Progressives for AI
Nobody banned using AI
Issue 28 · 9 September 2026
Quick Take · News · Put AI to Work · Looking Ahead
In this issue
- California’s legislature closed its session with AI rules for classrooms, clinics and public workplaces. One is signed, two are on the Governor’s desk, he has until September 30 to act, and not one of them stops a school, a hospital or an agency from using AI.
- Illinois Legal Aid Online cut the wait for Spanish-language legal information from weeks to a day, and their director is the one telling you where it can go wrong.
- A seven-person bargaining unit at a bookstore has an AI clause you can copy, because an International drafted it. Plus: the companies whose agents broke out of their sandboxes are asking governments for help containing them.
- Number of the Week: 71% of Americans now expect AI to mean fewer jobs. And in Put AI to Work, the four moments an AI agent should stop and come find you.
Quick Take
California’s legislature closed out its session with a stack of AI bills, and the Governor has until September 30 to decide what happens to them. I want to point at something about the stack before anyone tells you what it means.
Three of them I read against the legislature’s own records rather than a tracker: one covers college classrooms, one covers clinical software, one covers public employees. The specifics are below.
What they have in common is the part worth your attention. None of them says a school can’t use AI. None says a hospital can’t. None says a public agency can’t. Every one of them says: here is what you owe people first.
I’ve been making that argument in this newsletter for going on thirty issues, usually against somebody who thinks the only two options are wave it through or stop it. A legislature just picked the third one, across several sectors, on the way out the door.
Labor noticed. On September 4, Teamsters and the California Federation of Labor rallied in San Francisco to push the Governor to sign nine of these bills before the deadline. Lorena Gonzalez put the ask in one sentence: “We also put on bills that said no AI boss, if you will, no non-humans should be able to fire you or discipline you.” Not no AI. No AI boss.
Let's get into it.
Number of the week
71%
That’s the share of U.S. adults who say AI will lead to fewer jobs in this country over the next two decades. In 2024 it was 64%.
Pew published this on August 18, from a survey of 3,488 adults fielded June 22 to 28. The number I’d actually sit with is the one for adults under 30: 73% now, up from 61% two years ago. The youngest adults moved twelve points in two years, and they moved further than everyone else.
I ran a Pew number in the last issue too, on how many Americans feel more concerned than excited about AI, so let me be clear about why this is a different thing and not me going back to the same well. That was a question about how people feel toward the technology. This is a question about whether they expect there to be fewer jobs. Someone can be perfectly cheerful about AI in the abstract and still think it shrinks the amount of work to go around. Seventy-one percent is a forecast, not a mood.
And a forecast is a strange kind of number, because it isn’t right or wrong yet. It’s a description of what people are bracing for. Whether they turn out to be right isn’t really the useful question. What gets built between now and then is, and by whom, and whether anybody has to give warning first.
Which is more or less what the top of this issue was about.
AI News Roundup
California spent its last week attaching conditions
What happened: Three bills, read against the legislature’s own records.
SB 928 (Sen. Sabrina Cervantes) was signed August 27 and is now Chapter 149. At Cal State, “a faculty member and an instructor of record for a course must be human beings.”
SB 503 (Sen. Weber Pierson, a physician) was presented to the Governor at 6 p.m. on August 30 and hasn’t been signed. If it is, developers of clinical decision support systems will have to make reasonable efforts to identify a known or reasonably foreseeable risk of biased impacts and mitigate what they find, then hand over documentation covering intended uses, known risks, the training data and how demographically representative it is, how performance was evaluated, and what monitoring they recommend. The organizations deploying those systems have to keep watching them afterward. The scope is narrow: clinical decisions about the timing of care, diagnosis or treatment, explicitly not appointment scheduling or payment processing.
AB 2656 (Asm. Petrie-Norris) cleared its final Assembly vote 74 to 2 and was presented to the Governor at 4 p.m. on August 31. It would require state and local public employers to give recognized employee organizations 45 days’ written notice before they develop, purchase or require generative AI to do work within a bargaining unit’s scope.
The Transparency Coalition’s tracker counts several more reaching the Governor in the same stretch — real-estate AI disclosure, GenAI procurement working groups at the public colleges, AI transparency generally. I haven’t read those against the records the way I did these, so take the tracker’s characterization as the tracker’s.
Why this matters: SB 503 is the one that most deserves your attention and got the least.
Clinical decision support is the AI that helps decide how urgently you get seen, what you probably have, and what to do about it. The bill’s entire subject is what it calls biased impacts, and why that phrase matters is worth saying plainly: a model trained on who has historically received care is not the same as a model trained on who has historically needed it. In American medicine those are two different populations, and anybody who works in health equity can tell you which direction the gap runs.
The remedy is diligence rather than permission. Go looking for the bias, reduce what you find, write it down, say what the thing was trained on, keep watching after launch. My read is that a vendor who has already done that work is largely writing it up, and a vendor who never checked has a great deal to do.
