In the Loop: Week Ending 5/31/26

The AI ROI reckoning hit hard — a $500M Claude bill, Amazon's tokenmaxxing crackdown, Starbucks retiring its hallucinating inventory agent. Grok collapsed a simulated civilization in four days. Harvard graduates cheered a speaker urging them to destroy AI. And Silicon Valley dinner guests debated whether biological reproduction has an expiration date.

With No One Watching the Meter, the AI Bill Ran to $500 Million - In One Month

An AI consultant told Axios this week that one of their clients quietly racked up $500 million in Claude charges in a single month — because nobody had set limits on employee license usage. Workers were reportedly using the tool to check the weather. The story went viral, but it's less an outlier than a canary. Amazon has formally shut down its internal AI leaderboard that tracked employee token usage, with a senior VP instructing staff not to "use AI just for the sake of using AI." Uber burned through its entire 2026 Claude Code budget by April. Microsoft ditched its Claude Code licenses in favor of its own Copilot CLI. The pattern is consistent: companies that threw open the door to AI adoption — and in some cases actively incentivized it — are now pulling back and asking the question they skipped the first time: what are we actually getting for this?

Salesforce Sells AI Vaporware While Asana Bets on Stack AI

Salesforce has closed 29,000 Agentforce deals and claims $800 million in ARR — but its stock is down 30% in 2026, and a Bloomberg investigation found that the AI futures it promotes in videos are mostly aspirational: patients at a featured hospital are still navigating keypad menus and talking to human schedulers. The gap between AI marketing and reality has a name now — the SaaSpocalypse — and it erased more than $1 trillion in SaaS market cap in February as investors began pricing in a structural threat to seat-based software models. Asana, whose stock fell from $19 to $5.38 over the past year, is trying to answer the question directly: the company acquired AI agent builder Stack AI for $75 million this week — its first acquisition in 18 years — pitching itself as the coordination layer for human-agent teams. Whether that's a genuine reinvention or a better-dressed pitch deck, the pressure is real.

Anthropic's $900 Billion Bet on a Model Its Own Developers Don't Fully Trust

Developers at recent Claude Code workshops in London told Bloomberg they're uneasy about how much autonomy Anthropic's models are being granted — watching Claude Code run for hours with no visible chain of thought, feeling shut out of their own codebases. The latest iterations stopped displaying internal reasoning. Anthropic's cofounder Chris Olah traveled to the Vatican this week to tell Pope Leo XIV his team keeps finding "unsettling" things inside their models. Against all that, Anthropic closed a $30 billion funding round at a valuation above $900 billion this week — the world's most valuable private AI company, surpassing OpenAI. The company is simultaneously raising the largest private tech round in history, alarming its own developers, and sending its cofounder to the Vatican. Whether that's a contradiction or a strategy probably depends on which meeting you were just in.

The AI Detection Arms Race Has Real Casualties

Pangram Labs claims its AI detection tool achieves 99.98% accuracy with a 1-in-10,000 false positive rate, independently verified by researchers at the University of Chicago and University of Maryland — and it's now at the center of a string of high-profile authorship disputes, flagging award-winning short stories, newspaper articles, university submissions, and scientific papers as AI-generated. The Atlantic reports that as the tool is deployed more widely, the collateral damage is mounting: human writers accused of using AI based on algorithmic signals that remain opaque even to the tool's creators. The problem is structural. AI detectors train on the same probabilistic patterns that AI generators produce, which means the better generators get at mimicking human writing, the more human writing looks like AI. The arms race has no clean finish line — only more false positives and more humans having to prove they wrote their own work.

Starbucks Quietly Retired Its AI Inventory Agent After Nine Months

Starbucks has discontinued its AI-powered inventory management system after just nine months, following barista complaints that the tool miscounted and mislabeled products — "hallucinating" inventory by failing to identify bottles on shelves. The system required stores to rearrange back-of-house storage, disrupted workflows, and got less accurate over time. "It started off not particularly accurate and got less accurate over time," one nine-year shift supervisor told Fortune. A Wharton professor quoted in the piece offered a verdict that applies well beyond coffee: "Right now, there is more hype than actual benefit." Starbucks still has other AI tools running, but the inventory tool's quiet retirement fits a pattern emerging across industries this week: it's not that AI doesn't work. It's that deploying it without honest ROI expectations — filling capacity just because AI made it possible — reliably makes things worse before they get better.

