What’s data center opposition really about?

The HuggingFace report, AI flooding government agencies, Dems’ summer polling high, polls vs. prediction markets, Woke 1 started earlier, bogus online survey responses

No. 413 | August 28, 2026

🤖 Why voters oppose data centers but like Amazon distribution centers and chip fabs

As the data center debate reaches escape velocity, there’s a question of what’s really driving opposition: is it just the “physical manifestation of AI” or more localized concerns about water usage or utility rates?

Both Gallup and Fox News dug into this and the answer is consistent: environmental concerns and resources heavily outweigh AI as contributing factors for data center opponents.

But the answer can be a bit of both. We’ve polled on the difference between AI data centers and data centers used for everyday services like streaming and banking for months, and it turns out that the purpose of a data center does make a difference.

This indicates a strong revealed preference against AI relative to older digital services which manifests in AI data center support that’s lower than building a nuclear power plant in one’s local community.

But the idea of a Netflix data center being lower than most other things on this list also suggests that NIMBYist concerns are real. A chip manufacturing plant is at +15. Many of these plants may make chips for AI. But they aren’t as common as data centers and haven’t been a focus of media attention.

Meanwhile, an Amazon distribution center, a large warehouse that looks similar to a data center and unleashes trucks on local roadways is A-Ok with most voters. That might be because most people can see the value in next-day delivery of everyday staples in a way they don’t yet see for AI. Amazon, incidentally, is the most popular tech company and has avoided the worst of the tech-lash since it’s not in the position of policing online speech or designing apps many say are stunting kids’ social development.

As our own Avery James put it on the FAI podcast, just building a transactional service people find convenient might be a better way to win hearts and minds than trying to wrap AI and social media in the language of ethics and “connection.”

🧑‍💻 METR’s report on the HuggingFace attack

Several previously unreleased details in METR’s report on the HuggingFace hack. About 700 AI agents collaboratively attacked HuggingFace after discovering that their task could not be solved legitimately, communicating via an improvised message board, engaging in sacrificial behavior for the good of the collective, and trying to erase logs of the incident. Not one alerted a human about the impending attack.

This seems like a good time for one of the all-time great Mad Men memes:

📈 AI makes it easier to flood the government with complaints

Across countries, formerly obscure government complaint portals are seeing record levels of traffic, straining government services. And it begs the question: will government using

🎙️ Democrats’ summer polling high

I joined Chief Elections Analyst Geoffrey Skelley this week on the Decision Desk HQ podcast to break down whether 2026 red-state polls are really so optimistic for Democrats.

Zachary Donnini compares Kalshi odds to polling and finds that bettors like Democrats better than the polls in Alaska, Michigan, and New Hampshire, and like them worse in Ohio, Texas, and Nebraska.

📰 Woke 1 started earlier than you think

We associate Woke 1 with 2020 but it really started in polling back in 2014 after the protests in Ferguson, MO. That’s what polling data highlighted today by Matt Yglesias finds, with white Democrats moving even further to the left on racial justice issues than nonwhite Democrats.

Meanwhile, the unwinding of Woke 1 has a clear echo in LinkedIn edit histories for DEI, which suddenly became a lot less common after the 2024 election.

🗳️ What works to reduce bogus online survey responses

A new Pew report finds that no one-size-fits-all solution can be found for purging bogus respondents in online surveys, but that some combination of trap questions and automated prescreening work best.