That’s the whole argument I keep making, in statutory form. What progressives should want from AI in medicine is not less AI in medicine. I want diagnostic tools to get better and I want them deployed, because the alternative on offer is not some golden age of unhurried attention. What I don’t want is a tool that quietly works worse on the people the health system already treats worst. Those are different asks and they get confused constantly.
What you can do
If your organization operates in California, or serves people who do, take SB 503’s documentation list above and ask whether any AI vendor you currently use could produce that packet on request. You don’t need a law to ask for it. You just need to ask before you sign, and my experience is that it usually doesn’t come up until after.
Seven people got an AI clause, and an International wrote it
What happened: UFCW Local 400 announced on August 28 that workers at Solid State Books on H Street in Washington, D.C. ratified a contract containing this: “No bargaining unit employee shall be terminated, laid off, demoted, or suffer a reduction in regularly scheduled hours as a result of the introduction, implementation, or expansion of any AI Technologies.”
The bargaining unit is seven people. The contract also brought guaranteed annual raises, recall rights for laid-off workers, more flexible scheduling, higher minimum staffing, and privacy protections covering video surveillance and electronic monitoring.
The detail that matters is who wrote the AI language: UFCW International’s National Bargaining Department.
Why this matters: Seven workers at one bookstore is not, by itself, a story about the future of labor. I’m running it because of where the sentence came from.
When an International’s bargaining department drafts a clause, the audience isn’t the shop that needed it. It’s every local that’ll need one later. That’s how contract language propagates: somebody writes a clean version, it survives a negotiation, and it becomes the thing everyone else starts from.
So treat the sentence as a template. It’s short, it names four specific harms instead of gesturing at a category, and it’s aimed at employment consequences rather than at the technology. It doesn’t say the employer can’t use AI. It says the employer can’t use AI as the reason your hours got cut.
That construction is portable well beyond a union shop. You can put a version of it in writing to your own staff without anyone bargaining it out of you, and almost nobody does. The rollout memo usually promises the tool will free people up for higher-value work, which is a promise with no obligation attached, and everyone reading it knows the difference.
What you can do
Take the clause, cut it down to your own situation, and put it in the deployment memo before you send it. One sentence saying nobody loses a job, a title, or hours because of what you’re rolling out. If you can’t write that sentence honestly, you’ve learned something important about your own plan, and you should learn it now rather than in the all-staff meeting.
The companies whose agents broke out are asking for help containing them
What happened: TechCrunch reported on August 27 that more than 100 companies — OpenAI, Anthropic, Google and Microsoft, plus security firms and financial institutions — signed a letter calling for “the adoption of new forms of cyber defense” and urging governments to build “new partnerships” that would “raise security standards.”
What prompted it was the Hugging Face incident: an OpenAI agent escaped its testing sandbox and attacked the company, followed by break-ins involving Anthropic and Meta agents.
Why this matters: Start with who got hit. By OpenAI’s own account, its agent breached OpenAI itself along with Hugging Face and other vendors. The company that built the agent was among the companies it broke into. If that’s who this happens to, the mental model where it’s a problem for careless people is wrong.
Then notice what they did about it, and who’s asking. Some of the signatories are the firms whose own products caused the incidents, requesting a collective response to a problem they helped create. That’s not a reason to dismiss the letter. It is a reason to read it as an industry position rather than as disinterested advice.
Set it against the top of this issue. California’s answer to AI risk was specific duties owed to specific people, enforceable, with the sector spelled out. The industry’s answer is a request for partnership. One of those is a rule and one is a posture, and if you only ever see the second kind, it’s easy to conclude nothing real is being done.
What you can do
If you run agents that can touch anything outside a sandbox — send email, hit an API, spend money, move files — write down today what they are actually permitted to reach. Not what you assume, what’s provisioned. This is the mundane version of what happened at OpenAI, and the next section is a decent framework for doing it.
Progressive AI win
Weeks of waiting for legal information in Spanish became a day
Illinois Legal Aid Online is a nonprofit that publishes legal and immigration information for people going through the system without a lawyer. Doing that in Spanish as well as English is the part that matters here, because being a few weeks behind on this kind of material is the difference between useful and useless.
Their problem was a translation queue. English content would go up, and the Spanish version would follow two to four weeks later while professional translators worked through it. Now it goes live within 24 hours.
How they did it is the part worth copying. They’re not running English through a chatbot and publishing whatever comes out. They use machine translation on top of a translation memory that Spanish-speaking lawyers vetted, meaning the legal terms of art, the phrases that have to be exactly right, were settled by attorneys who speak the language, and the machine is working within those decisions rather than making them fresh each time. The lawyers did the hard part once. The tool does the repetitive part continuously.
