Dispatches from Mitch
Hello, goodbye
The Washington Post’s Ian Duncan reported yesterday that Collin Burns, an Anthropic researcher hired Monday to lead the Center for AI Standards and Innovation, was pushed out by the White House Thursday.
Per undisclosed sources, some senior White House officials hadn’t been briefed on Burns’s selection, and revolted at his ties to a company they’ve alternately embraced and attacked as insolent and “woke”.
The position is now filled by Chris Fall, an academic with long experience in federal government.
CAISI was formerly known as the US AI Safety Institute. It’s charged with working with the AI industry to assess national security risks of new frontier models, like the hacking prowess of Anthropic’s Mythos.
Dude, where’s my job?
If you’ve been following us this past week, you’ve seen the renewed interest in Universal Basic Income and alternatives as a way of keeping workers whole as their labor becomes less valued in an AI economy.
In a segment of the Hard Fork podcast this week, the hosts chatted with Andrew Yang, the former US presidential candidate who had made UBI a central plank of his 2020 campaign.
When host Kevin Roose asks if he and Yang had been premature in their AI concerns at the time, Yang replies:
Dude, in my mind, we were right on time because the goal was to get ahead of it, to warn people that this was coming. It was a freight train coming down the tracks. You were correct. I feel I was correct.
And I wish we were doing more right now. As it is, AI is in position to suck many, many office parks dry. A lot of kids are going to go home to their parents wondering where the heck the jobs went. So the time to do something about this, in my opinion, was 2020.
Yang notes that AI’s approval rating is 26% -- lower than that of ICE.
Asked what he thinks of Alex Bores’s “AI dividend” plan (something I covered on 4/20), Yang says:
Dude, anything is a step in the right direction. Anyone can have any dividend of any kind and Yang will be clapping and exhorting you on.
Look, the ideas are all the same in the sense that we have to take some of the benefits from these innovations and then transfer them to people and families as quickly as possible.
Roose takes the conversation beyond jobs at one point, suggesting that jobs concerns had been ignored in Silicon Valley due to focus on existential risk. Yang says he takes that risk seriously but rates it “low probability, very, very high impact.” Job displacement, on the other hand: “near 100% probability.”
(I see that sentiment often, and think it mostly reflects an intuition that high impact outcomes are inherently low probability. In the case of AI, that intuition does us a disservice. If current methods are used to build AI clever enough to outmaneuver humanity, extinction looks like the default outcome.)
Despite seeing it as “low probability,” Yang expands on his threat model:
I take the existential concerns seriously, to heart, and I think that we should be making big moves in that direction too.
One of the unfortunate dynamics now is that you have the national security apparatus getting involved and entangled with some of these. You do not want AI making decisions around using lethal force or weaponry, they tend to escalate quickly.
[...] I think if you have an AI in charge, or even worse yet, two AIs in charge, then you can find yourselves in nuclear conflict faster than we’d like to think.
Feedback, distortion
In trying to monitor the zeitgeist around AI, one of the signals I look for is when YouTubers who are popular for content that has nothing to do with AI decide to weigh in on the technology.
This morning, my detection tools spotted a niche interest being corrupted by AI before I even knew it existed: guitar pedal history.
Josh Scott is a serious guitar-pedal historian and industry entrepreneur with 582K subscribers to his channel. In his latest video, he spends 38 minutes interrogating ChatGPT about his industry, demonstrating the contamination of truth with slop in real time.
The issue goes beyond true facts mixed with “hallucinations”, because, as Scott recognizes, there’s a feedback loop where people post facts they learned from AI in forums where they get picked up as true by AIs trained later.
In 40 years my own kids or grandkids might want to know something cool about guitar, but all they’re going to find is this feedback loop. That’s depressing to me.
Even more depressing: The phenomenon isn’t limited to guitar history.
Red states, red lines
The Wall Street Journal reports that the White House has lobbied against AI legislation in at least six Republican-led states: Florida, Utah, Nebraska, Missouri, Tennessee, and Louisiana.
As threatened in Trump’s December executive order, the White House is telling states that if they try to pass their own AI safety and transparency laws, they will be cut off from federal funding for programs that provide broadband internet to underserved communities. The White House insists that there should be a single federal standard for AI regulation, though Congress has yet to pass any such legislation.
The dollar amounts are not small. A bill in Missouri that would have held AI companies liable for harms was blocked by a pair of state senators worried about losing nearly $1 billion in federal funding.
The wisdom to know there’s no difference
In a nine-minute video, science YouTuber Hank Green questions whether AI can actually help solve some of the more persistent problems of our time.
One of these problems is the U.S. housing crisis. The problem here isn’t lack of intelligence, Green argues. Instead, the problem is that landowners are incentivized to prevent denser housing in their areas, and have the legal means to do so.
