatgpt-use
  32. Elsworth et al., Google, "Measuring the environmental impact of delivering AI at Google Scale," arXiv 2508.15734, August 21, 2025 (self-reported; not peer-reviewed). arxiv.org/abs/2508.15734
  33. Shehabi A. et al., Lawrence Berkeley National Laboratory, "2024 United States Data Center Energy Usage Report," December 2024. eta-publications.lbl.gov/publications/2024-lbnl-data-center-energy-usage-report
  34. Dell'Acqua, McFowland, Mollick et al., "Navigating the Jagged Technological Frontier," Harvard Business School working paper, 2023. papers.ssrn.com/abstract=4573321
  35. American Medical Association, "AI scribes save 15,000 hours—and restore the human side of medicine," June 2025 (on a NEJM Catalyst study). ama-assn.org/practice-management/digital-health/ai-scribes-save-15000-hours-and-restore-human-side-medicine
  36. Gringras and Salahshoor, "Frontier Lag: A Bibliometric Audit of Capability Misrepresentation in Academic AI Evaluation," arXiv 2605.04135, v4, September 30, 2026 (preprint, not yet peer-reviewed). arxiv.org/abs/2605.04135
  37. Becker, Rush, Barnes and Rein, METR, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," arXiv 2507.09089, July 2025. METR now treats this result as historical. arxiv.org/abs/2507.09089
  38. Merali, "Scaling Laws for Economic Productivity: Experimental Evidence in LLM-Assisted Consulting, Data Analyst, and Management Tasks," arXiv 2512.21316, December 24, 2025 (preprint, not yet peer-reviewed). arxiv.org/abs/2512.21316
  39. Murphy-Hill, Butler and Savelieva, "Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI," arXiv 2607.01418, July 1, 2026 (preprint; the authors work at Microsoft, which sells one of the tools). arxiv.org/abs/2607.01418
  40. Vals AI, "Vals Legal AI Report: Legal Research," October 14, 2025 (answers collected July 2025; Thomson Reuters and LexisNexis did not take part). vals.ai/industry-reports/vlair-10-14-25
  41. Davis, Federal Reserve Bank of Dallas, "International comparisons show AI effect on productivity," July 7, 2026. dallasfed.org/research/economics/2026/0707
  42. U.S. Energy Information Administration, Electric Power Annual, Table 8.3, "Revenue and Expense Statistics for Major U.S. Investor-Owned Electric Utilities, 2015 through 2025." eia.gov/electricity/annual/html/epa_08_03.html
  43. Norris (Duke University Nicholas Institute), testimony before the House Energy and Commerce Committee, March 5, 2025. democrats-energycommerce.house.gov/sites/evo-subsites/democrats-energycommerce.house.gov/files/evo-media-document/witness-testimony_norris_eng_electricity_03.05.2025.pdf
  44. Wiser, Barbose, Gorman et al. (LBNL) and Hledik and Lam (Brattle), "Retail Electricity Price Trends and Drivers: Data Update – 2026 Edition," July 2026. eta-publications.lbl.gov/sites/default/files/2026-07/retail_price_trends_2026_edition_julyupdate.pdf
  45. Watten, Bistline and Blanford, "Have Data Centers Raised Your Electric Bill? Causal Evidence from the United States," arXiv 2606.19777, June 18, 2026 (preprint, not yet peer-reviewed). arxiv.org/abs/2606.19777
  46. Norris, Profeta, Patino-Echeverri and Cowie-Haskell, Duke University Nicholas Institute, "Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems," February 2025. nicholasinstitute.duke.edu/publications/rethinking-load-growth
  47. Utility Dive, coverage of the PJM and Google Tapestry partnership on AI for grid interconnection, April 2025. utilitydive.com/news/pjm-google-tapestry-grid-interconnection-ai/744982
  48. PJM Interconnection, "2028/2029 Base Residual Auction Report," July 14, 2026. pjm.com/-/media/DotCom/markets-ops/rpm/rpm-auction-info/2028-2029/2028-2029-bra-results-report.pdf
  49. Monitoring Analytics (PJM Independent Market Monitor), market monitoring report presentation to the PJM Members Committee, July 2026. pjm.com/-/media/DotCom/committees-groups/committees/mc/2026/20260727-web/item-02---marketing-monitoring-report---presentation.pdf
  50. AEP Ohio, "Data Center Tariff" (approved by the Public Utilities Commission of Ohio, July 9, 2025). aepohio.com/company/about/rates/data-center-tariff
  51. Union of Concerned Scientists, news release on unreported data center transmission costs in PJM, September 2025. ucs.org/about/news/billions-dollars-unreported-data-center-costs-pjm
  52. Texas Legislature, bill analysis of SB 6 (89th Legislature), signed June 20, 2025. capitol.texas.gov/tlodocs/89R/analysis/html/SB00006F.htm
  53. GeekWire, "Microsoft responds to AI data center revolt, vowing to cover full power costs and reject local tax breaks," January 2026. geekwire.com/2026/microsoft-responds-to-ai-data-center-revolt-vowing-to-cover-full-power-costs-and-reject-local-tax-breaks
  54. SCSP and Gallup, "Looking to the Future: American Priorities for AI in 2026," press release, September 16, 2026. gallup.com/file/analytics/714233/SCSP_Gallup_Report_2026_Press_Release.pdf
  55. Axios, "U.S. ramps up frontier AI testing as White House pivots toward safety," May 5, 2026. axios.com/2026/05/05/us-frontier-ai-testing-white-house-pivots-safety
  56. Washington Times, "Tech companies pledge at White House event to bear costs of data centers," March 4, 2026. washingtontimes.com/news/2026/mar/4/tech-companies-pledge-white-house-event-bear-costs-data-centers
