> daily_signal(2026_11_05)
Voters rejected AI data centers in 38 of 40-plus races. Also today: Google's school-AI ultimatum, a way to prove which model answered you, and a safer stroke robot.
PickBits Daily Signal · Wednesday, November 5, 2026
// tl;dr
- Voters rejected AI data centers in 38 of more than 40 races across 13 states on Tuesday. Six cities banned them outright and twelve capped the megawatts a facility can pull. Texas Governor Greg Abbott lost a fourth term to Gina Hinojosa by more than 700,000 votes after taking over $9 million from data-center-linked donors, and Elon Musk's roughly half-billion-dollar midterm spend didn't change the result.
- Starting in early 2027, a school that turns off Google's AI for students loses Google Search too. Gemini, Gemini Notebook, and Search will be off by default for student accounts; a school can switch each on, but plain Search without AI is gone. Turning Search on means AI Overviews and chat, or a "Guided Learning" mode. New York City and Los Angeles pulled back on student AI in September, and LA's request to keep Search without AI was refused.
- Anthropic open-sourced FlashProve, a way to prove which model actually generated a response without exposing its weights. It's aimed at the router or reseller that quietly serves a cheaper model than the one you paid for. The prototype runs at about 110 tokens per second on one H100, produces a roughly 19 MB proof that verifies in 17 seconds, and for now only covers short contexts and greedy decoding, not the sampling real chatbots use.
- A bench study of a stroke-clot robot found it works far better with a clinician steering than on its own. Across six control modes on a patient-specific vascular phantom, full autonomy scored 81% in simulation but collapsed to 23% in the physical model, while shared clinician-AI control hit 92-97%, near direct robotic control at 95% and approaching manual at 100%. It is a 13-participant preprint on one phantom, not a patient trial.
We've spent months on this one, and it comes down to one line on your power bill: who pays to wire up a hyperscaler, you or the company running servers you'll never touch? In June we covered the Seattle engineers Amazon pulled into HR after they testified for a moratorium, and the House fight over whether Big Tech or ratepayers cover data-center power. Tuesday it stopped being a council fight and turned into an actual vote, and the data-center side lost almost everywhere it was on the ballot. The boom's real limit isn't chips or power. It's whether the town next door says yes, and this week a lot of towns said no.
Google is pulling the one setting that let a school keep Search without turning its AI on. Anthropic shipped a way to check that the model you paid for is the one that answered, the exact worry we raised about Copilot in June. And a stroke-robot study found the safe setup isn't the machine on its own, it's a doctor with a hand on it.
Today voters in 13 states put AI data centers on the ballot and turned them down in 38 of more than 40 races.
1. Voters rejected AI data centers in 38 of more than 40 races this week.
Six cities banned them outright, twelve capped the megawatts a site can draw, and Texas's governor lost his seat after taking millions from data-center donors.
Data-center proposals were explicitly on the ballot in more than 40 races across 13 states on Tuesday, and voters rejected them in 38 and counting. Six cities banned new data centers outright; twelve capped the megawatts a single facility may draw from the grid. The backlash reached local offices too: a ban on high-impact data centers in Emporia, Kansas, a recall of Yukon, Oklahoma's mayor over a hyperscale project, and anti-data-center mayors elected in Asheville, North Carolina and San Marcos, Texas.
The loudest result was in Texas. Governor Greg Abbott lost a bid for a fourth term to Democrat Gina Hinojosa by more than 700,000 votes after accepting over $9 million from data-center-linked donors, and Elon Musk's roughly half-billion-dollar midterm spend didn't change the outcome. For months this fight lived in rate cases, council hearings, and that fight in Congress. This week it got to the ballot box in 13 states, and the data-center side lost it almost everywhere.
Why this matters: The data center going up near you now has to get past your vote, not just a zoning board. A data center means cheap land and a huge power draw, and it lands on your electric bill. This week people said no. My own read: you can't buy your way out of this one. Musk spent half a billion dollars in these races and the data-center side still lost, because people were voting on their own power bill.
Action this week: Pull your state utility commission's open data-center cases before the next hyperscale request lands, because the megawatt caps and who-pays rules that just won at the ballot get decided in those filings first, not at the next election. Your commission is in the naruc.org directory, and datacentersontheballot.com is tracking how the live ones land. When I've dug into these dockets, the cost-shift is never buried; it's written right there in the filing.
404media.co: American voters overwhelmingly reject AI data centers (November 4, 2026)
wired.com: The results are in: voters hate data centers (November 2026)
technologyreview.com: Texas was the new epicenter of the AI data-center buildout. That might be kaput. (November 4, 2026)
2. From 2027, a school can't keep Google Search unless it keeps Google's AI on.
Student accounts get Gemini and Search off by default, each one re-enabled separately, but the plain-Search-without-AI setting is being removed for good.
Starting in early 2027, Gemini, Gemini Notebook, and Google Search will be off by default for student accounts, and a school can switch each on individually. The setting that disappears is the one many districts actually wanted: Search with the AI stripped out. Turn Search on and the only options are full AI Overviews and AI chat, or a "Guided Learning" mode. A version of search without AI is, in the reporting, "no longer on the table." Google is at the same time reversing a recent change that had switched Classroom AI on by default after that drew backlash.
Google already owns this market, through Chromebooks, Gmail, and Classroom. New York City and Los Angeles, the country's two largest districts, pulled back on student AI in September; LA specifically asked to keep Search without AI Overviews and was told no. Google's education VP, Chris Phillips, basically said to deal with it: search and AI are too tightly fused to pull apart now. We've watched Google push AI further into school accounts all year, from the NotebookLM rename in July to a student hub in August. This is the one where opting out of the AI costs you the search bar too.
