Progressives for AI
California made checking AI a real job
Issue 29 · 15 September 2026
Quick Take · News · Put AI to Work · Looking Ahead
In this issue
- California is turning AI auditing into a state-registered profession. From January 1, 2029, offering a covered AI audit in the state will require being on a public register and meeting standards for independence.
- A New York Fed survey found 61% of service firms now using AI, and among the firms using it, 4% reported laying anyone off because of it.
- A federal judge refused Elon Musk’s request to pause Minnesota’s law against AI nudification apps, so it stays in force. The law protects adults, not just minors, which most of the coverage got wrong.
- Put AI to Work: have AI read back your own email thread and tell you what you already promised.
Quick Take
California signed two AI bills on September 9. Together they turn auditing AI into a recognized occupation with a state register behind it.
Under SB 813, the Government Operations Agency has to build a framework for designating what the statute calls independent verification organizations, outfits qualified to assess AI systems for compliance with state law. Under AB 1405, auditors go on a public state registry and have to meet standards to stay on it. They have to be independent of the company they’re auditing, and they can’t go looking for a job with that client while the audit is running. The Governor’s own release describes the pair as establishing “a framework for independent third-party evaluation and audits.”
People already audit AI systems, at universities, at a handful of consultancies, at a few nonprofits, and both statutes lean on standards that national, international and professional bodies have already written. What California didn’t have was its own public list of who does this work and its own criteria for designating them. That’s what these two bills build.
I’ve spent a lot of issues arguing that the useful question about AI regulation isn’t how much, it’s what kind. This is a kind I didn’t expect to see first.
Let's get into it.
Number of the week
4%
That’s the share of AI-using service firms that told the New York Fed they had laid anyone off in response to AI over the past six months. The denominator matters, so I’ll be precise: that 4% is out of the firms actually using the technology, not out of every firm surveyed.
Adoption is the other half. 61% of all service firms reported using AI this year, up from 40% last year and 25% in 2024. Manufacturers went from 26% to 51%, and reported no AI layoffs at all.
The Fed’s researchers, Jaison Abel, Richard Deitz, Natalia Emanuel and Nick Montalbano, published this on September 1, drawing on the Empire State Manufacturing Survey and the Business Leaders Survey, with the AI questions fielded each August since 2024. Two caveats I want on the record. This covers the Fed’s Second District, which is New York, northern New Jersey and southwestern Connecticut, so it’s regional rather than national. And the post doesn’t publish the number of firms that answered, so I can’t give you one.
What the firms did instead of firing people: just over a third of the service firms retrained workers. About 15% slowed down hiring. About 13% hired more.
The authors’ own summary is blunter than anything I would write: firms “are continuing to adapt and change due to AI use, but not by eliminating vast numbers of jobs. Instead, this new technology is reshaping work itself, with firms investing in their existing workforces rather than replacing large swaths of people.”
I ran a Pew number last issue on how many Americans expect AI to mean fewer jobs, so let me be clear about why this isn’t the same well twice. That was a forecast, what people think is coming. This is a count of what firms say they have already done. The gap between 71% of the public expecting fewer jobs and 4% of AI-using service firms reporting a layoff isn’t proof that the expectation is wrong. It’s a measurement of where we actually are, which is a different thing and much harder to come by.
Source: Federal Reserve Bank of New York, Liberty Street Economics, 1 September 2026
AI News Roundup
California is building the job of checking AI
What happened: Governor Newsom signed two AI bills on September 9. Read together, they give AI auditing a state framework and a public register.
SB 813 (Sen. McNerney), now Chapter 179, puts the work with the Government Operations Agency. By January 1, 2028 the agency has to develop and publish the criteria for designating what the statute calls independent verification organizations, outfits qualified to assess AI systems for compliance with state law, along with the application requirements and the procedures for suspending or terminating a designation. SB 813 also requires a disclaimer that publishing those requirements and criteria is not a state endorsement of any AI system or model. AB 1405 carries its own, separate disclaimer: an auditor’s registration is not the state endorsing that auditor.
