The AI safety floor is labor.

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

In this issue

600 Google employees signed a letter against the Pentagon AI deal. Their CEO went silent.
Maryland just became the first state to ban surveillance pricing in grocery stores. The fine is $10,000 a violation.
A Harvard study says AI beat ER doctors at diagnosis. The researchers say not so fast, and that answer is the whole point.
One small thing you can build this week to replace a SaaS tool you don't need.

Quick Take

The federal regulators didn't show up this year. The Pentagon picked its AI partners by who's loudest about being unregulated, and Anthropic — the company that's been most vocal about safety guardrails — got left out of the contracts.

Meanwhile, the actual AI floor in 2026 is being set somewhere else.

It's being set by 600 Google employees who signed a letter to Sundar Pichai on April 27 telling him the new Pentagon deal is, in their words, blatantly stupid. Pichai's only public activity that day was a tweet about Google Translate's 20th anniversary.

It's being set by the screenwriters who got AI clauses into their last contract. By the screen actors who require consent for digital replicas. By a Maryland law that just made it illegal to charge you a different price than the person behind you in line because of what your phone knows about you.

The regulators didn't do this. The workers and the states did.

Let's get into it.

$10,000

The first-violation penalty under Maryland's new Protection from Predatory Pricing Act, signed by Governor Wes Moore on April 28. Maryland is the first state in the country to ban grocery stores and delivery apps from using your personal data to charge you a higher price than the next shopper. Repeat violations: $25,000.

Source: Maryland HB 895, Office of the Attorney General

Google's own AI researchers are revolting

On April 27, more than 600 Google employees signed an internal letter to CEO Sundar Pichai opposing the company's new Pentagon AI contract. The deal lets the Department of Defense use Google's AI for "all lawful use." The employees note this is non-binding "should not" language rather than enforceable "shall not."

This isn't the first letter. Back in February, more than 100 DeepMind employees urged Jeff Dean, the head of Google's AI research, to refuse the deal outright. He didn't.

Some of the workers spoke on the record. Andreas Kirsch, a DeepMind research scientist:

"The contract includes some meaningless weasel words to allow for PR spin, but it seems so blatantly stupid... I do not understand how this is 'doing the right thing,' and I think this violates 'don't be evil' quite clearly."

Alex Turner, also a research scientist:

"If OpenAI offered a fig leaf, Google said 'imagine we offered a fig leaf.'"

Their objections are specific: the contract permits Google AI to be used for domestic mass surveillance and autonomous weapons systems. The "safeguards" are guidance, not requirements. And nobody at the company asked them.

Pichai's response: silence. The day the deal was reported, his only public activity was a tweet celebrating Google Translate's 20th anniversary.

Why this matters. This is what corporate accountability looks like when the regulators have stopped showing up. The federal AI procurement decisions in the last six months rewarded the loudest-on-safety company, Anthropic, by excluding it from major DoD contracts. They rewarded Google with a deal its own researchers think is indefensible. The check on that wasn't a federal regulator or a Senate hearing. It was 600 of the people who actually build the technology, signing their names to a letter.

That's the floor. Not a regulatory floor. A worker floor.

What you can do

If you work at a tech company, this is a reminder that internal organizing still works. Google's leadership had to respond, even if their response was a 20-year-old translation product. If you work in advocacy, support the tech worker organizing groups (Tech Workers Coalition, Coworker.org's tech campaigns) and keep an eye on procurement policy at every level of government. Procurement decisions are quietly shaping which AI companies thrive and which don't. Industrial policy by other means.

More wins this week

The Academy banned AI-generated actors and scripts from Oscar consideration. Performances must be "demonstrably performed by humans with their consent" and screenplays must be "human-authored." The Academy reserves the right to ask for verification. This comes after a year of high-profile AI revivals (Val Kilmer) and AI-generated "actors" like Tilly Norwood being shopped to studios. The institutional version of what unions like the WGA and SAG-AFTRA spent the last two years bargaining for.

Artisan, the AI startup whose entire ad campaign is "Stop hiring humans," got caught using a human's stolen art. The artist is KC Green, creator of the "This Is Fine" webcomic (the dog in the burning room). Artisan put the comic in one of their own ads without permission. Green's reaction:

"It's not anything I agreed to. It's been stolen like AI steals. Please vandalize it if and when you see it."

The company that wants every business to fire its workforce was happy to take a worker's labor for free. The hypocrisy is the entire story.

The harder question — what about when the AI is actually better?

Here's where the lead's argument gets harder.

A Harvard study published in Science this week found that OpenAI's o1 model outscored two doctors on diagnostic accuracy in real ER cases. The model was 67% accurate. The two physicians scored 55% and 50%. Sample size was 76 patients at Beth Israel.

The progressive instinct on a story like this is automatic: protect healthcare workers from being replaced by AI. But here's the thing: that's the same answer the doctors are giving themselves.

The lead author, Adam Rodman:

"There's no formal framework right now for accountability."