Why this matters: Pretty soon a kid on a school Google account won't be able to keep AI out of the search bar at all. The tool the school can't drop now comes with the AI attached, so "off by default" stops meaning you can keep it off.
Action this week: Ask your district in writing, before the early-2027 rollout, which Search mode student accounts will default to and whether any AI-free path survives. Put it to curriculum and IT in the same email, because when I've asked institutions this out loud, nobody can name who actually owns the setting. And if the IT is yours to run, the real call is whether "Guided Learning" Search is good enough for minors or whether you need a non-Google path. I'd want that one settled before the switch flips, not after.
Listen to the Daily Signal podcast at pickbits.ai/daily-signal.
3. There's a new way to prove which AI model actually answered your query.
Anthropic open-sourced FlashProve, which uses a zero-knowledge proof to show a response came from the claimed model without revealing its weights.
Anyone paying for a named model has had no real way to check that the named model is what ran. A routing layer or reseller can serve something cheaper, and the stream looks identical. Anthropic's FlashProve is an open-source research prototype that attacks exactly that gap: the provider fingerprints the model's weights once, then ships a compact zero-knowledge proof (a zkSNARK) with each response that shows it came from the claimed model, without exposing the weights themselves.
It's early, and the overhead is brutal. FlashProve runs at about 110 tokens per second on a single H100 for a 400-billion-parameter model, produces a proof of roughly 19 MB that verifies in 17 seconds on an ordinary CPU, at something like 70 to 95 times the inference cost. That's enormous. But it's a huge step down from where this started: a prior comparable system needed over 14 hours and a 93.6 GB proof for 1,000 tokens. For now it only covers contexts up to 1,024 tokens, only works with greedy decoding rather than the sampling real chatbots use, and hides the weights but not the architecture. In June we flagged that Microsoft could quietly swap a cheaper model into Copilot and you wouldn't know; this one's real code you can download, not a white paper.
Why this matters: You pay for a top-tier model through an API or a reseller, but you've been trusting a label you can't actually check, and whoever is in the middle has every reason to swap in something cheaper and pocket the difference. FlashProve doesn't fix that today, because it can't yet cover production chat. But "is the model you billed me for the one that answered" is something you can actually check now, instead of just hoping.
Action this week: Put one line on your vendor checklist now: how do you verify the model that served this response is the one we are paying for? The honest answer today is "you mostly can't," and that is fine for the moment. I wouldn't accept a cloud bill with no line items either, so I don't see why inference gets a pass. Just don't let anyone sell you "verified inference" as shipped yet: test FlashProve's 1,024-token and greedy-decoding limits against your real workloads first, because those constraints rule out most live chat and agents.
the-decoder.com: Anthropic's new prototype proves which model wrote an AI response without exposing its weights (November 4, 2026)
github.com/anthropics/FlashProve: code, Lean proofs, and docs under Apache 2.0
4. A stroke-clot robot worked best with a clinician steering, not on its own.
Across six control modes on a vascular phantom, full autonomy collapsed from 81% in simulation to 23% in the physical model, while shared clinician-AI control hit 92-97%.
Pulling the clot out of a stroke, a procedure called mechanical thrombectomy, needs a specialist most hospitals just don't have on call, which is the whole reason anyone is looking at a robot for this at all. Researchers tested a robotic navigation system for it across six modes, from fully manual to fully autonomous, with 11 expert clinicians and two students on a patient-specific vascular phantom. Fully autonomous navigation looked strong in simulation at 81% success, then collapsed to 23% once it had to work in the physical model. The clinician-AI shared-control modes recovered almost all of that, hitting 92 to 97% success, close to direct robotic control at 95% and approaching hands-on manual at 100%.
The authors aren't pushing for more autonomy. They're asking who's in control, and when the doctor can take over. The setups where the clinician decided when to take over worked best, and the doctors trusted them more. It is a bench study, 13 participants on one phantom, not yet tested in animals or patients. This could one day put expert stroke care in a hospital that has no specialist, and the team's read is to keep the doctor on it. We've been covering autonomous surgical robots for weeks, and the fear they'll push doctors out of the room. This is the one that put a number on it: 23 percent alone, 97 with a doctor.
Why this matters: The safe robot here is the one a doctor is steering, not the one left to run itself. Full autonomy aced the simulation and fell apart on the physical model, a 58-point drop, and that's exactly where having a doctor on the controls mattered.
Action this week: Watch whether this team moves from a phantom to animal and then patient studies, and whether the "shared control beats full autonomy" result holds once a real patient is on the table. If you build or evaluate surgical AI, test the human-in-the-loop setups against full autonomy on real hardware, not just in simulation, since that sim-to-real gap is where I'd expect most autonomy claims to quietly fall apart. The preprint is public on arXiv if you want the mode-by-mode numbers.
» What to watch this week
- Which legislatures turn the ballot wins into law, and which hyperscalers sue the city bans. Losing 38 of 40 races and a governor's seat is a real result, but the developers weren't on the ballot and they have far more money than the towns do. The city bans are the ones most likely to end up in court.
- Google's "no" to Los Angeles, and whether a big enough district can change it before 2027. LA already asked for Search without AI and got refused. My bet is no district actually walks, and this just pushes schools onto non-Google search and classroom tools instead.
- FlashProve's token-context cap and greedy-decoding constraint. Those two limits are what keep it out of real chat today. If they loosen, inference proofs stop being a lab demo. And watch whether a single provider decides "we can prove it" is worth selling as a feature.
- The stroke-robot team's next step off the phantom, into animal and then patient studies. It's 13 people on one phantom, not a trial yet. I'd watch whether that sim-to-real collapse shows up again as the tests move closer to a real artery.
Tomorrow's signal lands here.