AB 1405 (Asm. Bauer-Kahan), now Chapter 178, builds the register. By January 1, 2029 the Government Operations Agency must run a public AI Auditor Registry, and the statute sets the terms of staying on it. An auditor “shall not conduct a covered AI audit if the auditor has a financial, business, employment, or other interest or relationship that would reasonably be expected to impair the auditor’s independence or objectivity.” They cannot “seek, solicit, negotiate for, or accept employment with the auditee while participating in the audit.” They must “exercise independent, objective, and impartial professional judgment,” and they cannot audit their own prior work. Every audit produces a signed report to the client covering scope, results and limitations.
One line in AB 1405 does more work than the rest: once the chapter takes effect on January 1, 2029, people who are not on the register will not be permitted to offer covered AI audit services. That’s what separates a register from a directory.
Now the part you should know before anyone sells this to you as tougher than it is. SB 813 says in plain words that the chapter does not “require any person, partnership, or corporation that develops, deploys, or operates an AI system or model to engage an IVO or to undergo a covered AI audit as a condition of developing, deploying, or operating an AI system.” Developers decide whether to be audited. What the statute adds is a consequence in court: if you’re sued over harm, the fact that an audit was done to an identified standard is “relevant to, but not conclusive of, the action.”
Why this matters: People audit AI systems today, and both statutes explicitly reference existing national, international, professional and industry standards rather than inventing a method from scratch. What California didn’t have was a public register of who does this work, state criteria for designating them, and a rule keeping the auditor at arm’s length from the company paying the bill. Bauer-Kahan’s line in the Governor’s release is the short version: “third-party auditors are essential to ensuring AI is safe for our communities.”
The pressure on developers here is indirect. It runs through liability. An audit is evidence you can put in front of a judge, and its absence is a thing a plaintiff’s lawyer will notice. That’s weaker than a requirement. It also means the first companies to get audited may be the ones with the most legal exposure, which isn’t the same as the ones doing the most harm.
But look at what this builds rather than what it commands. Every regulation that arrives later, in California or anywhere copying it, needs somebody qualified to check compliance, and that capacity has to exist before the rule does. Right now the people doing this work in California are scattered across universities, consultancies and a few nonprofits, with no public list and no state criteria a buyer can check them against. A register with independence rules is what turns that into a recognized field with a route into it.
How much work it actually becomes depends on how much demand there is for audits, and the statute doesn’t create that demand. But the route now exists, and it leads to a job that consists of holding powerful companies to their claims. Progressives should want a great many more of those.
What you can do
If your organization buys or deploys AI, start asking vendors now whether they have had an independent audit, who did it, and whether you can see the report. You don’t need the register to exist to ask, and asking before you sign is the only time the question has any leverage. If you work in evaluation, data science, civil-rights compliance or accessibility testing, the registry has to be established by January 1, 2029, and that’s the date the restriction on unregistered audits begins. A little over two years to decide whether this is your next move.
Briefly — the Justice Department filed against Minnesota
On August 18 the Department of Justice filed a statement of interest supporting xAI’s challenge to Minnesota’s AI nudification law, arguing the state statute “goes further than existing federal restrictions” and warning that states “must be careful not to impose excessive measures that hinder American national and economic security.” Notably, DOJ stopped short of asking the court to grant xAI’s injunction. The judge ruled against xAI anyway, which is the next item.
Source: Gizmodo, 20 August 2026
Progressive AI win
Minnesota’s law is still standing
On September 4, Governor Tim Walz posted four words about a court ruling: “Good news: he lost.”
The “he” is Elon Musk. His company xAI has been trying since July to block a Minnesota law. Both forms of interim relief it asked for in that motion were refused: the temporary restraining order on July 31, and the preliminary injunction on September 4, denied by Judge Donovan Frank. One lawsuit, and the law has stayed in force throughout.
That’s a real win. It’s also a law that a lot of people have been told the wrong thing about.
What the law actually covers. HF 1606, signed May 7 and effective August 1, makes it illegal for apps, sites and programs to let users “nudify” an identifiable person, meaning altering or generating an image to show an “intimate part” that wasn’t in the original. The summary Walz was replying to said the law was about protecting minors. There is no age limit anywhere in the definition. Sen. Erin Maye Quade carried the bill after adult women came to her whose photographs had been used to make AI nudes of them. The part that protects them is exactly the part the summary dropped.
