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

Quick Take

For most of this newsletter's life we've been asking a version of the same question. Who gets to decide how AI is built, who benefits, who gets left out. Those are good questions and we're going to keep asking them.

But this issue is about people who stopped asking.

A group of ten organizations, including Hugging Face, Mozilla, and MIT Media Lab, put together an open-source AI assistant in seven weeks. It comes out of an organization whose whole push is the languages and communities that the big commercial models treat as an afterthought. Nobody granted them permission. They just built it.

Back in June we covered Bernie Sanders' proposal for public ownership of AI, a one-time tax on the biggest labs paid in stock into a fund the public would own. That's a demand made of the companies that already exist. What's happening here is different, and I think more interesting: not a claim on someone else's stack, but a parallel one.

The same thing is showing up in smaller ways. Legal aid organizations are building their own tools for benefits navigation. Public health agencies can now apply to pilot AI. A union struck over AI and came back with contract language about it.

Let's get into it.

Ten jurisdictions will be selected to pilot it, and the use cases are the unglamorous ones that actually matter: disease surveillance, mapping social determinants of health, multilingual translation, pulling clinical data.

Worth being precise: this is a commitment, not a deployment. The jurisdictions haven't been chosen yet and pilots don't start until the fall. But applications are open right now, through August 7, to state, territorial, tribal, county, and large city health departments. If you know someone in one, that's a real deadline worth passing along.

---

AI News Roundup

A public AI stack, assembled in seven weeks

What happened: At the AI for Good Summit in Geneva, the nonprofit Current AI launched Alpha Chat, an open-source chatbot put together in seven weeks by a coalition of ten organizations. Hugging Face, Mozilla, and MIT Media Lab were among them, and the model here is worth noticing: each contributor supplied a piece of the stack rather than one organization building the whole thing. A language model from one, safety tooling from another, computing power from a third.

The point of the organization behind it is who gets served. Current AI has been putting grant money toward AI that works in more than 50 African languages, in Arab cultural archives, and for Indigenous communities in the Amazon, with $3.2 million going to four organizations across Kenya, Lebanon, and Brazil. A separate partnership put 22 Indian languages onto a device that works offline. The organization has been around since February 2025, led by Ayah Bdeir, who came over from Mozilla, and it has roughly $400 million committed to it, seeded by the French government alongside the Ford and MacArthur foundations.

Why this matters: The usual progressive move when a technology leaves people out is to demand inclusion. Make the companies serve these languages. Make the training data represent these communities. Those demands are correct and I've made them in this newsletter more than once.

The trouble is that a demand puts you on someone else's timeline. You ask, they consider, you wait. Meanwhile the languages that don't have a commercial case attached keep not getting served.

What Current AI did instead was treat the stack as something you can assemble. Seven weeks, ten organizations, each contributing what it already had. That's not a moonshot budget or a decade-long research program. It's closer to what a coalition of well-organized nonprofits could actually attempt, and that's the part I'd want progressive technologists to sit with.

Here's the honest caveat: an open chatbot assembled quickly is not the same as a model that competes with the frontier labs, and nobody involved is claiming it is. The interesting claim is smaller and more useful. The public interest sector can build its own AI infrastructure, and this one got assembled in seven weeks out of parts the contributors already had.

---

Legal aid shipped an AI that helps families keep their food benefits

What happened: Frontline Justice, a national nonprofit, teamed up with the legal automation company Josef to launch Frontline Q, an AI assistant that answers questions about SNAP eligibility and appeal rules. It's rolling out in Arizona, Texas, and Alaska. The tool pulls together federal, state, and local regulations, and it runs under the supervision of legal aid lawyers rather than on its own. The people using it are community justice workers who get training, onboarding, and ongoing supervision from local partner organizations in each state. Frontline Justice CEO Nikole Nelson says the organization's advocates have recovered more than $20 million in SNAP benefits for families since late 2022. Medicaid and housing support are on the list of possible expansions.

Why this matters: Longtime readers will recognize the shape of this. Back in May we covered Tipping Point Community raising $40 million to put AI tools in the hands of frontline nonprofit staff. This is the same idea, further along: not funding to eventually get AI to direct service workers, but a working tool aimed at one specific problem.

The design choice I keep coming back to is the supervision. Benefits law is exactly the kind of domain where a confidently wrong answer does real harm, and somebody clearly thought about that. Lawyers oversee it. Trained workers operate it. The AI handles the part it's genuinely good at, which is holding an enormous, fragmented body of rules in memory and answering questions about them quickly.