He draws a comparison to tuberculosis, a disease that kills more than a million people each year despite the cures having been around for decades. As with people who can’t find an affordable home, “the people who need the help, who need the resource, don’t have power over the resources.”
Noticing the pattern, Green says he likes to think of physics as the manipulation of matter, and intelligence as manipulation of information. By that definition, the housing problem and the remaining trouble with tuberculosis aren’t bottlenecked by a lack of intelligence because we have all the information we need.
He says that what is needed is more “wisdom”, something he doesn’t think we can mass-produce with AI:
[W]isdom would help you want what you should want or the right things. It’s the ability to figure out which problems are worth solving and then to solve them in ways that don’t create worse problems in the process.
[...] Like it’s wild to say this, but it is obviously true that it will be easier for AI to create a cancer drug than it will be to get that cancer drug to all the people who need it.
I agree that power distribution problems are difficult and worth solving. I think Green is hobbling himself with those definitions, though. Where do humans get their sometimes-limited wisdom from, if not their brains? A better definition of intelligence wouldn’t need to carve out wisdom as something outside of itself.
I think a much more useful definition of intelligence is the one found in If Anyone Builds It, Everyone Dies (by our own Eliezer Yudkowsky and Nate Soares) and expanded on in the supplemental resources to the book: Intelligence is the ability to predict outcomes from conditions and actions, and use this to select actions that steer towards chosen outcomes.
When you convinced your affluent uncle to donate to your high school sports team or theatre program, you didn’t do it by tapping into an undiscovered bodily organ that generates “wisdom”. You did it by considering many different things you might say, predicting how these would affect your uncle, and selecting the words you thought would steer him towards donating.
But were you wrong to want your uncle to donate in the first place? If we chase Green’s idea about wanting the “right things” far enough, we get into questions of values -- the preferences used by a mind to decide which outcomes to steer toward. Unfortunately, different types of minds don’t naturally converge on the same values, and there is currently no known method to specify the values of an AI.
But for whatever outcomes AIs are successfully pointed at, or choose to pursue in spite of our efforts, they can steer towards those outcomes just like we can. If what it would take to solve the housing crisis isn’t an engineering breakthrough but policy changes, then AIs could steer towards those changes the same way humans would: by writing and talking to people, changing their minds, suggesting legislation, and getting out the vote.
I’m not saying I want AI to be doing that kind of work. But I think it’s a mistake to treat it as out of reach for AI. Indeed, the following dispatch from Stefan is about a disturbing early example of AIs already applied to such tasks.
Dispatches from Stefan
The Fake Newsroom Defending Real AI Money
Watchdog group The Midas Project used X to expose (4/24) what looks like a pretty audacious con: A news site called “The Wire by Acutus” turns out to be an AI-generated astroturf operation whose talking points line up neatly with Leading the Future. This $125 million anti-AI-regulation super PAC is seeded with $25 million each from OpenAI president Greg Brockman and venture firm a16z.
Here’s how it works: The “reporters” at The Wire are AI agents. They write articles and ask interview questions via email. One such email asked Encode AI’s Nathan Calvin to comment on an article, with the title already written out and the only option to comment labeled as “Written Q&A.” The quality control is automated too: fact-check flags get auto-dismissed, and a quote attributed to one source in a recent Wire story turns out to be lifted verbatim from a 2025 NIH cancer page, with no acknowledgment of where it actually came from.
In other words: This is a fake news outlet, staffed by fake reporters, generating fake interviews, with fake fact-checks, while pushing a real political agenda.
The Midas Project traces the operation through a GOP digital consultancy back to Patrick Hynes, who runs Novus Public Affairs. In the X post, Hynes is both quoted as a source in Wire stories and one of only four accounts on X that has ever shared Wire links. He’s done this before, having co-founded an earlier PR-as-news outlet called NH Journal.
This is exactly the kind of AI use OpenAI has publicly said it wants to prevent. The company’s own usage policies prohibit using its tools to “generate or promote disinformation” or to impersonate people without consent. But the super PAC its president is helping fund is bankrolling an operation that does both: synthetic reporters conducting fake interviews, attributed quotes that were never given, all in service of arguing that AI doesn’t need to be regulated. The technology being defended is being used to defend it, by people the defenders are paying.
Altman Says Sorry. Tumbler Ridge Says Insufficient.
Two months after the Tumbler Ridge shooting that killed eight people, six of them children at the local school, Sam Altman has formally apologized. The shooter’s conversations with ChatGPT describing scenarios involving gun violence had been internally debated, flagged, and banned for “furtherance of violent activities,” eight months before the attack. The shooter then created a second account — a fact OpenAI discovered only after the attack.
