The AI that rejects you everywhere at once

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

In this issue

Stanford found AI hiring screeners reject Black and Asian applicants at higher rates, and when one vendor screens for many employers, a single biased model can follow you from job to job.
Hugging Face is hosting deepfake-nude tools aimed at real political figures, with women across the political spectrum the main targets, a direct hit on the threat that pushes women out of public life.
More than $27 million in AI PAC money, both industry and safety, flooded a single Democratic House primary. AI is now a live financial faction inside our primaries.
A roundup of free AI tools that run on your own computer, so the data never leaves your device, useful for legal aid, domestic violence, and immigration staff who can't use cloud AI.
Progressive AI Win: Stripe is funding a $500 million nonprofit, with Anthropic and the OpenAI Foundation backing it, to prevent colds and flu, including air cleaning in schools.
Put AI to Work: audit the AI in your own hiring stack and draft a hiring-AI policy in an afternoon.

Quick Take

We spend a lot of energy here on how AI gets built. This week is about something quieter and, honestly, scarier: how a single AI gets used by everyone at once.

Stanford researchers studied real hiring screeners and found they reject Black and Asian applicants at higher rates. That's bad on its own. But the part that should change how we think is what happens when one vendor's tool does the screening for many employers. Get filtered out once and the same model can keep saying no at the next employer that uses it. They call it an algorithmic monoculture, and it turns one biased decision into a door that stays closed.

Here's the hopeful half, because there always is one. The same independent research that exposed the bias also points to the fix. Researchers got access, measured the harm, and put a number on it. That's the lever. Audits, transparency, the right to see how the machine decided. Let's get into it.

40,000

That's how many more job applications would have moved forward if the hiring AI had recommended candidates at equal rates across racial groups, in a single study covering four million applications. Not a rounding error. Forty thousand real chances, lost not to a human's bad day but to one model applied at scale. Fix the model and you don't help one applicant, you help forty thousand. Stanford HAI has the study.

AI News Roundup

One algorithm, every rejection

What happened: Researchers at Stanford's Institute for Human-Centered AI ran the numbers on AI hiring screeners and found they discriminate by race. In the data they studied, 26 percent of Black applicants and 15 percent of Asian applicants ran into postings where the AI worked against their racial group. The study was large: 3.4 million people, four million applications, 1,700 job postings, 150 employers, across eleven industries. The team (Rishi Bommasani, Sarah Bana, Kathleen Creel, Dan Jurafsky, and Percy Liang) also flagged a second problem that's bigger than any one biased model. When a single vendor's screening tool is used across many employers, the same flawed judgment repeats everywhere it's deployed. They call it an algorithmic monoculture. One in ten applicants who applied to four jobs got rejected from all four.

Why this matters: A biased human hiring manager is a problem for the people who apply to that one company. A biased model used across an industry is a problem for all of them at once. We're used to thinking of discrimination as a series of separate decisions by separate people, where a no from one employer still leaves the next door open. Monoculture closes those doors together, and it does it quietly, behind a screen that looks neutral and efficient. Here's what gets me about this one: the harm was invisible until someone with access went looking. The most useful thing in the story isn't the bias, it's the fact that independent researchers could study the real systems and put a number on it. The progressive ask falls out of that directly. Hiring algorithms that decide millions of people's shot at a job should be open to that kind of scrutiny by default, not by accident.

What you can do

If your organization uses any AI in hiring, including resume screeners and ranking tools baked into your applicant tracking system, ask the vendor one question this week: has this been tested for disparate impact across race, and can you see the results. If they can't answer, that's your answer. And if you work in policy or organizing, "independent audit access for hiring algorithms" is a clean, concrete ask to put in front of legislators who want to do something on AI but don't know where to start. Source: Stanford HAI.

Hugging Face is hosting deepfake-nude tools aimed at public officials

What happened: A Transformer investigation found deepfake nudification tools hosted on Hugging Face, one of the largest open platforms for AI models, built to generate fake nude images of real political figures: sitting members of Congress, one of the country's top judges, a former Trump cabinet official, and a prominent woman in the White House. Transformer named Rep. Alexandria Ocasio-Cortez as one target and held back others who haven't spoken publicly about being targeted. The tools weren't age-gated or hidden, they sat in the open alongside the platform's legitimate research models. One striking pattern in the set: women were disproportionately represented among the targets.

Why this matters: Women in public life already face image-based harassment at a scale men in politics mostly don't, and tools like these industrialize it. The targets here run across the political spectrum, from AOC on the left to women close to the Republican Party, which tells you this isn't a partisan weapon, it's a gendered one. The chilling effect reaches women who haven't even decided to run yet. Groups that recruit and support women candidates, like EMILY's List, She Should Run, and state-level pipeline organizations, have direct standing here. And it points at a gap progressives can press on: platforms that host open models write acceptable-use policies, but whether they enforce them is hard to see from the outside. That's a fixable accountability hole, not a law of nature.

What you can do

If you fund or support candidates, especially women candidates, make sure your digital-safety person knows these tools exist and where. Then push the platforms where it counts: a short, specific coalition standard on how model repositories should handle identity-targeted tools gives reporters and policymakers something to point at the next time this surfaces. Source: Transformer News.

AI money is flooding Democratic primaries, and not all on one side

What happened: The Democratic primary for New York's 12th congressional district drew more than $27 million from AI-focused political action committees, a live test of how AI money behaves inside a progressive primary. It didn't line up the way you might expect. A pro-industry PAC, Leading the Future, spent about $8 million against candidate Alex Bores. Pro-AI-safety PACs spent more than twice that, over $19 million, on the other side. And the results were murky: Bores' opponent already had structural advantages, and it's genuinely unclear how much any of the spending moved voters.

