Progressives for AI
Congress is writing rules for the child-facing product
Issue 25 · 11 August 2026
Quick Take · News · Put AI to Work · Looking Ahead
In this issue
- Parents want useful AI for their children — with rules they can trust.
- Secret cyber tests still need public, democratic rules.
- If AI takes on more work, workers deserve a say in what happens next.
- A useful agent puts boundaries and a stop button in people’s hands.
Quick Take
Parents are not asking Congress to turn AI off.
In a new Common Sense Media survey, 63 percent of parents were interested in AI tools that help children learn. At the same time, 68 percent wanted strong laws requiring companies to make AI safe. Only 24 percent preferred trusting companies without government rules.
That is not a contradiction. Families want the good stuff — homework help, creativity, and accessible tutoring — with protective defaults: clear controls, less data collection, no targeted ads or emotional hooks, and a hard line on self-harm.
Last week, the Senate Commerce Committee advanced two bills aimed at those protections by voice vote. The package has real strengths: duties for companies, protective defaults, limits on manipulative design, room for stronger state laws, and no federal ID mandate.
It is not finished. During markup, senators removed an explicit anti-profiling rule and a private right to sue. They also let parents authorize indefinite chatbot memory. And the CHATBOT Act still makes every teen account depend on parental consent — a barrier for young people with unsafe, unsupportive, or unavailable parents.
A progressive bill should put the burden of safe design on companies, give families real control, and protect a teenager’s path to confidential human help. It should not be that hard. Let useful AI earn trust by meeting the standard.
Source: Common Sense Media, 9 March 2026 · U.S. Senate Commerce Committee, 5 August 2026 · EPIC, 5 August 2026
Let's get into it.
Number of the week
20%
One in five employed U.S. adults said AI now does most or all of at least one task they once handed to a coworker or contractor. This is self-reported task handoff, not a count of lost jobs. But productivity gains do not distribute themselves. Workers should help decide how jobs change — and share the payoff through better work, better pay, or more time of their own.
AI News Roundup
Secret benchmarks do not require secret rules
What happened: The White House wants to know whether frontier AI models can carry out serious cyberattacks. Fair enough. Executive Order 14409 calls for classified testing and a voluntary framework that lets developers give the government up to thirty days of early access. It explicitly rejects mandatory licensing or preclearance.
But the rules around that process remain murky. Michelle De Mooy has five basic questions: Which models qualify? How long can review take? Who gets early access? Can developers appeal? What will the public learn afterward?
Why this matters: Security may require a secret test. It does not require a secret process. Clear criteria, deadlines, and appeal rights are basic democratic accountability. They keep national-security power from becoming informal leverage that only the biggest companies know how to navigate.
What you can do
When officials invoke security, ask what specifically must stay secret. Protect the test. Publish the rulebook.
The model you audit is not always the model people run
What happened: A safety team tests the full-sized model. The version on your laptop may be compressed to make it faster and cheaper. Same family, different build — and sometimes different behavior.
In a new preprint, Emilio Ferrara and his coauthors found that familiar refusal and multiple-choice tests looked reassuring across compressed models. Longer, open-ended conversations did not. An independent judge flagged stereotypes in roughly 24 to 27 percent of answers across the tested versions and eight languages.
Why this matters: Do not read 24 to 27 percent as proof that compression causes bias everywhere; this is a preprint centered on two models. The stronger finding is simpler: short tests missed behavior that longer conversations exposed. Affordable, local AI can spread access beyond the largest companies and wealthiest users. That promise only works if the safety standard travels with the model. Keep compression — and test the build that ships.
What you can do
Add one line to your evaluation plan: name the exact model and configuration in production. If the vendor cannot do that, keep testing.
Briefly — workplace AI needs due process
California’s SB 947 heads into Thursday’s Assembly Appropriations suspense hearing. The bill targets automated employment decisions — the systems that can help decide who gets hired, promoted, disciplined, or fired.
This is where progressive AI policy belongs: tell workers when automation shaped a high-stakes decision, give them access to the data used, require independent human review, and preserve a path to challenge the result. The goal is accountable workplace technology, not a ban on useful software.
Source: California Legislative Information, accessed 11 August 2026
Briefly — Claude Code is changing the default
Starting Friday, new Claude Code sessions on Pro, Max, and Team plans will default to auto mode. Pinned defaults stay pinned, and Enterprise remains opt-in for now.
The reason is more interesting than the setting. In Anthropic’s test, humans stopped just 13.6 percent of planted dangerous commands. Auto mode stopped 89 percent. It still missed 11 percent, so this is not perfect automation. It is evidence that a company cannot dump safety work onto people through an endless series of approval boxes and then call that human oversight.
Source: Anthropic, 7 August 2026
Progressive AI win
Organized labor got a seat at the infrastructure table
AI infrastructure is going to be built. The better fight is over who gets the work and whether it leads to a real career.
This week, North America’s Building Trades Unions, BlackRock, and the AI Infrastructure Partnership agreed to plan for apprenticeships, training, recruitment, and project labor agreements around new data centers and energy projects. Labor is in the room before the money lands. That matters.
The agreement is nonbinding, so nobody has won prevailing wages or union jobs yet. The next test is what appears in the contracts.
Put AI to Work
Practical ways progressives can use AI this week
Keep the stop button in human hands
A good agent should give people more control over their work, not quietly take control away. That starts with knowing when it is finished — and when a person must take over.
Pick one repeatable task and write four things:
- The output: What should exist when the work is done?
- The evidence: How can the agent prove it worked?
- The boundary: What always requires your approval?
- The exit: When should it retry, stop, or ask for help?
Then test one normal case and one failure. That is a worker-controlled agent loop: plan → act → verify → retry or escalate → done.
Autonomy should remove drudgery without removing human authority. Keep the boundaries, evidence, and stop button in the hands of the people responsible for the work.
Source: The Neuron, 7 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
- Thursday: California Appropriations takes up SB 947 and other AI bills. Watch whether lawmakers preserve worker protections when the bills face fiscal triage.
- Friday: Claude Code switches new eligible sessions to auto mode. The test is whether automation makes human judgment more meaningful — or simply moves responsibility away from the company.
Until next time,
Jordan
