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

Quick Take

Two of the most powerful people in AI spent last week on the same subject, from opposite ends of it.

Dario Amodei called it a crisis of trust, and said the most accurate criticism of AI companies, his own included, is that they haven’t delivered on what they promised. Sam Altman went on a podcast and said the field has done a poor job explaining what AI is good for. Pew, meanwhile, has 52% of Americans now more concerned than excited about AI in daily life, up from 37% five years ago.

Altman thinks this is a communications failure. I don’t. I think people are perfectly capable of hearing a promise and noticing it didn’t arrive, and no amount of better framing fixes that.

But here’s the part I keep turning over. While the labs were workshopping their message, a lot of other people spent the week quietly writing down terms they could actually live with: a bargaining committee in Maine, a routing layer at a phone company, a redaction tool that costs twenty-nine dollars once, a social worker asking permission before starting the recording.

That’s what closes a trust gap. Not a better story about the technology, but better terms attached to it.

Let's get into it.

AI News Roundup

The trust problem got named by the people who caused it

What happened: TechCrunch reported on August 19 that Pew found 52% of Americans more concerned than excited about the increased use of AI in daily life, up from 37% in 2021. The piece ran the polling alongside remarks from two tech executives, one of whom runs a frontier AI lab.

Anthropic’s Dario Amodei described a “crisis of trust,” and said this: “I think by far the most accurate criticism of AI companies, including Anthropic, is that we haven’t yet delivered on our big promises to benefit the world.”

Airbnb’s Brian Chesky split the difference. “I think part of it’s a narrative issue that we’re not talking about AI correctly. But part of it is we need to actually be developing more products that just regular people can use and say, ‘I love AI because AI allows me to have a doctor on demand.’”

And on August 23, The Neuron wrote up Sam Altman’s version, delivered on David Senra’s podcast. Altman’s diagnosis is that the field has “not as a field done a very good job” explaining the benefits, and that the right framing is one about “more power and personal freedom,” with AI setting off “the greatest boom in people starting smaller businesses that we have ever seen.”

Why this matters: I want to give Altman partial credit and then take most of it back.

The partial credit is real. I’ve thought for a while that the doom framing was a strategic disaster for everyone, including the people making the argument in good faith, and Issue 22 made a version of the case: a rulebook written around extinction risk is cheap for a lab to accept and does nothing about the harms readers actually face. When your loudest public message is that your product might end civilization, you have told the public two things at once: that the thing is dangerous, and that you’re building it anyway. Nobody hears that and concludes you’re trustworthy. Altman is right that the field talked itself into a corner.

Where he loses me is the implied fix. If the problem is that people misunderstood, then the solution is better explaining, and better explaining is the cheapest remedy available. It costs a comms budget and nothing else. Nobody has to change what they’re building, or who they’re building it for.

Amodei’s version is harder and I think it’s the correct one. The problem isn’t that people misheard the promise. It’s that they heard it fine and it hasn’t shown up. Chesky’s example is the tell: a doctor on demand. That’s the thing people actually want, and it is conspicuously not what got built. What got built was a lot of chat interfaces, a lot of enterprise pilots, and an enormous amount of infrastructure spending. Some of that will eventually turn into the doctor. None of it has yet.

So I’d put it this way. Trust isn’t a story you tell about a technology. It’s a set of terms a person can point at: who can see my data, who’s accountable when the output is wrong, what happens to my job, and whether anyone asked me first.

Every remaining story in this issue is somebody answering one of those questions in writing, which is why I’m not especially gloomy about a 52% number. Concern isn’t rejection. Concern is what people feel when they can see the upside and don’t yet trust the terms. Terms are fixable.

The AI rules that actually bind are being written at bargaining tables

What happened: Gleb Tsipursky argued in Allwork.Space on August 21 that the enforceable rules governing workplace AI in the United States are not coming from Congress. They’re coming from union contracts, one at a time. Citing an Axios report from July, he puts the NewsGuild-CWA alone at roughly 85 to 90 contracts with explicit AI provisions.

Two examples in the piece are worth sitting with. At Microsoft’s ZeniMax studios, the contract requires management to notify the union and bargain before introducing certain AI systems, framing AI as something that should support workers rather than replace them. At Politico, the Washington-Baltimore News Guild challenged AI tools that were producing inaccurate material. An arbitrator found management had violated the collective bargaining agreement, and Politico later dismantled the tools.