Why this matters: The takeaway isn't "AI industry money is buying our nominees." It's messier and, honestly, more important: AI has become a live financial faction inside Democratic primaries, with industry money and safety money both pouring in, sometimes against each other. That lands on district organizers, small-dollar donor networks, and field staff who plan around money they can see coming. The good news is that this is knowable. Campaign finance filings are public. The orgs that map who's spending, and on which side, this early won't get blindsided in 2026.

What you can do

Ask someone on your team to pull the FEC filings for races you care about and flag which AI PACs are spending, and which side they're on. An afternoon of public-records work now beats getting blindsided mid-cycle. Source: Transformer News.

Private AI you can run on your own computer

What happened: Jeremy Caplan of Wonder Tools rounded up free AI tools that run locally, on your own machine, instead of in the cloud: Jan, LM Studio, Msty, and AnythingLLM. The pitch is simple. Because the model runs on your device, your conversations never get sent to a big tech company's servers, and several of them work fully offline.

Why this matters: A lot of the progressive workforce handles information they can't responsibly paste into ChatGPT, Claude, or Gemini: legal aid attorneys, domestic violence service coordinators, immigration case managers, harm-reduction outreach workers. Local AI changes that calculus, because the data stays on the machine. It isn't a compliance guarantee, and Caplan is clear these tools aren't a substitute for professional judgment, so treat them as a safer option for sensitive work, not a green light to feed in anything. But for staff who've been responsibly sitting AI out, a tool that doesn't phone home is a real opening, and these are exactly the workers least likely to have an IT department sorting it out for them.

What you can do

Forward Caplan's roundup to the direct-service orgs in your network with one line: here are AI tools that run on your own computer, so the data never leaves the building. Then offer to help one team set up a single tool. Source: Wonder Tools.

Briefly: AI starts programming biology

A cluster of new tools (NVIDIA's BioNeMo, Nabla Bio's JAM-2, and the Arc Institute's Proto) is converging into systems that can read, design, and test biological processes far faster than the old lab cycle. Worth watching for one progressive reason: private health systems are likely to adopt this years before public clinics, community health centers, and rural health districts can afford to, which would widen the diagnostic gap right where it already hurts most. File it under "get someone tracking this before it hits appropriations." Source: The Neuron.

Progressive AI Win

AI labs are helping fund the fight against the next outbreak

Stripe is funding a new $500 million nonprofit called Intercept, aimed at preventing the common cold and the flu, with Anthropic, the OpenAI Foundation, Flu Lab, and Bill Gates among the backers. The plan funds prevention that helps everyone and disproportionately helps the people who can least afford to get sick: vaccines, plus large-scale air cleaning for schools, offices, and other public spaces.

This is the version worth amplifying: half a billion dollars of real money going into public-health infrastructure, not a flashy demo. Progressives should treat it as a precedent. The next time we're pushing AI firms to invest in community benefit, whether that's housing, mental health, or environmental monitoring, this is the example to put on the table. You already did it for respiratory infections. Do it here too. Source: MIT Technology Review.

Put AI to Work

Practical ways progressives can use AI this week

Audit the AI in your own hiring

The lead story is about hiring algorithms gone wrong somewhere else. Here's how to make sure yours aren't, and how to turn that check into a policy your org can stand behind. Any HR lead or operations person can do this in an afternoon, no data scientist required.

1. Inventory what you're actually using (about 20 minutes). List every place AI touches your hiring: resume screeners, ranking or "match score" features in your applicant tracking system, AI-assisted interview tools, anything that filters or sorts candidates before a human sees them. A lot of this is on by default and nobody turned it on deliberately. Write it all down.

2. Pressure-test the vendors in Perplexity (about 20 minutes). For each tool, search for whether it's been independently tested for bias, any disparate-impact audits, and any enforcement actions or lawsuits. Ask Perplexity to give you sourced links, not a summary, so you can read the primary material yourself.

3. Draft a hiring-AI policy in Claude or ChatGPT (about 30 minutes). Feed it your inventory and what you found, then ask for a one-page internal policy covering: which AI tools are allowed in hiring, what bias testing a vendor must show before you use them, where a human must review before a candidate is rejected, and how a rejected applicant can ask for that review. Tell it to write a usable internal policy, not a press statement.

4. Make one change this week (about 15 minutes). Pick the single riskiest tool from step one, the one doing the most filtering with the least oversight, and either turn off the auto-reject or add a human checkpoint in front of it. Small, concrete, done.

The payoff: you've protected the people applying to you, and you've built something you can share. A progressive org that can hand a peer a real hiring-AI policy is modeling the accountability we're asking everyone else to adopt.

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

Three things to keep an eye on.

Watch for movement on hiring-algorithm transparency. The Stanford team's whole point is that the harm only became visible because researchers got access. Watch whether the EEOC, state attorneys general, or state legislators pick up audit access as a concrete ask. That's the fight this story sets up.

Watch which candidate-support orgs respond to the deepfake-tool problem. The platforms hosting these tools won't move on their own. The interesting question is who builds the first real accountability standard and gets it in front of reporters.

Watch the public-health funding model. One $500 million nonprofit is a nice headline. If a second or third public-good investment follows, it's a precedent progressives can build a real demand around.