Switched off overnight

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

In this issue

The federal government switched off Anthropic's two most powerful AI models overnight, using national-security authority and no public process. A licensing power a lot of us asked for arrived without us in the room.
A progressive researcher got rare direct access to Congress to make the case that gig platforms are rewriting nursing rules before workers can organize.
Google DeepMind is worried about AI agents acting on each other with no human watching. For benefits claimants and tenants, that future is already here.
The Roosevelt Institute published its own rules for using AI, and handed every advocacy org a template.
Put AI to Work: real-time translation is finally usable, and it can open your services to people you could not reach last month.

Quick Take

On Friday the federal government reached over and switched off the two most powerful AI models in the country. No hearing, no vote, no public notice. One company's CEO raised a concern, and three days later the most capable AI Anthropic had ever shipped went dark for everyone.

Your gut reaction tells you something. If it was "good, somebody finally reined them in," or if it was "see, this whole thing is dangerous," sit with that for a second. Both reactions hand the keys to someone else. The power to decide who gets to use AI just got demonstrated in the most dramatic way possible, and the question that matters is not whether you like this one decision. It is who gets to make the next one, and whether any of us are in the room when they do.

That is the thread running through everything this week. The people most affected by AI are the least likely to be at the table. Our whole job is to change that. Let's get into it.

Three days

That is how long Anthropic's Fable 5, the most capable AI model the company had ever released, was available to the public before the federal government forced it offline. It launched on a Tuesday and was dark by Friday. A generational decision about who controls the most powerful AI in the country, made in less time than it takes most agencies to answer an email. Fortune has the reporting; Anthropic confirmed it.

The government just invented AI licensing, and we weren't in the room

On June 12, the Trump administration used national-security export-control authority to bar all foreign nationals from using Anthropic's two newest models, Fable 5 and the more capable, less-restricted Mythos 5. The trigger was specific. Amazon researchers found they could prompt the Mythos-class model into producing restricted information about cyberattacks, and Amazon's CEO flagged it to senior officials. Because Anthropic could not reliably wall off every foreign user in real time, including some of its own foreign-born employees, it shut both models off for everyone to comply. Fable 5 had been public for three days. As of this writing there is a fast-moving negotiation to bring it back.

Why this matters

Strip away the specifics and look at what got demonstrated. The federal government can decide, on its own, who is allowed to use the most powerful AI in the country. That is a licensing power, and it is exactly the kind of governance a lot of us have been asking for. The catch is how it arrived: by executive fiat, in seventy-two hours, with no law behind it, no public comment, and no one representing workers, communities, or the public interest anywhere near the decision. The first real exercise of AI gatekeeping in America happened entirely inside the executive branch and a couple of corporate boardrooms. Cheering the outcome or mourning it both miss the point. The precedent being set right now, about who decides and through what process, will outlast this particular shutdown by a decade.

What you can do

When your representatives or allied policy shops talk about "AI safety" or "national security" framing, push the question one level down: who decides, and how? A governance process any of us could live with has public notice, a comment period, and a seat for labor and civil society. Ask for those things by name now, while the template is still wet, because the version that calcifies is the version nobody objected to in time.

A progressive researcher took the nursing fight straight to Congress

AI Now Institute senior fellow Dr. Katie J. Wells testified before the House Subcommittee on Workforce Protections about how gig-work platforms, the "Uber for nursing" model, are lobbying states to rewrite healthcare staffing rules. The result would leave a largely female, disproportionately immigrant nursing workforce with less pay, fewer protections, and less control over their shifts. The strategy is legislative as much as technical: lock in the AI-enabled labor arrangement before workers have any organizing foothold.

Healthcare AI usually gets debated as a technology question, which keeps it safely abstract. Wells reframed it as a worker-rights question and got a progressive voice into the congressional record on it, which almost never happens. The platforms are not waiting for the technology to mature before they reshape the rules. They are doing both at once, and the people whose working conditions directly affect patient safety are the ones being cut out.

What you can do

When the written testimony posts, pull it, strip out the procedural framing, and turn the core argument into a one-page brief for any healthcare-worker or nursing-union partners you have. This is exactly the evidence base local labor coalitions need when a gig platform shows up at their statehouse.

DeepMind is worried about AI agents talking to each other. For some people that's already today.

Google DeepMind's AGI safety director Rohin Shah is funding research into what happens when large numbers of AI agents start interacting at scale, taking instructions from each other rather than from humans and producing behavior nobody can fully predict. It is being studied as a frontier safety problem, something to get ahead of before it arrives.

