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

Quick Take

We spend a lot of energy in this newsletter arguing for rules. This week is a reminder that the rules are the easy part.

Congress already requires every federal agency to publish a list of the AI systems it uses. Not a suggestion, a requirement. A reporter found one of those listings in July, asked the agency about it, and ten days later the listing was gone.

The good news is real, and there’s a lot of it. California’s labeling law came into force on Sunday, timed to Europe’s. A bill requiring human review before AI can fire someone faces its make-or-break committee this week. And a New Jersey law signed a couple of weeks ago went after the pricing software landlords use to move rents in the same direction at the same time. Three recent moves, all of them making AI more accountable without making it smaller.

Then there’s the one that was already on the books, and quietly stopped working. That’s the one I keep thinking about, because writing rules turns out to be the part we’re getting good at.

Let's get into it.

AI News Roundup

The disclosure was required. It disappeared anyway.

What happened: On July 10, two Christian Science Monitor reporters, Sarah Matusek and Aaron Glantz, found an entry in the Department of Health and Human Services’ public AI inventory. The system was called Unaccompanied Child Sponsor Identity Verification. It uses computer vision to help vet adults who apply to sponsor unaccompanied immigrant children in government custody.

HHS had classified it “high-impact,” which under Office of Management and Budget guidance means it serves “as a principal basis for decisions or actions that have a legal, material, binding, or significant effect on rights or safety.”

The reporters asked HHS about it the same day they found it. The agency didn’t respond. By July 20, the description was gone from the inventory.

Congress requires agencies to maintain a public inventory of their AI use cases, current and planned. Jeramie Scott of the Electronic Privacy Information Center put it bluntly: “It would appear that HHS is in violation of the law.” Deirdre Mulligan, a faculty director at UC Berkeley, made the narrower point that lands harder: “when a reporter asks about something and it’s removed, that’s problematic.”

Why this matters: I want to be careful about what the argument is here, because the obvious version isn’t the right one.

The obvious version is that the government shouldn’t use AI to vet the adults who want to sponsor these kids. I don’t think that’s clearly true. Somebody has to do that vetting, and if the alternative is a slower manual process, slower could mean more children spending more time in federal custody. That’s a real cost, and it would fall on the children.

The argument is narrower and stronger. This tool helps decide whether a child gets placed with a sponsor at all. Congress decided systems like that have to be listed publicly, and it was. Then a reporter asked a question and the listing came off the site. Whatever explains that, it isn’t more disclosure.

And the specific worry advocates have is worth understanding, because it explains why the transparency matters here more than in most places. If sponsor information collected for child-welfare purposes can be routed to immigration enforcement instead, then the sensible move for a qualified aunt or cousin without status is to not come forward at all. The kid stays in custody. And you would never see that cost anywhere in the tool’s own performance numbers, because the sponsor who didn’t apply isn’t in the data.

That’s the whole case for the inventory. Not that these systems are bad, but that you cannot ask the question until you know the system exists.

California’s labeling rules went live, on the same day as Europe’s next phase

What happened: On August 2, the California AI Transparency Act became operative. It’s SB 942, amended last October by AB 853, and legal trackers read that amendment as having moved the start date to line up with Europe: August 2 is also the day the remainder of the EU’s AI Act came into application, one narrow exception aside.

If you run a generative AI system that’s publicly accessible in California and has more than a million monthly visitors or users, you now have three obligations. The covered content is images, video, and audio; text is excluded, as are video games, television, streaming, and film. Within that scope, everything your system generates has to carry a hidden identifier recording “the name and version of the GenAI system used, and…the date the content was created or altered.” You have to offer users a way to add a visible AI-generated marking that is “permanent or extraordinarily difficult to remove.” And you have to publish a free detection tool that lets anyone check a file, a URL, or an API call to see whether your system made it.

Penalties are $5,000 per violation, and each day counts separately. There’s no private right of action, so only the state can enforce it.

Why this matters: The free detection tool is the part I’d underline for anyone doing communications or rapid response work.

Up to now, verifying a suspicious image meant reverse image search, reading shadows, and a lot of judgment. Starting this week, for content from the largest systems, there’s supposed to be a provider-backed check you can run. Not a read on the vibes, an actual answer from the company that made the thing. That’s a genuinely new capability for the people who need it most and can afford it least.

It’s also a good example of what we keep arguing for. This law doesn’t make less AI. It makes AI’s output legible, which is a precondition for trusting it at all. The companies still ship whatever they want. They just have to sign their work.

The obvious question is whether the detection tools actually work well and stay up, and whether “more than a million monthly users” leaves the systems doing the most damage below the line. Both worth watching. But the shape of this is right.

The No Robo Bosses Act reached the room where bills die

What happened: This is suspense-file week in Sacramento. California’s Senate Appropriations Committee ran its list on August 3, and the Assembly’s runs on August 5. That’s the last fiscal gate before floor votes, and the place where bills quietly stop existing. Roughly thirty AI bills were still alive going in, and anything that doesn’t clear can’t be revived this session.

On Wednesday’s Assembly list: SB 947, the No Robo Bosses Act, by Senator Jerry McNerney and sponsored by the California Federation of Labor Unions, AFL-CIO. It passed the Senate 29 to 9 in May and was re-referred to Assembly Appropriations on July 2, which is where it sits now.

The bill does three things. It bars employers from relying solely on an automated system to fire or discipline a worker. It requires human oversight and independent verification where AI assists those decisions. And it prohibits systems that use a worker’s personal information to predict their future behavior.

McNerney, who spent sixteen years in Congress and co-founded its Artificial Intelligence Caucus, framed the Senate vote this way: “The commonsense guardrails in SB 947 will ensure that California businesses do not rely entirely on robo bosses to fire or discipline workers.”

