Quick Take  ·  News  ·  Put AI to Work  ·  Looking Ahead

Quick Take

Here’s the thing that struck me this week. The people already using AI at work are not asking anyone’s permission, and they’re not asking for a ban either. They’re asking for help.

Municipal staff in small Pennsylvania towns are drafting meeting notes with it and would like to know what a good policy looks like. Half of American workers are using it on the job. Meanwhile the rules and the training mostly haven’t shown up.

That gap is an opening, and there’s a lot of room in it.

Let's get into it.

AI News Roundup

AI got to local government before the rules did

What happened: The University of Pittsburgh’s Institute for Cyber Law, Policy, and Security partnered with the Local Government Academy, a Pennsylvania nonprofit that runs education programs for local governments, to survey small municipalities, mostly senior staff, about how they’re using generative AI. Mostly townships and boroughs serving between 5,000 and 20,000 residents. It’s a small survey, 35 respondents, so hold it loosely. The answers are still worth reading.

Two-thirds said they use generative AI at least monthly. More than half said their workplace had already rolled out enterprise tools. What they’re using it for is unglamorous and completely believable: writing, summarizing, meeting notes. Most of them have no formal policy governing any of it, and more than half of the officials without a policy said they’d find one valuable.

The barrier isn’t what you might guess. Half of respondents named a lack of staff expertise as the biggest obstacle, ahead of budget at 38 percent. When asked what would help, they didn’t ask for a framework handed down from the state. They asked for training, support, and live peer-to-peer learning, with examples from governments that look like their own.

Why this matters: There’s a version of the progressive AI conversation that assumes the fight is about whether these tools get used at all. In small-town government, I think that question is already settled. It got settled quietly, by staff who needed to get a newsletter out and a meeting summarized, while the rest of us were arguing about it.

What’s missing is the part progressives are historically good at: showing up with practical help for under-resourced public institutions. My guess is that a township with a part-time manager isn’t a promising target for a consulting firm, because there isn’t much money in it. So the guidance comes from civic organizations and peer networks, or it doesn’t come.

And notice what the officials asked for. Not a rulebook. Peer learning, from places like theirs. That’s a request for a network, and building networks is something we already know how to do. Worth noting that one of the survey’s own partners, the Local Government Academy, exists to do exactly that kind of peer education. The supplier and the demand are already in the same room.

There’s a cost to leaving it alone, too. A policy vacuum doesn’t stay empty. It fills up with whatever the vendor’s onboarding deck said, and a resident’s data could end up somewhere nobody chose on purpose.

A four-part test for whether “open source” means anything

What happened: Writing in Tech Policy Press, JJ Jasser proposes a clean test for a word that’s getting stretched past usefulness. A model is open source, the argument goes, only when its architecture, training code, model weights, and training data are all publicly available under licenses that permit unrestricted use, modification, and redistribution. All four. Not three.

Jasser applies it to Moonshot’s Kimi K3, which has been marketed as open. The weights were promised but not released as of writing. The license is a non-standard “Modified MIT” that isn’t OSI-approved. The training data is closed. The training pipeline is closed. Calling it open source, Jasser writes, is “a category error.”

Why this matters: Last issue we covered a coalition of ten organizations assembling an open-source AI assistant in seven weeks, and I argued that the public interest sector can build its own infrastructure rather than waiting to be included. I still think that’s right. This piece is the necessary follow-up, because that strategy depends entirely on “open” meaning something specific.

If you’re a nonprofit choosing a model to build on, the four criteria map directly onto risks you actually carry. Closed training data makes it much harder to audit for bias or check where the material came from, and you’re going to get asked about both. A non-standard license means your lawyer has to read it, and might tell you no. Weights that haven’t shipped mean you’re planning around a press release.

You don’t need to get cynical about companies that overclaim. You need a checklist, so “open” becomes a question you can answer instead of an impression you pick up from a launch announcement.

Briefly — the “Avoiding AI” workshops are packed, and that’s worth understanding

Public librarians in South Philadelphia and Bangor, Maine have been running workshops on how to turn off AI features across phones, laptops, and platforms. They’re full. The Bangor sessions drew about 70 people each counting the livestream, with in-person spots capped at 30 and a waitlist. The South Philly library’s post about it got more than 2,000 likes.

Before anyone in our audience reads this as the enemy: both librarians describe it as digital literacy work. “As a librarian, I think it’s important to see this as advancing digital literacy and helping people reclaim their autonomy,” Charlie Bailey told TechCrunch. His counterpart in Bangor talks about teaching people the basics of what’s happening so they can make informed choices.

That’s our position too. One attendee’s line explains the whole phenomenon: “I keep getting AI shoved down my throat at work.” That’s not a complaint about AI. It’s a complaint about not being asked. The demand these workshops are meeting is for agency, and the people meeting it are librarians, which should surprise nobody.

Source: TechCrunch, 25 July 2026

Put AI to Work

Practical ways progressives can use AI this week

Stop researching which AI to use

Ethan Mollick published a guide this week that I want to boil down to its most useful instruction, because a lot of people reading this are stuck at the starting line comparing options.

His advice: pick Claude or ChatGPT, pay the $20, and give an agent a real task from your real life.

That’s it. That’s the step.

A few things worth knowing around it:

For low-stakes work, the free tier is fine. Drafting a routine letter, asking a casual question, getting unstuck on some copy. You don’t need a subscription for that.

For anything that matters, pay and use the good model. Mollick points to Claude’s Opus or Fable, or ChatGPT’s GPT-5.6 Sol set to “High” thinking, for medical, legal, or complicated professional questions. The error rates are meaningfully lower. That’s a real distinction and it’s worth $20 a month if a wrong answer costs you anything.

Treat the output like a first draft from a new team member. Mollick’s framing, and it’s the right one. He suggests reviewing the result critically and asking for revisions the way you’d give notes to someone on your team, rather than taking what you get. That’s the habit worth building early.

Leave the approvals on. For anything that sends, spends, or deletes, keep the confirmation step in place until you genuinely trust it. That’s Mollick’s caution and it’s a good default for an organization where somebody else’s data is on the line.

The reason I’m leading with “just pick one” is that tool comparison is a very comfortable way to avoid starting. The learning happens on the first real task, not in the evaluation spreadsheet.

Source: One Useful Thing, 23 July 2026

Looking Ahead

• PULSE applications close August 7. From last issue: state and territorial health departments, local and county health departments, tribal authorities, and large city health departments can apply for donated enterprise AI seats. If you have any connection to a public health agency, this is the last useful week to forward it.
• Watch whether anyone supplies what those township officials asked for. A model AI policy plus a peer learning session for small municipalities is cheap to produce, and the survey suggests the demand is going largely unmet. If a civic tech group or a municipal association picks it up, that’s a template worth copying in every state.
• The evidence question is still open. Steve Newman published a piece on July 20 called “Anecdotes Everywhere, Evidence Almost Nowhere,” arguing that the effects of AI “are not (yet) large enough to stand out unambiguously in hard data.” Both things can be true at once: half of workers using it, and the results not yet visible in the economic statistics. Worth keeping in mind as the adoption numbers climb, in either direction you’re inclined to argue.

Until next time,
Jordan

Read past issues  ·  RSS  ·  Website