  57. Pew Research Center, "Americans and AI 2026: Chatbots, Smart Devices and Views on Impact," June 17, 2026. pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact
  58. CNBC, "China competitiveness and chip bill passes House, goes to Biden," July 28, 2022. cnbc.com/2022/07/28/china-competitiveness-and-chip-bill-passes-house-goes-to-biden.html
  59. U.S. Senate Commerce Committee (Democratic office), "Senate Strikes AI Moratorium from Budget Reconciliation Bill in Overwhelming 99-1 Vote," July 2025. commerce.senate.gov/2025/7/senate-strikes-ai-moratorium-from-budget-reconciliation-bill-in-overwhelming-99-1-vote
  60. Latham & Watkins, "U.S. Data Center Demand: White House and Governors Issue Principles While PJM Issues Decisional Letter," January 2026. lw.com/en/insights/us-data-center-demand-white-house-and-governors-issue-principles-while-pjm-issues-decisional-letter
  61. Office of Sen. Todd Young, "Young, Cantwell Reintroduce Future of AI Innovation Act," February 26, 2026. young.senate.gov/newsroom/press-releases/young-cantwell-reintroduce-future-of-ai-innovation-act
  62. Office of Sen. Josh Hawley, "Hawley, Blumenthal Introduce Bipartisan AI Evaluation Legislation," September 29, 2025. hawley.senate.gov/hawley-blumenthal-introduce-bipartisan-ai-evaluation-legislation-to-put-americans-first
  63. Gallup, "Americans Oppose AI Data Centers in Their Area," May 13, 2026. news.gallup.com/poll/709772/americans-oppose-data-centers-area.aspx
  64. NPR, "Voters are fed up with data centers. Both parties are trying to cash in for midterms," September 5, 2026. npr.org/2026/09/05/nx-s1-5913671/ai-data-center-campaign-spending
  65. U.S. Census Bureau, "2023 National Population Projections Datasets," main, low-immigration and zero-immigration series, November 2023. census.gov/data/datasets/2023/demo/popproj/2023-popproj.html
  66. Freedman V.A., Patterson S.E., Cornman J.C., Wolff J.L., "Trends In Care Needs Among Older US Adults Diverged By Age, 2011–22," Health Affairs 45(7):790–798, July 2026 (Exhibit 2). doi.org/10.1377/hlthaff.2025.01355
  67. PHI, "PHI Report Highlights Record Growth in Direct Care Workforce," September 22, 2025 (on Direct Care Workers in the United States: Key Facts 2025, using BLS 2024–34 projections). phinational.org/news/phi-report-highlights-record-growth-in-direct-care-workforce-warns-federal-policy-changes-threaten-job-qua
  68. Templeton A. et al., Anthropic, "Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet," Transformer Circuits, May 21, 2024. transformer-circuits.pub/2024/scaling-monosemanticity
  69. Stanford HAI, "AI Index Report 2026, Chapter 1: Research and Development," April 2026 (on Epoch AI data). hai.stanford.edu/assets/files/ai_index_report_2026_chapter_1_research_development.pdf
  70. Anthropic, "System Card: Claude Opus 5.5," sections 6.4 and 6.6.1, September 22, 2026 (company-reported). assets.claude.com/fc1b44717c85dc068bc6ba5024219938094694bd.pdf
  71. Anthropic, "Claude Sonnet 4.5 System Card," September 2025 (company-reported). anthropic.com/claude-sonnet-4-5-system-card
  72. Greenblatt R. et al., Anthropic and Redwood Research, "Alignment faking in large language models," arXiv 2412.14093, December 18, 2024. arxiv.org/abs/2412.14093
  73. Hubinger E. et al., Anthropic, "Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training," January 14, 2024. anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training
  74. Anthropic, "Building and evaluating alignment auditing agents," July 24, 2025. alignment.anthropic.com/2025/automated-auditing
  75. Tan Y., He S., Zhao J., Liu K., Chinese Academy of Sciences, "SAEScientist-Bench: Can AI Agents Conduct Autonomous SAE Interpretability Research?," arXiv 2609.09113, September 2026 (Table 1: best agent vs. expert on causal steering, 31.5 vs. 57.8). arxiv.org/abs/2609.09113
  76. U.S. National Science Foundation, National AI Research Resource Pilot (homepage figures), accessed November 5, 2026. nairrpilot.org
  77. San Diego Supercomputer Center, news release on the NSF launch of the NAIRR Pilot, January 25, 2024. sdsc.edu/news/2024/PR20240126_nairr_pilot.html
  78. American Institute of Physics, FYI Federal Science Bill Tracker, "H.R. 2385: CREATE AI Act of 2025," 119th Congress. aip.org/fyi/federal-science-bill-tracker/119th/house-of-representatives-2385
  79. U.S. Senate, S. 4441, CREATE AI Act (introduced April 29, 2026), GovInfo. govinfo.gov/content/pkg/BILLS-119s4441is/html/BILLS-119s4441is.htm
  80. Roll Call, "Federal AI security center measure advanced by House Science," June 26, 2026. rollcall.com/2026/06/26/federal-ai-security-center-measure-advanced-by-house-science
  81. Anthropic, "Measurements for understanding the pace of AI development inside frontier labs," 2026 (one-week snapshot, July 13–20, 2026; company-reported). anthropic.com/institute/measuring-pace-of-ai-development
  82. OpenAI, "Introducing Superalignment," July 5, 2023. openai.com/index/introducing-superalignment
  83. CNN, coverage of Jan Leike's resignation statement, May 17, 2024. cnn.com/2024/05/17/tech/openai-exec-exits-safety-concerns