> daily_signal(2026_09_20)
Government came at AI from three directions, from Virginia's data centers to a Senate camera hearing to federal heart-care money, while a top lab clocked its AI running a quarter of its research.
PickBits Daily Signal · Sunday, September 20, 2026
// tl;dr
- Virginia, the world's largest data-center hub, signed Executive Order 22 and a Data Center Accountability Framework. It bars state executive-branch agencies from signing or requiring non-disclosure agreements on data-center projects, pushes local sign-off on giant power draws, orders a cumulative-impact review of diesel backup generation, names lower ratepayer costs as a goal, and stands up a state AI Task Force.
- Sen. Josh Hawley summoned the CEOs of Flock Safety, Axon, Motorola Solutions, and Verkada to testify about their AI camera networks. He says the systems pool records into a searchable national database and that the overwhelming majority of Americans caught in them "did nothing wrong." A Data for Progress poll found roughly 41% view Flock cameras unfavorably, with Democrats and Republicans nearly even.
- Anthropic says its AI, Claude, now "leads" 26% of the company's R&D, up from under 1% in February, and touches more than 90% of it. But "leads" is one rung (AL4) on a 0-to-5 autonomy scale where a human still signs off; full autonomy is reached nowhere; and its own raters agreed on the level only about a third of the time.
- ARPA-H's ADVOCATE program awarded $62.7 million over four years to build an agentic AI "member of the care team" for heart-failure patients between clinic visits. Tempus AI, UpDoc, and Atman Health build the patient-facing system (FDA package due within two years); Stanford builds a supervisory AI; Duke and Kaiser test it in real clinics. It targets the roughly half of U.S. counties with no cardiologist.
Three different arms of government are pushing back on AI, each from a different direction. Virginia, of all places the state with the most data centers on earth, moved to slow its own gold rush and drag the deals into daylight. A Senate subcommittee called the four biggest AI-camera vendors toward a hearing table. And a federal health agency put real money behind AI aimed at patients the specialist system never reaches. None of these arcs is new to us. We've been on the data-center power bill, on Flock's quiet spread, and on ARPA-H's access bets all year. The difference this week is that each one finally has a date on it.
The fourth story is the odd one out, and the one I wouldn't skip. Anthropic published its own count of how much of its research its AI now runs, and the number, 26%, is built to be quoted out of context. The honest version is more interesting than the scary one: even the lab measuring it can't cleanly say what "leads" means, and its own scorers disagree a third of the time. What gets me is the timing. Governments spent years treating this as too abstract to touch, and they moved the same week the people building it admitted they can't measure how fast it's going.
Virginia just made it illegal for the state to sign a secret NDA on the data center going up next to your town.
1. Virginia moved to slow its data centers and end the secrecy around them.
Executive Order 22 bans state NDAs on data-center deals, pushes local sign-off on giant power draws, and names lower ratepayer costs as a goal.
We have been on the data-center power bill all summer, from Dominion Energy being ordered to assign transmission costs directly to data centers, to the reports of residents watching their supply charges climb to wire up someone else's servers. The missing piece in almost every one of those fights was the same: the public couldn't see the deal. When a hyperscaler and a utility strike an arrangement, the terms sit behind non-disclosure agreements and redacted filings, and the county board voting on the land is often the last to know what the thing will actually pull off the grid.
On September 18, the state with more data-center capacity than anywhere on the planet went at exactly that. Governor Abigail Spanberger signed Executive Order 22 and a Data Center Accountability Framework her office calls the most comprehensive and aggressive in the country. The move that actually matters is the ban on secrecy: no executive-branch agency or employee may enter into or require an NDA on a data-center project. Around it sits a community-engagement toolkit and local planning power, expedited noise rules that had stalled for years, and a cumulative-impact review of the diesel backup generators these campuses run, with lower energy costs for ratepayers written in as an explicit goal. The order also stands up a rapid-response state AI Task Force aimed at workforce displacement, data privacy, and cybersecurity, the risks everybody talks about and few states actually put anyone on.