For a lot of people it has already arrived. Benefits claimants, tenants, and low-income consumers already deal with automated decision systems in housing, public assistance, and debt collection, and they have no visibility into whether those systems are now being driven by upstream AI agents they will never see. The safety concern DeepMind is studying in the lab is a street-level reality for the people with the least power to contest it. The gap between "emerging frontier risk" and "thing happening to my client right now" is the gap progressives are built to close.

What you can do

Book a 60-minute call between your tech-policy people and your direct-services people. Ask the direct-services team to name two or three automated systems their clients already run into, then check whether any of those have quietly gained an AI layer. That inventory will shape your advocacy priorities better than any white paper.

The Roosevelt Institute wrote its own AI rules, in public

The Roosevelt Institute, one of the most credible progressive economic policy shops in the country, published a public statement defining how it uses AI internally, setting norms for transparency, oversight, and intellectual integrity in its research and communications.

Plenty of nonprofit and foundation staff are getting quiet pressure from boards and donors to either adopt AI fast or avoid it entirely, with no framework for navigating either. Roosevelt modeled the third option: make a considered, public commitment instead of drifting into incoherence. The point is not that they use AI. It is that they decided how, said so out loud, and gave everyone else something to adapt.

What you can do

Read the statement, then block a couple of hours this month to draft a one-page internal AI-use policy for your own org. Even a rough draft circulated to staff does more to build trust and cut anxiety than any all-hands meeting about AI. Roosevelt's framing is a clean template to start from.

One to read: "The Anti-AI Trap"

Writing on his Doomscrolling Babel newsletter, Manoel Horta Ribeiro makes an argument every progressive who is sick of AI hype should sit with. Reflexively dismissing AI as a scam or a fad is its own kind of trap. You cannot reshape a technology you refuse to take seriously, and mockery is not a strategy. It is the case this newsletter has been making since issue one, argued from the resistance side rather than the enthusiast side, which makes it more persuasive. Ribeiro is sympathetic to people worried about AI. His point is that the worry only translates into power if you engage with what the technology actually does, instead of waiting for it to go away. It will not.

What you can do

Send it to the colleague who keeps telling you AI is all smoke. Not to win an argument, but because the people most skeptical of AI are often the ones whose voices we most need inside the rooms where it gets governed.

Put AI to Work

Practical ways progressives can use AI this week

Open your services to people you couldn't reach last month

Real-time translation finally works well enough to use. Low latency, conversational, good enough for a real exchange rather than a clumsy back-and-forth. For any org that serves people across a language barrier, this is one of the rare AI capabilities with immediate value and almost no prerequisites. No procurement, no infrastructure, no technical staff. Just a phone. The Neuron has a good rundown of where it stands.

Intake and front-desk conversations. If someone walks in or calls and you do not have a staffer who speaks their language, real-time translation lets you actually help them now instead of scheduling a callback that may never get returned. Test it with a trusted bilingual volunteer first so you know where it stumbles.

Community meetings and canvassing. Multilingual organizing has always been bottlenecked on who you can pay to interpret. This does not replace a skilled human interpreter for high-stakes settings like legal or medical work, and you should not pretend it does. But for a neighborhood meeting or a door conversation that otherwise would not happen at all, it opens doors that were closed.

Know-your-rights and benefits help. Pair translation with the plain-language explainer workflow from earlier issues: get the accurate English source, have AI render it simply, then deliver it in the language the person in front of you actually speaks.

The honest caveat: this is a tool for expanding reach you did not have, not a reason to stop paying professional interpreters where accuracy is non-negotiable. Used that way, it is one of the most directly useful things AI can do for the communities we serve.

From our friends

Your org deserves its own AI. Not Big Tech's.

Change Agent is a private AI platform built for nonprofits, unions, and advocacy orgs. Your data stays yours, it plugs into tools you already use (Google Drive, Slack, ActBlue), and it handles the tedious stuff so your team can focus on the mission. Starts at $35/month. Small nonprofits under $1M can apply for discounted pricing.

Learn more

Looking Ahead

Next week we're going deep on data centers.

The fight over where AI actually gets built is heating up, and it doesn't break along the usual party lines. Bernie Sanders and AOC have a bill, the AI Data Center Moratorium Act, that would pause new large data centers until Congress puts real protections in place for the workers, communities, and environment these projects affect. And AOC just sat down with Republican voters in a deep-red data center town to make that case to their faces. It's worth fifteen minutes of your time.

This is one of those rare AI fights where progressive priorities and local conservative anger point the same direction. Next week we'll get into what the bill actually does, where that coalition is real and where it isn't, and what your org can do while the politics are still in motion.

Until next time,
Jordan

Know someone who should read this?

Share it with anyone who thinks the only two options on AI are cheerlead or panic.