And on August 23, nearly 4,500 workers at Bath Iron Works represented by IAM Local S6 ratified a five-year agreement that, alongside a 17% first-year raise and five years of locked healthcare costs, includes what the reporting describes as “stronger protections from new technology, artificial intelligence, and robotics.” The contract language isn’t public, so that phrase is all any of us know. But note where it is. This is a Navy shipbuilder in Maine, not a newsroom and not a tech company.

Why this matters: Look at the Politico case again. It’s the argument in one story.

A tool was deployed. It produced bad work. Somebody with standing filed a grievance, an arbitrator agreed, and the tool came out. No new statute was required. No agency rulemaking, no floor vote, no waiting for the next session. The mechanism was already sitting there in a document both sides had signed.

That’s a very different theory of AI governance than the one most of us have been operating under, where the good outcome arrives as legislation and everything before that is just waiting. And I’d argue the contract route has a property the legislative route doesn’t: it produces a specific, enforceable answer at a specific workplace, written by the people who have to live with it, and it can be enforced by somebody whose job is enforcing it.

The obvious limit is just as real. This route only reaches people who have a union, and the Bureau of Labor Statistics put the 2025 membership rate at 10.0% of wage and salary workers, about 14.7 million people. So most American workers can’t get an AI clause this way at all. Tsipursky’s argument is that the contracts are worth reading anyway, because the same five disciplines transfer to a workplace with no union at all: give advance notice before a material deployment, put frontline workers in the design group, write down which uses are off limits and which require a human sign-off, connect any productivity gain to an actual workforce plan, and build a channel where an employee can raise a problem and get an answer.

Read that list and notice something. Not one of those five is a restriction on the technology. They’re all restrictions on how it gets introduced. Which is the same thing the lead story is about: the objection was never really to the tool.

AT&T is reportedly running 40% of its internal AI on models it can host itself

What happened: The Information reported on August 20 that AT&T now routes about 40% of its employees’ AI queries to open models rather than to Anthropic or OpenAI, with a stated goal of pushing that to 60 or 70%. On coding and some other advanced tasks, the reported result was a 56% cost reduction with roughly a 2% decline in quality. Mark Austin, the AT&T vice president who oversees AI for employees, is quoted in PYMNTS’s write-up of the report saying open-source models are “just as good or better” than older models from Anthropic and OpenAI. The reporting separately puts today’s open models about six to ten months behind today’s frontier ones. Those two statements fit together: the open model you can download today is roughly where the paid ones were last winter, and last winter’s models were already useful.

One sourcing note before we go further. I have not read The Information’s piece, which is paywalled. What I’ve read is PYMNTS and half a dozen other outlets relaying it, which sounds like corroboration and isn’t. Many articles, one source. Treat the numbers as reported rather than established.

Why this matters: The cost figures are the hook. They aren’t the reason this is in the newsletter, and I suspect “open model” is doing a lot of unexplained work for some of you, so let me back up.

When you use ChatGPT or Claude, you are renting access to a model you cannot see. It runs on the company’s servers. Your text travels to their infrastructure, gets processed there, and the answer comes back. You have no copy of the thing, and if the price changes or the terms change or the company decides your use case is no longer welcome, your options are to accept it or start over.

An open-weight model is different in one specific way: the trained model itself is published as a file you can download. Meta’s Llama, Google’s Gemma, and Nvidia’s Nemotron are the ones AT&T is reportedly using. Anyone can take the file. That means you can run it on a server you control, or a laptop, or a rented machine in a data center you chose. The words you type never leave the hardware you’re running it on, because there’s no outside service involved.

For most organizations the practical version isn’t a server in a closet. It’s that you get real choices you didn’t have before. You can pay a hosting provider to run an open model for you, which is cheaper than frontier-lab pricing and lets you switch providers without switching models. You can run a smaller one locally for the sensitive work and use a frontier model for everything else. And you stop being exposed to a single vendor’s pricing decisions, because the model you’re depending on is a file, not a subscription.

The catch, and it’s a real one, is that the best open models genuinely do trail the best closed ones. The reporting puts the gap at six to ten months. What their result suggests is that the gap doesn’t matter for most of the work. They didn’t drop the frontier models. They built a routing layer that sends easy work to the cheap model and hard work to the expensive one. That’s the actual insight, and it scales down: the question is never “which model is best,” it’s “which model is enough for this particular task.”

For a progressive organization I’d add one thing the cost story leaves out. Everybody in our world has data we shouldn’t be pasting into a rented chatbot. Member lists. Case notes. Anything a client told you in confidence. The reason to care about open models isn’t mainly the 56%. It’s that for the sensitive half of your work, there is now a serious option where the data never leaves.