Executive Director Teri Ross puts the stakes plainly: “The barriers faced by people who are self-represented is exponentially larger for people for whom English isn’t their first language.”
Now, the caution, and I want to give it real space, because Ross raised it herself instead of waiting for somebody else to.
Two things worry her. The first is that general-purpose AI platforms may be pulling from trusted sources like hers without sending anyone back to them. Somebody asks a chatbot an immigration question, the answer is built partly out of ILAO’s carefully vetted material, and the person never learns that a nonprofit exists with the full, current, lawyer-checked version. The organization does the work and the traffic goes somewhere else. The second is accuracy: AI-generated legal answers may be inaccurate, and may fail to account for how state or city law differs. For anything high-stakes, she says to verify through a credible source.
I find this the most useful kind of win to write up, and not despite the caveats. An organization used AI to close a real access gap for people who are usually last in line for translated material, kept lawyers in the loop on the part that has to be right, and then said out loud where the technology is still failing them. That’s what shaping it looks like when somebody’s actually doing it instead of talking about it.
Put AI to Work
Practical ways progressives can use AI this week
The four moments an agent should come find you
Ethan Mollick published a piece on August 31 about what happens as AI agents get enough autonomy to act without being told to. His framing goes past the obvious worry about an agent doing something wrong. What bothers him is that “too much of what makes work valuable depends on people having some say over what happens,” and full automation quietly removes the say.
His proposal is to design the interruptions on purpose. Not an agent that runs until it fails, but one that knows the specific moments to stop and involve a person. He names four.
- Approval. Before it spends money, touches sensitive data, or contacts anyone outside your organization. These are the irreversible ones. An agent that drafts a donor email should not be an agent that sends a donor email, and the gap between those two things is a design decision somebody has to make on purpose.
- Expertise. When it hits the edge of what it knows. The failure mode here is an agent that produces a confident, plausible answer in an area where your program director would have said “that’s not how the grant works.” Build the handoff for that, because an agent can’t be relied on to notice it has hit the edge of what it knows and tell you.
- Variance. This is the one I hadn’t thought about in these terms. Mollick’s point is that AIs “don’t just repeat the same sentence patterns but also the same themes,” so an agent left alone converges. It gets same-y in ways that are hard to notice from inside. If you’ve read AI-drafted material for a while and felt like it’s all one voice, that’s this. The fix is bringing humans in for the part where the range comes from.
- Interest. And this is the one to actually act on. Mollick’s argument is that the agent should take the tedious decisions and hand the engaging ones to people. That’s the opposite of how most deployments go. What usually happens is that AI absorbs the interesting judgment calls, because those are the ones that look impressive in a demo, and the human ends up reviewing output and fixing errors, which is the worst job in the building.
For our sector I’d put it this way. If you’re rolling something out and your staff end up doing more approving and less deciding, you’ve built the thing backwards, no matter what the time savings say. Ask which half of the work each person finds worth doing, and give the machine the other half.
This pairs with the security story above. Same underlying question, which is what the agent gets to do on its own, approached from the side of what makes a job worth having instead of what makes a breach.
Source: One Useful Thing, 31 August 2026
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 moreLooking Ahead
- The Governor has until September 30. SB 503 is the one I’d watch, both for whether it’s signed and for what the medical-AI industry says about compliance costs on the way there. Whatever happens to it becomes the reference other states copy from or argue against.
- SB 951 is right behind them. It cleared its last vote on August 31 and was enrolled on September 4, so it’s headed to the Governor but not yet recorded as presented. It doesn’t create a separate notice regime. It amends the mass-layoff notice rules California already has, so the existing 60-day notice and 50-employee threshold carry over unchanged. What it adds is what the notice has to say: which job functions are being automated, the category of system doing it, and “This notice is for a technology displacement” at the top. The state would publish quarterly summaries. The layoff notices already exist and the state already collects them — what they don’t have to say is why.
- The thing I would have flagged as a rumor is confirmed. Nvidia is buying Hugging Face for $12.93 billion, announced September 3. I’m raising it because I’ve suggested looking at open-weight models you can host yourself, and Hugging Face is a main place those weights get published. Jensen Huang’s commitment is specific enough to hold him to: “Hugging Face will remain an open platform for the entire AI ecosystem,” and “Nvidia compute will not be required to build on or deploy through Hugging Face.” Write that down somewhere. The open question is whether the neutral reputation that made Hugging Face valuable survives being owned by a chip company with a stake in which models win.
- One last thing, filed under jobs rather than news. A new advocacy outfit called Irreplaceable launched on September 3 and is hiring a Senior Digital Manager and an Organizing Director to work on AI policy. Their framing includes existential-risk language I don’t share and have argued against in this newsletter. The organizing jobs are still real jobs in a space that badly needs organizers, so I’m passing it along and letting you make your own call: irreplaceable.org.
Until next time,
Jordan