He explicitly said the study does NOT mean AI is ready for clinical decisions. He called for "urgent... prospective trials to evaluate these technologies in real-world patient care settings."

The emergency physician Kristen Panthagani, who wasn't part of the study, went further. She called the headline overhyped and pointed out two specific problems. First, the AI was compared against internal medicine doctors, not actual ER specialists. Second, initial ER triage isn't really about precise diagnoses anyway. It's about identifying life-threatening conditions fast.

So here's the thing to sit with. In the same week 600 Google researchers said not without us about military AI, the doctors are saying not without us about ER AI. The instinct is the same. The form it takes (peer-reviewed papers calling for trials and accountability frameworks) is what worker-led caution looks like in healthcare.

A useful progressive answer here probably isn't "ban it" or "deploy it." It's: back the workforce's call for prospective trials, accountability frameworks, and equitable access if and when the technology is ready. Push for healthcare AI to be treated like a public utility, built once and available everywhere, instead of letting it become another premium service that cuts the underserved emergency rooms last.

I don't have a tidy answer here. Neither do the doctors. That's actually the point.

Progressive AI Win

Maryland just banned surveillance pricing

On April 28, Governor Wes Moore signed the Protection from Predatory Pricing Act into law. Maryland is the first state in the country to do this.

Here's what the law actually does. It bans grocery stores over 15,000 square feet (plus the third-party delivery apps that work with them) from using your personal data to set a price just for you. No more "you bought baby formula three weeks in a row, here's a slightly higher diaper price." No more "your zip code says you're price-insensitive, here's the marked-up version."

What's still allowed: temporary promotions, loyalty programs, supply-and-demand pricing, perishability adjustments. Normal pricing. Just not personalized pricing based on what your phone is telling them about you.

Penalties: $10,000 per first violation, up to $25,000 per repeat. Enforced by the Maryland Attorney General's Consumer Protection Division. Effective October 1, 2026.

Two things worth noting. First, the bill text doesn't say "AI." It says personalized pricing using personal data, which is what AI-driven pricing actually is. The legal mechanism is broader than the AI conversation. Second, the practice it bans is pretty new in the food sector but already routine in airline pricing, ride-share, and online retail. Maryland is the precedent. If it survives industry challenges, expect California, Washington, and New York to follow.

Take the win. Then push your state.

Put AI to Work — Build the audit tool you keep wishing existed

Here's a small thing I built last week that any nonprofit comms or fundraising team could replicate.

A client needed to audit 58 historical donation pages for any that had promised donor premiums (stickers, totes, a tote bag at the $100 level, a face mask) as part of an internal accounting review. The platform exports the pages as a CSV. Reading them in a spreadsheet is unbearable. Reading them in the platform's UI requires 58 separate clicks.

I asked Claude to write a Python script that takes the CSV and renders it as a single browsable HTML page. First pass put everything in a flat list. Second pass refactored it: known-merch pages at the top with a gold badge, everything else underneath. I served it on the local network and reviewed it in a browser. Total build time: about 20 minutes.

That's the whole pattern. Pick the boring data task you keep putting off. It might be a donor list cleanup, a contract review, a content inventory, a vendor audit, a board roster reconciliation. Export it as a CSV. Tell Claude what you're looking for and how you want it highlighted. Ask for a one-file HTML render. Open it in your browser.

If your org is on Claude Cowork (Anthropic's desktop app, available on every paid plan), the workflow is the same. Cowork can read the CSV right off your Mac.

The point isn't that this particular tool is special. The point is the same one as the lead story. When the people doing the work decide what their tools should do, the tools fit the work. When somebody else decides — usually a SaaS sales team with a $30,000 minimum — the tool fits the sales pitch and you bend your work to match.

Worker-shaped tools are the floor. Build one this week.

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

October 1, 2026 — Maryland's Protection from Predatory Pricing Act takes effect. First test of how aggressively the AG enforces, and whether other states follow.
Summer 2026 — California's SB 947 (employment automated decision systems) and AB 2027 (worker data: AI prohibitions) both moved out of committee in late April. Floor votes likely this summer. If they pass, California is suddenly the worker-AI policy floor for the whole country.
Anthropic shipped its most powerful model yet — and you can't have it. Claude Mythos Preview, released to 11 enterprise security partners (including Google, Microsoft, Apple, JPMorganChase) plus 40+ critical-infrastructure orgs. Anthropic explicitly said they "do not plan to make Claude Mythos Preview generally available." Worth watching: this is a real shift toward concentrating frontier capability with enterprise security teams instead of releasing it broadly. Same procurement question as the lead story, just from the other side of the table.

The pattern is the same in every story this week. Workers, professionals, and states are doing the AI accountability work that the federal government isn't. They're doing it in letters, in collective bargaining, in peer-reviewed papers, and in state legislation.

If you work somewhere that uses AI, you're part of this. Sign the letter. Ask the question. Negotiate the clause. Build the small tool yourself instead of paying for the wrong one.