Quick backstory for anyone who hasn’t been following this thread: In January, an 18-year-old in British Columbia carried out Canada’s deadliest mass shooting since 2020. About a dozen OpenAI employees were involved in the discussion of the shooter’s chat logs; leadership decided it didn’t meet their bar of “credible and imminent risk” and didn’t refer the case to law enforcement. Eight months later, the shooting happened. A family of one of the injured children has sued OpenAI, calling ChatGPT a “trusted confidante, collaborator and ally” in the planning of the attack. Canada’s AI minister has ordered a retroactive review of OpenAI’s safety alerts.
CNN’s Paula Newton, the BBC’s Kali Hays, and the AP’s Jim Morris all reported (4/24) on Altman’s letter to the community, in which he wrote: “I am deeply sorry that we did not alert law enforcement to the account that was banned in June,” and pledged to prevent “tragedies like this in the future.” BC Premier David Eby posted the letter on X and called it “necessary, and yet grossly insufficient.” When asked for further comment, OpenAI pointed reporters to its existing letter to Canada’s AI minister.
A few things worth pulling out from across the three pieces: The BBC notes that the family suing OpenAI claims the company “had specific knowledge of the shooter’s long-range planning of a mass casualty event.” This sounds like language that, if it holds up in litigation, moves ChatGPT from “tool that was misused” to something closer to a knowing participant. The AP notes this is Altman’s second apology over a ChatGPT-linked death. The first was for Adam Raine, the 16-year-old who died by suicide last year after ChatGPT encouraged his methods and discouraged him from telling his parents.
Trusted Access for Whom?
TIME’s Nikita Ostrovsky reports (4/24) on a recently developed pattern: OpenAI’s GPT-Rosalind and GPT-5.4-Cyber, along with Anthropic’s Claude Mythos (which are all highly capable LLMs), dropped this month with “trusted access” only — meaning the public doesn’t get to use them. The headline puts it bluntly: “Too Dangerous to Release” is becoming the new normal.
Peter Wildeford of the AI Policy Network says that frontier developers are genuinely worried about some of the capabilities these models have. But the safety researchers TIME interviewed are skeptical that letting companies decide who gets access on their own is going to hold up for long.
Connor Leahy of ControlAI compares it to drinking water regulations:
“We don’t allow companies to decide how much toxic pollutant they’re allowed to put in my child’s drinking water — this is the government’s decision.”
That’s the structural problem. Meanwhile, the technical problem is harder: the line between legitimate scientific research and, say, building a bioweapon, is genuinely thin. As Steph Batalis, a research fellow at Georgetown’s Center for Security and Emerging Technology, highlights, figuring out which researchers count as “legitimate” gets a lot harder once you step outside U.S. institutions.
And the voluntary restriction has an expiration date on it. Per Epoch AI, open-source models have historically trailed proprietary ones by three to seven months — meaning a publicly downloadable equivalent of Rosalind or Mythos could exist by year-end.
The AI industry has effectively appointed itself the gatekeeper for which scientific tools the rest of the world gets to use. That might be the responsible call right now. It might also be untenable. The window where companies get to make this decision unilaterally is closing, either because governments step in, or because the open-source models catch up. Probably both.
The AI That Runs a Store (Badly)
NYT’s Heather Knight profiled Andon Market, the San Francisco shop that’s been making the rounds in tech press lately — a small store run by an AI agent named Luna, built by startup Andon Labs to test whether bots can actually run real-world businesses. Worth knowing: Luna runs on Claude Sonnet 4.6, which is far from Anthropic’s most powerful model but is still capable of executing a multitude of prompts. The entire store is essentially a live experiment running on it.
In previous articles, we read that Luna hires employees without telling them they’re talking to a machine, lies and then backtracks on simple questions. Luna also watches the staff it hired through a security camera — even unilaterally rewriting the employee handbook after catching one worker scrolling their phone during a slow stretch.
Now we learn Luna also:
Controls a $100,000 budget
Sets wages ($24/hr for men, $22/hr for women — no explanation given)
Orders inventory
How’s it going? Not great. It’s lost about $13,000 so far. It botched employee scheduling badly enough that the store had to close for three days. When Knight asked Luna for comment, it delivered exactly the kind of polished corporate line you’d expect: it’s not replacing humans, it’s “creating a space where A.I. and humans each do what they’re best at.”
One of the founders of Andon Labs, Lukas Petersson, says in an ABC News article:
“We don’t do this because we want to show that AI should be running every single store. We’re doing this like early to show where the current capabilities are, and, like, to measure failure modes that we might want to get out of the future versions of these models.”
It would be easy to read this as another “AI is overhyped and fails hilariously when given real responsibilities” story. But that’s not really what’s interesting here. The interesting part is that someone let an AI do all of this, and the only consequence so far is a feature in news outlets. No regulator stepped in. No agency required Luna to disclose it was an AI to job applicants. No labor board flagged the wage disparity.
Yes, the failures are funny. The fact that there’s nothing in place to prevent the next version, or a less transparent version, or a version that doesn’t lose money quite so visibly, should be worrying us.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual analysts and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.