Why this matters: The reason this reaches past Virginia is the NDA ban, not the noise rules. Once one state, and the biggest one, refuses to let these deals hide, residents everywhere can point at Virginia and ask their own utility commission for the same disclosure. The catch is that an executive order isn't a law. It binds this administration and can be undone by the next governor with a signature, which is why the thing I'll be watching is whether the legislature writes any of it into statute before the politics turn.
Action this week: Ask your county board what a proposed data center will draw off the grid and who pays to build for it, because in Virginia that answer can no longer be sealed. My own read, after watching the Mississippi ratepayer fight go the way it did, is that these deals get decided in utility-commission dockets almost nobody attends, so find out if yours has a data-center case open right now and read who is on the witness list.
2. A senator summoned the AI camera companies to testify.
Nobody voted to build a searchable national record of where cars go; it accreted camera by camera, and now the Senate wants the vendors under oath.
AI-powered cameras reading license plates are spread across the country now, and the striking thing about how they got there is that no one ever voted for a national system. It grew one county at a time, until the local records quietly pooled into something searchable coast to coast. On September 18, Sen. Josh Hawley, who chairs the Senate Judiciary Committee's Crime and Counterterrorism Subcommittee, invited the CEOs of the four biggest vendors to testify: Flock Safety's Garrett Langley, Axon's Rick Smith, Motorola Solutions' Greg Brown, and Verkada's Filip Kaliszan, who says he can't attend but will provide information.
Hawley opened a formal investigation in August into how Flock collects, retains, and shares the data its cameras generate, arguing the systems pool records into a searchable national database rather than serving one discrete local case, and that the overwhelming majority of people swept in "did nothing wrong." This isn't new ground for us, we have covered the reporting on police hiding their use of Flock cameras and the leaked Flock guide coaching officers on how to talk city councils into a contract. What is new is who is angry. A Data for Progress poll found about 41% of people view Flock cameras unfavorably, with Democrats at 40% and Republicans at 39%, close enough to call even. You don't often see both parties land in the same place on a tech question, and here they did, hard enough to reach a Senate subcommittee.
Why this matters: A hearing isn't a rule, and a summoned CEO can decline and often does. But this one puts the real question on the record: did anyone actually agree to pool every local plate read into one nationwide, searchable database? Whether a single camera is legal was never the hard part. Plenty of people want a camera that catches a car thief, and the same people don't want a permanent, searchable log of everywhere they've driven.
Action this week: Find out whether your town contracts with Flock, Axon, Motorola, or Verkada, because many city-council agendas list it and Flock runs a transparency portal. Then get one thing answered before the next renewal: the retention period and who else can query the data. When I have pushed local officials on that, the honest ones admit they never asked, which is exactly how a local safety tool becomes a national one nobody signed off on.
rollcall.com: Flock cameras controversy set for Senate hearing (September 18, 2026)
3. Anthropic says its AI now "leads" a quarter of its research.
The 26% headline is the part everyone will quote; the part that matters is that even the lab's own scorers can't agree on what "leads" means.
If your work is knowledge work, this is the number that lands in your next planning meeting, so it is worth getting right before someone extrapolates it to your team. On September 18, Anthropic published its own internal measurement: Claude now "leads" 26% of the company's AI research and development, up from under 1% in February, and participates in more than 90% of it. The figures were released to back CEO Dario Amodei's call for a coordinated slowdown and to give the public a way to track how fast this is moving. Taken at face value, it reads like a job posting for a machine.
Then you read how Anthropic actually defined "leads." It scores work on an autonomy scale from Epoch AI, running AL0 to AL5, where "leads" is AL4: an engineer sets the direction and Claude executes independently, but nothing ships without a human's approval. Full autonomy, AL5, with no human in the loop, is reached nowhere. And the metric is soft by Anthropic's own admission: employees rating the same piece of work landed on the same level only about a third of the time, and Claude grading itself matched human judgment 59% of the time. That is a wobbly foundation when a single rung is the whole difference between "collaborates" and "leads." Anthropic is being unusually candid about how soft the number is, which I respect. It's still a soft number.