Why this matters: Read the actual mechanism, because it’s more interesting than the name.

The bill doesn’t stop employers from using AI in personnel decisions. It says a human has to be in the loop before the consequence lands, and it says you can’t use someone’s data to forecast what they’re going to do. That first part is a procedural right, the right to have a person look at your case. Progressives have spent a century winning rights shaped exactly like that, in labor law, in benefits determinations, in immigration proceedings. This is the same right, applied to a new decision-maker.

The prediction ban is the part I’d watch spread. There’s a meaningful difference between an algorithm summarizing what you did and an algorithm estimating what you might do. The second one is where workplace surveillance has been heading, and it’s much harder to appeal, because there’s no incident to dispute.

Third bill this year we’ve covered that lands in this territory, and the pattern across all three is the same: not whether the software runs, but who is accountable for what it concludes.

Briefly — New Jersey went after the software that sets your rent

Two weeks ago, on July 20, Governor Mikie Sherrill signed the Forbidding the Algorithmic Inflation of Rent Act. Coverage of the signing put New Jersey as the fourth state to regulate algorithmic rent-setting.

The definition in the bill is worth reading, because it’s precise in a way these laws often aren’t. It targets a “coordinator,” meaning anyone operating software that performs a “coordinating function,” meaning it collects “the competitively sensitive information of two or more rental property owners” and runs it through an algorithm “used to set or recommend rental prices, material lease terms, or occupancy levels.”

That’s the actual harm, stated exactly. Not “an algorithm set your rent,” which is legal and ordinary. It’s that your landlord and the landlord across the street fed their private numbers into the same system and got back prices that move together. We had a word for that before there was software involved.

Enforcement runs through the New Jersey Antitrust Act, which is the right home for it. Violations expose operators to civil penalties, injunctive relief, and treble damages. The Attorney General has to set up an online complaint database. The law takes effect July 1, 2027.

Source: Office of the Governor of New Jersey, 20 July 2026

Briefly — Google wrote a rulebook, and drew the line around itself

Writing in Tech Policy Press on August 3, Gregory Gondwe examined Google’s proposal for a Frontier AI Regulatory Organization, or FARO, a body that would oversee advanced models posing national security or public safety risks.

His critique isn’t that frontier risk doesn’t matter. It’s about what the boundary leaves outside it: discrimination through automated decisions in hiring, lending, and content ranking; harm distributed thinly enough that no single violator is traceable; and power imbalances in creator compensation and labor conditions.

On the copyright opt-outs the proposal offers, Gondwe is sharp about who can actually use a right like that: “A large media company may have lawyers, engineers, and bargaining power. A freelance journalist, community newspaper, local artist, or small African publisher may not know its work entered a training system.”

His recommendations are structural rather than prohibitive. A public-interest majority on the board rather than mixed representation. Independent funding. Published audit summaries. Limits on staff moving between the regulator and the industry it regulates. One line is worth keeping: “Expertise alone should not determine public policy.”

Source: Tech Policy Press, 3 August 2026

Put AI to Work

Practical ways progressives can use AI this week

Find out what AI your government is running

The lead story rests on a public dataset most people don’t know exists. Here’s how to read it.

Under Executive Order 13960 and OMB Memorandum M-25-21, federal agencies have to report the AI systems they use, and those reports get consolidated and published. The 2025 edition covers 3,611 individual use cases from 41 agencies, of which 445 are flagged high-impact.

Step 1: Open the inventory. It lives on GitHub at the OMB repository. The file you want is /Data/2025_individually_reported_AI_use_cases.csv. There’s an Excel version in the same folder if that’s easier. Note that the old ai.gov inventory page is a dead link now. The repository is the working copy.

Step 2: Filter to an agency you care about, then to high-impact. The data dictionary is in /Validation/data_dictionary.md. High-impact is the designation that means the system is a principal basis for decisions affecting someone’s rights or safety, so it’s where the consequential stuff is. For scale: Veterans Affairs alone lists 215 high-impact systems.

Step 3: Read the descriptions critically. You’re looking for a specific thing. Does this entry tell you what the system actually decides, or does it describe a capability? “Improves processing efficiency” tells you nothing about who gets denied. That gap is the story, when there is one.

Step 4: Check the agency’s own page too. Several publish their inventories directly, and it’s worth comparing what’s there against the roll-up: DHS, DOJ, Energy, FHFA, and the Federal Reserve.

Step 5: Use AI to do the reading. This is a large CSV and you don’t have to go line by line. Drop it into Claude or ChatGPT and ask it to summarize every high-impact system at one agency, then flag the ones whose descriptions don’t specify what decision the system informs. Check anything it surfaces against the raw row before you rely on it. But as a way to make 3,611 entries tractable in an afternoon, it works well.

Why this is worth an afternoon: if your organization does policy, oversight, or direct services in any federal domain, there is almost certainly a system in this file that touches your people. Knowing its name is what makes every subsequent question askable.

Source: OMB, 2025 Federal Agency AI Use Case Inventory

Looking Ahead

• Watch what comes out of California’s suspense hearings. Roughly thirty AI bills were in play across the August 3 Senate list and the August 5 Assembly list, SB 947 among the latter. Bills that don’t clear are done for the session. Worth checking on the ones you care about by the end of the week.
• New Jersey’s rent law has an eleven-month runway. The FAIR Act doesn’t take effect until July 1, 2027. Long lead times are when definitions get tested and workarounds get designed, so the interesting question is what the pricing vendors do between now and then.
• A new federal inventory edition will land eventually. The 2025 one is what’s public now, and it more than doubled the 2024 count. The useful thing about a list that gets reissued is that you can compare editions: which systems appeared, which quietly went away. That’s a job someone should be doing on purpose.

Until next time,
Jordan

Read past issues  ·  RSS  ·  Website