Briefly — What a union asks for when it gets to write the list

The University of Washington’s labor relations office published a bargaining recap on August 21 covering three sessions with SEIU 925. Buried in it is the union’s opening AI proposal, and it’s a useful document precisely because it’s an opening position: this is what workers ask for when nobody has negotiated them down yet.

Sixty days’ notice before AI changes are implemented, up from the thirty in the current agreement. Human oversight required for automated decisions about hiring, firing, discipline, pay, scheduling, workloads and job applications. A prohibition on using AI to move bargaining-unit jobs into non-bargaining positions. Then: consent for data collection, limits on monitoring, mandatory bias and privacy audits, training, protections for workers who report AI harms, labeling of AI-generated content, and retraining for anyone affected.

Nothing on that list bans the technology. Every item is about notice, review, or being told the truth. The next session was set for August 26, and UW hasn’t published a response, so treat this as an ask rather than a win.

Source: University of Washington Office of Labor Relations, 21 August 2026

Put AI to Work

Practical ways progressives can use AI this week

Use AI without handing over your members’ data

Jeremy Caplan published a privacy toolkit on August 22, and it pairs so neatly with the open-models story that I’m running it in full.

Start with why this matters more for us than for most people. His warning is the plain one: “If you’re on a free AI plan, what you feed a chatbot might later be used to train its model.” He also notes that AI assistants have leaked sensitive material publicly. Now think about what actually sits in a progressive organization’s files: member lists, immigration status, case notes, somebody’s testimony about their landlord, a donor’s health situation mentioned once in an email. Most of us have pasted something into a chatbot that we would not have put in a public Google Doc, and the distinction between those two things is thinner than it feels.

Four ways to fix that, in the order I’d do them.

1. Strip the names before the text goes anywhere. Weeve removes identifying data from text and files before you hand them to an AI, and it’s free. Its Mac note-taker processes entirely on your device. This is the habit I’d build first, because it works no matter which AI you use afterward. Most of the time you don’t need the model to know who the person is. You need it to summarize a situation.
2. Keep transcription on your own machine. MacWhisper transcribes locally on a Mac, with a free tier and a €64 lifetime license. Vibe does the same on Windows and Mac in more than 100 languages, free. Nothing uploads. If you record interviews, intake calls, member meetings or anything covered by a confidentiality promise, this is the single change I’d make this week.
3. Redact properly, not visually. Basalt redacts PDFs by destroying the underlying text rather than drawing a black box over it. Free trial, $29 lifetime. The black-box-in-Preview approach has embarrassed better-resourced organizations than yours, repeatedly, because the text is still in the file and anyone can select it.
4. Use a chat tool that isn’t keeping your conversation. Duck AI offers anonymous chat with no login and no training on your inputs, free, with a $10 a month tier. Privatemode encrypts processing during the chat. Venice does private encrypted chat, free at ten prompts a day and $18 a month for Pro. And Jan, which Caplan flags as his preferred free offline app, runs an open model entirely on your own computer. That last one is the open-weight story from the news section, showing up as something you can install this afternoon.

One more from his list that’s worth five minutes on its own: Lumo AI Paper Trail analyzes what you’ve already exposed in your AI chat history. Uncomfortable, useful.

If you do one thing from this section, do the first one. Strip identifying details before the text leaves your hands. It’s free, it takes a week to become a habit, and it makes every other decision here less consequential.

Source: Wonder Tools, 22 August 2026

Looking Ahead

Steve Newman published a piece on August 21 arguing that we’ve been thinking about cybersecurity wrong. Security that depends on millions of developers and billions of users each doing the right thing every time isn’t a strategy, it’s a hope. His alternative is to systematically retire the mechanisms that keep failing, passwords and API keys among them, the way the web collectively moved to HTTPS: not by asking everyone to be more careful, but by making the old thing stop working.

I keep thinking about that in the context of this issue. Amodei says there’s a crisis of trust. Altman says we need to explain better. Newman’s model points at a third answer: you fix a trust problem by changing the defaults, so the untrustworthy version stops being available. A contract clause does that. So does a model running on your own hardware, and so does a consent prompt before a recording starts. None of them ask anyone to be more careful.

• SEIU 925 and the University of Washington are back at the table on August 26. The union’s opening AI proposal is above; UW hasn’t answered it in public.
• California’s SB 947 sits on the Assembly’s third reading file. It was read a third time and amended on August 21, so the floor vote is genuinely next. If it passes, California will have written a human review requirement into employment law.
• CHAI’s PULSE pilot jurisdictions still haven’t been announced. When they are, that’s the deployment story Issue 22 could only cover as a promise.

Until next time,
Jordan

Read past issues  ·  RSS  ·  Website