That's also the answer to a question this newsletter gets a lot: where does AI actually fit in progressive work? Not replacing the advocate. Giving the advocate a faster way through the rulebook so more families get helped.

---

Illinois passed a frontier AI safety law, and the labs said thank you

What happened: Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act on July 6. It applies to what the law calls large frontier developers, meaning companies above $500 million in yearly revenue training models with massive computing power. Those companies have to publish safety plans addressing catastrophic risk, which the law defines as incidents that could contribute to the death of or serious injury to more than 50 people, or more than $1 million in property damage. They have to report critical safety incidents within 72 hours, or within a single day if there's imminent risk of death or serious injury. There are whistleblower protections and confidential reporting channels for employees. Penalties run $1 million for a first violation and $3 million after that. It takes effect January 1.

The genuinely new piece is an annual independent third-party audit requirement, described as the first of its kind in the country. And the bill passed with broad bipartisan support, with both Anthropic and OpenAI backing it on the way through. An Anthropic spokesperson called it "an important step toward the accountability this technology demands."

Why this matters: I want to be careful here, because the easy version of this story is that regulation is working and the companies have come around, and I don't think that's quite it.

Look at what the law regulates. Catastrophic risk. Fifty deaths. A million dollars in property damage. Those are real things to worry about and I'm glad someone is thinking about them. But they are not the AI harms that show up in the work most of you do. Nobody's SNAP application got denied by a catastrophic risk event. The hiring screener we covered three weeks ago that rejected Black and Asian applicants at higher rates wouldn't trigger a single provision of this law.

So when two of the largest AI labs publicly back a bill, the useful question isn't whether they've had a change of heart. It's what this framework asks of them, and what it doesn't. My read: a catastrophic risk regime is expensive in reputation and fairly cheap in practice for a company already running safety teams, while rules on algorithmic discrimination, on labor displacement, on surveillance pricing would cost real money and change real products. That's my judgment, not something either company has said.

None of which makes the Illinois law bad. The third-party audit requirement is a genuine precedent and the whistleblower protections could matter a lot. Lawmakers estimate that Illinois, California, and New York together account for roughly 40 percent of the U.S. AI market, which if anywhere near right makes state law the operative standard in the absence of a federal one. Just notice which bills the industry shows up to support, and which ones it doesn't.

---

---

Put AI to Work

Practical ways progressives can use AI this week

Build a rapid-response verification kit in an afternoon

Every organization doing advocacy work eventually needs to answer some version of "is this real?" A screenshot going around. A photo attached to a claim. A statistic that sounds too convenient. Jeremy Caplan and Craig Silverman recently pulled together three free tools that make this faster, and you can have all three working before the end of the day.

1. Set up scheduled monitoring with Yutori Scouts (about 20 minutes). You tell it what topics to watch and how often, and AI agents go out, collect links, and send back a summary. Point it at the issues you'd otherwise be checking manually every morning. The setup cost is small and it runs on its own after that.

2. Install Search by Image for reverse image checks (about 5 minutes). It's a Chrome extension that right-clicks any image and runs it across Google Lens, Bing, TinEye, and Yandex at once. This is the fastest way to find out whether a photo is what someone says it is, or whether it's a real photo from a different event five years ago.

3. Learn SearchWhisperer for hard searches (about 20 minutes). It turns a plain-English question into advanced search operators, so you can narrow by document type, by domain, by exact phrase, without memorizing the syntax. Useful for finding a specific filing, an archived page, or a document somebody would rather you didn't find.

4. Run one real check today (about 15 minutes). Take something currently circulating in your issue area and run it through all three. The tools stick when you've used them once under real conditions instead of bookmarking them for later.

The payoff: when something breaks and everyone's asking whether it's real, you're the person who can answer in ten minutes rather than tomorrow.

Source: Wonder Tools

---

---

Looking Ahead

• PULSE applications close August 7. State, territorial, tribal, county, and large city health departments can apply. Ten jurisdictions get selected, pilots start in the fall. If you have any connection to a public health agency, this is the concrete thing to forward this week.
• The Illinois law takes effect January 1. The thing to watch is the third-party audit requirement, and whether any other state copies it. That provision is the real precedent, more than the catastrophic risk framing around it.
• Watch whether the coalition model gets repeated. Ten organizations each contributing a piece of the stack is a replicable structure, and if a second coalition forms around a different gap, that's a pattern rather than a one-off.

Until next time,
Jordan

---

Read past issues on the web · Subscribe via RSS · Website