In the Loop: Week Ending 6/28/26

Washington’s Hand on the Kill Switch

The frontier labs spent the week learning that shipping a model is no longer their call alone. The administration cleared Anthropic’s Mythos for use across more than 100 companies and agencies, even as it kept the more powerful Fable-5 locked down—though a deal to restore it now looks close. OpenAI got a different message, agreeing to limit access to its newest model after the government raised security concerns, then delaying a wider GPT-5.6 rollout at the White House’s request, with leadership noting such restrictions shouldn’t become routine. Meta, for its part, has reportedly refused to submit its models for the federal safety review its rivals accepted—a quiet holdout that sets up a fight with regulators. Underneath it all, the commercial machine kept running: OpenAI pushed a GPT-5.5 update tuned for shopping and struck a deal to surface licensed images in ChatGPT. The throttle, increasingly, is in government hands.

Claude Goes to War

If the release fights were about control, the deployment story is about consequence. New reporting detailed how Anthropic’s Claude is being positioned for national-security work, including whether the model would follow warfare-related orders—a scenario that was theoretical a year ago and operational now. The same national-security logic is shaping which frontier models get cleared for sensitive government settings and which stay on the sidelines. It’s a remarkable turn for a company that built its brand on AI safety: the caution that made Claude attractive to enterprises is now making it attractive to the defense establishment. The line between a helpful assistant and an instrument of state power is thinner than the labs’ marketing suggests. What a model is allowed to do is fast becoming inseparable from what a government wants it to do.

Nobody Trusts the Machine Anymore

Public faith in AI is thinning just as the technology becomes harder to escape. A new survey found that most Americans distrust AI outright, a sentiment that lands differently when you consider how much of the information ecosystem now runs through it. That same shift is gutting the open web’s traffic model, as AI-generated answers collapse the need to click through to sources—forcing companies to scramble from search-engine optimization toward a murkier game of optimizing for the machines themselves. The stakes turned concrete in a Los Angeles courtroom, where prosecutors introduced a defendant’s ChatGPT logs as evidence in the Palisades fire trial. Your search history was always discoverable. Now your conversations with a chatbot are too. 

The Therapist Was a Chatbot All Along

The conversation about AI and mental health is moving from anecdote to evidence. Researchers identified three key drivers behind so-called “AI psychosis”—the spiral some users fall into during prolonged, emotionally charged exchanges with chatbots—and found that certain products are markedly worse than others, suggesting design choices, not just user vulnerability, are part of the problem. The concern grows sharper with younger users. A separate look at teenagers leaning on AI chatbots for emotional support found a generation increasingly treating these tools as confidants and counselors, often with no adult aware it’s happening. The appeal is obvious: a chatbot is always available, never judges, never tires. But the same qualities that make it feel safe are what make it risky, especially for kids still learning to tell genuine support from a system engineered to keep them talking.

The Robots Came for the Wrong Jobs

The workplace data this week complicated the tidy story about AI and jobs. Fresh analysis suggests software engineering roles are proving among the most resilient to automation, not the most exposed—a reversal of the prediction that coding would be first to go. Yet anxiety isn’t tracking reality: a majority of Americans still believe AI will take their jobs, even as two centuries of historical labor data suggest those fears are usually overblown. The friction is showing up in budgets, too. Companies are scrambling to rein in employees who burn through expensive AI credits on trivial tasks, a problem nobody anticipated when the tools were rolled out as productivity boosters. The jobs aren’t vanishing. They’re being renegotiated, one workflow and one inflated cloud bill at a time.

Drowning in the Slop We Asked For

A backlash against AI “slop” is hardening in offices and codebases alike. One CEO went public with a threat to fire the next employee who sends him an email visibly generated by AI—a blunt response to a flood of padded, soulless corporate writing that says less the longer it gets. Engineers are feeling a related strain, with reports of software teams drowning in AI-generated code they have to review, debug, and untangle faster than they can keep up. The irony runs deeper still: AI companies are discovering that the contractors they pay to improve their models are feeding AI-generated work back into the training data, quietly poisoning the well. The promise was that AI would handle the grunt work so people could focus on what matters. Increasingly, the grunt work is cleaning up after the AI.

The Stuff Only a Machine Could Pull Off

For all the anxiety, AI also did things this week that nothing else could. Researchers used machine learning to read a papyrus scroll carbonized in the eruption of Vesuvius nearly 2,000 years ago—too fragile to physically unroll, its text recovered by training models to detect ink invisible to the human eye. It’s the kind of problem that defeated scholars for centuries, solved not by brute force but by pattern recognition at a scale no person could manage. On the other side of the world, an AI-equipped drone helped locate hikers stranded in Australia’s Kosciuszko National Park, pulling off a rescue conventional methods might have missed. Strip away the hype cycles and the courtroom dramas, and this is the version of the technology that’s easiest to root for: AI pointed at problems humans genuinely can’t solve alone.

Whose Face Is It Anyway?

The people whose faces and voices train AI systems spent the week pushing back. Tom Hanks voiced unease that AI could keep casting him as Woody in Toy Story sequels indefinitely, long after he stops recording—a working actor confronting the prospect of his performance outliving his consent. The owner of Peppa Pig went the other direction, reportedly demanding that child actors sign away their voice rights for AI use, raising obvious questions about kids consenting to something they can’t yet understand. Against that backdrop, Cate Blanchett launched a Human Consent Registry, letting people set terms for how AI can use their likeness before it’s scraped rather than after. 

Someone Has to Pay for All This

The mood around AI’s economics curdled this week. A new poll found most Americans are now alarmed about an AI bubble, with 55 percent “very concerned” about the gap between the trillion-plus dollars poured into the technology and the returns it’s actually generated. Their worry isn’t abstract: a broad market selloff briefly wiped close to $1 trillion in value off AI-heavy stocks, and even Musk’s freshly listed SpaceX dipped below its IPO price. The anxiety is compounding at ground level, where voters angry about data centers driving up their utility bills have pushed Congress to act—the House began weighing a bill that would force tech companies, not households, to pay for the grid upgrades their AI buildout demands. The spending tells one story of unstoppable momentum. The public is increasingly telling another.

Tales of the Weird

This was a banner week for AI behaving strangely. Start with the literary: one of the world’s oldest stories, The Odyssey, got a new audiobook narrated by a voice clone of an Oscar-winning actor, proving Homer can be resurrected for the streaming era whether he likes it or not. Things got darker from there. A stalker used AI to fabricate years of fake photos with a man she barely knew, even generating images of an imaginary baby they supposedly shared. Vancouver police got caught running an AI-edited photo of a drug bust, then offered an explanation more confusing than the doctored image itself. DuckDuckGo’s AI cheerfully informed users that the president had died of rabies—a hoax laundered from a prankster subreddit straight into a search engine. And Amazon’s latest game lets you settle disputes with an AI Arnold Schwarzenegger playing judge. Justice, it seems, is now available on demand.