Two more things in the text. It exempts tools that require “the technical skill of a user,” which the statute defines as real human skill and judgment, so the target is services that do the work for you. And enforcement runs two ways: a victim can sue for up to treble compensatory damages plus punitive damages and fees, and the attorney general can separately seek up to $500,000 per unlawful access, download or use, with those penalties funding victim services grants.
Why the judge ruled this way. Frank found xAI hadn’t shown irreparable harm, and noted the company waited roughly three months after the bill was signed before suing. He weighed that against Minnesota’s interest in a law its legislature passed “democratically and nearly unanimously.” The case isn’t over. A motion to dismiss is still pending, and xAI has appealed to the Eighth Circuit.
The critique worth hearing. The argument against this law isn’t only Musk’s. Mike Masnick at Techdirt and Elizabeth Nolan Brown at Reason both make the case that the drafting sweeps too wide, that the definitions could reach political parody, images people made of themselves, or ordinary photos of people in swimwear. I’m not persuaded that sinks the statute, but they’re arguing in good faith about real text, and I’d rather you hear it from me than find it later and wonder what else I left out.
So here’s the whole thing. A state passed a law against a genuine harm, nearly unanimously. A very well-resourced company sued to stop it and has twice failed to get it paused, with the Justice Department backing the company before the second ruling. That’s the system working. The work left over is the drafting, and you can’t improve drafting nobody has read.
Sources: Minnesota Session Laws 2026, ch. 72; MPR News, 4 September 2026
Put AI to Work
Practical ways progressives can use AI this week
Ask your own email what you already promised
Last week I was turning a client proposal into a signed contract scope. Before I finalized it, I had Claude read back through the email thread with that client, 21 messages over several weeks, with one question: what did I commit to in here that isn’t in this document?
It came back with three. A 24-hour response time I’d offered in August. Uncapped maintenance hours, in my own words, “whatever it takes to make sure it’s working perfectly.” And a 90-day quiet period I agreed to after the client countered my 60. All real, all mine, and none of them in the scope I was about to sign.
This is the most useful thing I do with AI that almost nobody talks about. Your organization’s email is the record of every promise anyone made to a funder, a vendor, a coalition partner or a member. Nobody can hold that in their head, and nobody has time to reread it. A search tool can.
Try it this week. Before your next funder report, board meeting or contract signature, take the relevant email thread, paste it into whatever AI tool you already have, and ask that one question. Then check what comes back against the actual emails. It will miss things, and every so often it will invent one.
That last part isn’t a disclaimer, it’s the job. A silent 18-second video went through my own transcription setup last month and came back with a transcript. The model wrote words where there was no audio at all, and it read as success. The fix was to check the output against something independent rather than trusting that a result meant a real result. Same principle here: the AI finds candidates, you confirm them against the source.
Two limits on where you point this. Don’t paste member or client personal information into a general-purpose AI tool, which is a separate conversation with whoever owns your data policy. And if your email lives in Google Workspace or Microsoft 365, find out what your admin has already turned on before you sign up for anything new.
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
Newsom still has the rest of the enrolled AI package in front of him, and the deadline is September 30. SB 503 is the one I’m watching. That is the clinical decision support bill I wrote about last issue, which would require developers to look for biased impacts and document what their systems were trained on. SB 951, the amendment requiring layoff notices to say when automation is the cause, and AB 2656, the 45-day notice to public-sector unions, are both still unsigned as well. Two weeks left.
Colorado’s revised rules on automated decision systems are due to circulate by September 23. September 4 was the cutoff for comments to be reflected in that revised draft, but the comment period itself runs through October 26, so there’s still time to file if this is your area.
And xAI’s appeal to the Eighth Circuit is the one to keep an eye on past this issue. A district judge declining a preliminary injunction isn’t a ruling that the Minnesota law is constitutional. It’s a ruling that xAI didn’t clear the bar for stopping it while the case proceeds. The constitutional question is still open, and an appellate answer to it will matter well beyond Minnesota.
One more, filed under the lead story. California’s AI auditor registry doesn’t have to exist until January 1, 2029, and the designation criteria are due a year before that. That’s a long runway, and long runways are when the standards get written and when the people who’ll do this work decide whether to train for it. If that’s your field, this is the part of the process where showing up counts.
Until next time,
Jordan
Until next time,
Jordan
