You're using it wrong.

Progressives for AI Issue #12 — You're using it wrong.

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

In this issue

Most professionals use AI like a vending machine. The ones who get real value out of it iterate.
Sora's dead. Not for democracy reasons, but the result is the same. We'll take it.
Four SaaS subscriptions you could cancel this month — and what I built instead of paying them.

Quick Take

Most professionals use AI the same way they use a vending machine. Type a prompt, take the answer, leave. If the answer is mediocre, they shrug and write the email themselves.

The pros getting real value out of AI work it like a thinking partner. They iterate. They tell it the specific context of their work. They push back when it's vague. They ask it to critique its own answer before they accept it.

This month, three teachers in the AFT showed exactly what that looks like: Gabriela Aguirre, a first-grade dual-language teacher in San Antonio, used ChatGPT, Gemini, and Claude to build bilingual English/Spanish flashcards for her students. AFT teachers in New York are building agentic tools to monitor IEP compliance. None of them took the first answer. All of them taught the AI their context. That's the difference.

Let's get into it.

92% / 7% / 81%

92% of nonprofits use AI. Only 7% report it's had a major impact. 81% use it solo, with no shared workflows. That's the vending-machine problem in a single stat — lots of usage, almost no leverage.

Source: Virtuous + Fundraising.AI, late 2025 survey of 346 nonprofits

Progressive AI Win — We'll take it.

OpenAI killed Sora — their text-to-video app, and a major source of AI-generated slop.

The consumer app shut down on April 26. The API closes September 24. For a tool that was supposed to define the AI video moment, the wind-down is fast and quiet.

Be honest about why: it wasn't democracy. Sora was burning roughly $1 million a day. Its user base had cratered to under 500,000. Sam Altman is redirecting the compute to compete with Claude Code. OpenAI didn't kill Sora to protect the 2026 election. They killed it because it was bleeding money.

But the result is real, and we'll take it.

NewsGuard tested Sora 2 against twenty known false claims. It generated convincing video for sixteen of them, including five Russian disinformation narratives. The watermark could be stripped in four minutes with a free tool. Public Citizen called the system out by name as a 2026 midterms threat back in November.

We already have three Republican deepfake ads in the cycle: Talarico in Texas, Ossoff in Georgia, Spanberger in Virginia. There's no federal law on this! (Twenty-eight states have something, but most of them only require disclosure.)

So one less easy-to-use political-deepfake tool walking into the next six months is a real win, even if it's an accidental one. The caveat: the tools that survived — Runway, Veo, Hedra — don't have meaningful election guardrails either. Sora is gone. The field still has no floor.

Take the win. Then keep pushing for regulation.

More wins this week

WGA's new contract landed Tuesday. The Writers Guild ratified its 2026 MBA on April 24 with 90.4% voting yes. On top of the AI guardrails from 2023, the new contract adds notification and compensation requirements when studios use writers' work to train AI systems. Big union, fresh win, clean precedent.

Washington passed the strongest worker AI law in the country. HB 1672 requires human oversight for employment decisions made with AI, bans emotion-prediction AI in the workplace, and carries $10,000 penalties per violation. Companion bills (HB 1770, HB 2225) added AI content disclosure requirements and chatbot safety rules.

Illinois HB 3773 is now in force. Employers using AI in hiring decisions have to give applicants notice. It's a precedent more states can build on.

Three wins, three weeks — advocacy works.

Put AI to Work and stop renting SaaS by the seat

I'm not a developer, but I've still built more than fifty tools with Claude Code over the last few months. Almost every one of them replaced something I used to pay a SaaS company for.

Anthropic just made this kind of work easier: Claude Cowork lives on the desktop, reads and edits your local files, and runs multi-step tasks across your apps. Available on every paid plan. Anthropic's own pitch is the right frame:

Most AI tools are built around the prompt. Claude Cowork is built around the outcome.

So the move is two-step.

Step 1. Set up Cowork (or Claude Code or ChatGPT Codex if you want to go even further). Then ask it how to set itself up to help your specific work. Tell it about your work and ask it to find ten things it could automate for you. Iterate on the one that surprised you.

Step 2. Once it's running, point it at one of these. Each one is a SaaS bill I'm not paying anymore.

You're paying for...	Roughly	What I built instead
Social media scheduler Buffer, Hootsuite, Sprout	$30–100/mo	An N8N + Claude pipeline that posts to Bluesky and Twitter, exactly the cadence I want. How I built it →
Task manager Asana, Todoist, Things	$5–15/mo per seat	Drift — built in one session, ADHD-friendly, lets me bounce between projects without losing track. /ai → Drift
Domain monitor DomainTools, RegistrarSafe	$10–200+/mo	Domain Locker — tracks expirations, DNS, SSL, and WHOIS for the 47 domains I manage across personal and client work. /ai → Domain Locker
Email/newsletter platform Mailchimp, Constant Contact	$30–300/mo	Sendy + custom templates running an online community after AWS SES denied production access. /ai → Email Infrastructure

The full list of fifty-something projects is at jordankrueger.com/ai. The point isn't to copy any one of them. It's that you're allowed to build the tool you actually need instead of paying someone else's monthly fee for 80% of what you wanted.

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

May 13 — Colorado AI Act enforcement deadline. The country's only law requiring algorithmic discrimination accountability is under DOJ attack. Worth watching.
September 24 — Sora's API officially closes. One less surface for political deepfakes. The other tools still need work.

If you start treating AI like a coworker instead of a vending machine, your output changes pretty quickly. Try it this week. Tell it what you actually do. Ask it what it could take off your plate. Iterate on the one that surprised you.