Why this matters: A frontier lab is now quantifying how much of its own cognitive work its AI carries, and publishing the trend line on purpose. That is genuinely useful, right up until the number gets ripped from its definition and used to justify a headcount cut somewhere that has nothing to do with AI research. The gap between "Claude participates in 90% of our work" and "Claude does 90% of our work" is the entire ballgame, and the second version is the one that will travel.
Action this week: Read "26% leads" against Anthropic's own definition before you repeat it, because AL4 still means a human directs the work and approves every ship. If you benchmark AI's effect on your own function, adopt an explicit autonomy scale rather than a single "percent automated" figure, given even the vendor's raters split on the level a third of the time. I have watched too many "AI does X%" claims collapse the moment you ask what X actually counted, and this one tells you upfront that it would.
PickBits Daily Signal is free. If it lands in your inbox every day and it is worth something to you, the best way to support it is to forward it to someone who would read it. Subscribe today!
4. The government funded an AI "member of the care team" for heart failure.
It is aimed at a heart patient in a county with no cardiologist, on their own for the weeks between visits when things quietly go wrong.
Picture someone discharged after a heart-failure scare, living in one of the roughly half of U.S. counties with no cardiologist. The dangerous stretch isn't the hospital, it is the weeks between appointments, alone, when a small change goes unnoticed until it is an emergency. That is the exact gap the government's health advanced-research agency, ARPA-H, put money behind this month. Under its ADVOCATE program, Agentic AI-Enabled Cardiovascular Care Transformation, it awarded $62.7 million over four years to build an agentic AI meant to act like a member of the care team, checking on heart-failure patients between visits and escalating to humans when something looks wrong.
It is the same strategy we saw earlier this month, when ARPA-H funded AI to shorten the rare-disease "diagnostic odyssey": aim federal money at the patients the specialist system leaves behind. What I like is how carefully they built it. Tempus AI, UpDoc, and Atman Health build the patient-facing system, which must file an FDA authorization package within two years. Stanford builds a separate supervisory AI whose only job is to catch the first one when it makes an unsafe or out-of-distribution call. And Duke, across five health systems including rural sites, plus Kaiser Permanente, tests it in real clinics. It answers a complaint you hear from every clinician who has tried this: most health AI dies before the bedside because nobody translated it into something a doctor can actually use.
Why this matters: The hopeful part has nothing to do with replacing a cardiologist. It is that an agentic AI aimed squarely at scarcity, at the counties that have no specialist to replace, is the version of this technology that can do the most good with the least downside. A tool that watches a patient no human was going to watch anyway is a very different bet from one sold to a hospital that already has a full cardiology wing.
Action this week: Watch the access metric, not the announcement, because the honest caveat is that this is a research bet and not a product anyone can get. The milestone that counts isn't the press release or even the FDA filing due in two years; it is the first rural patient a pilot actually reaches. My own read is that the safety architecture, a second AI built to overrule the first, is the smartest thing in this award, and I will believe the impact when a county that never had a heart doctor tells me the thing caught something in time.
medtechdive.com: ARPA-H to award $62.7M for AI in cardiovascular care (September 2026)
» What to watch this week
- Whether Virginia's legislature writes any of EO 22 into law. The NDA ban only outlives this governor if it becomes a statute; watch the next session, and watch which other data-center states copy the framework.
- Whether the Flock CEOs actually show. A summons isn't a subpoena. If the vendors decline, the next question is whether Hawley's subcommittee escalates, and whether any of them disclose retention and data-sharing practices under oath.
- Whether any other lab publishes an AL-scale number. Anthropic proposed the autonomy scale as a public yardstick. It only becomes a yardstick if a rival lab reports a comparable figure the same way, disagreement rate included.
- The first ADVOCATE pilot in a cardiology desert. Duke's rural sites and Kaiser's clinics are where this stops being a contract and starts being care, or stalls the way most medical AI does between funded and deployed.
Tomorrow's signal lands here.