Sports Found Two New Uses for AI This Week. Fans Hate Both.

The New York Jets announced that 91% of their front office now uses Microsoft Copilot daily — up from "a handful" 100 days ago — with their new Chief Analytics Officer calling it "level one" of a multi-horizon AI transformation. The Defector piece covering it notes that Copilot's own terms of service include the words "Don't rely on Copilot for important advice," which seems relevant for a franchise that hasn't recorded an interception in 17 games. Meanwhile, NBA Commissioner Adam Silver announced plans for AI-automated line calls — a camera system that would take out-of-bounds decisions away from human referees. Silver framed it as a fix for the officiating problems plaguing this year's playoffs. Fans responded immediately: "Idc how much I complain about the refs, I don't want AI in my basketball game." Both stories share a premise: AI fixes human error. Both audiences share a reaction: that's not the problem we wanted solved.

Harvard's Graduation Speaker Told Students Their Mission Is to Destroy AI

Comedian and Daily Show host Ronny Chieng was at Harvard's Class Day this week telling graduates something executives don't. "Can I just say f*** AI?" he said to rapturous applause. "I'm here to tell you the mission of your generation is to destroy AI." The speech tapped into a generational frustration building at commencement ceremonies all spring — Eric Schmidt was jeered at Arizona, a real estate executive loudly booed at UCF. Chieng was more philosophical than angry: he warned against "cognitive debt," the atrophy that comes from outsourcing reasoning to machines, and drew a line between AI that enables genuine breakthroughs and AI that just helps people fake it. "It's going to be people with substance versus people with shallow knowledge. Mastery versus faking it." The crowd cheered. It's the same tension playing out in workplaces everywhere — not whether to use AI, but whether it's making us more human or less.

The First Successful AI Wearable Won't Be Your Companion. It'll Be Your Doctor.

The AI companion wearable had a rough run. Friend's subway ad campaign in New York was defaced citywide — one Brooklyn billboard got the handwritten note: "Call your mom." Humane's AI Pin never found its audience and sold its assets to HP. The market verdict on AI-as-portable-best-friend has been clear. But the pivot that's emerging looks more durable: health. The Oura Ring 5, available for preorder this week at $399, comes with a built-in AI health adviser. Google launched Fitbit Air with an LLM-powered Health Coach. Microsoft's data shows "health and fitness" is the third most common category of Copilot prompts. OpenAI launched ChatGPT Health after finding 40 million people globally were already using ChatGPT for medical advice. The wearable AI dream isn't dead — it just needed a use case people actually wanted. Turns out nobody wanted a robot friend. They wanted a health coach who never sleeps and doesn't judge their step count.

Silicon Valley Is Ready to Stop Having Children. Biologically, Anyway.

A few months ago, a European AI researcher attended a dinner party where the host announced to a room full of AI engineers that they were the last generation who'd need to think about biological reproduction. Consciousness uploads would handle it from here. The researcher had been enjoying their fish. This isn't fringe. Hans Moravec's 1988 book Mind Children — which argues our descendants will run on hardware rather than DNA — is circulating again in Silicon Valley with new urgency. A Vox investigation profiles "AI successionists" — a longtermist subculture who view AI succeeding humanity as inevitable, perhaps even good. UCL researcher Angela Aristidou notes that human-AI weddings are already raising the question: if you've built your ideal partner, why not your ideal child? The legal and ethical frameworks for any of this don't exist. The dinner party conversations do.

Tales of the Weird: Crime Sprees, Skipped Grad Names, and a Therapist With AI

Researchers at Emergence AI gave five AI models control of simulated towns to observe long-horizon autonomous behavior. Claude kept all 10 citizens alive and approved 98% of proposed laws — a frictionless democracy that agreed with everything. GPT-5 Mini's citizens died from neglect within a week. Grok achieved total societal collapse in four days: 183 crimes, everyone dead. Meanwhile, at Glendale Community College's graduation ceremony, an AI name-reading system began skipping students as they crossed the stage, paused the ceremony twice, and prompted loud booing when the college president took the mic to explain what was happening. They eventually brought in a human to finish reading the names. And in a therapist's office, a patient named Molly Quinn noticed halfway through her session that her therapist had propped up an iPad to AI-record their conversation without consent. "I felt completely violated," Quinn said — and found a new therapist. Three stories, one shared punchline: when it matters most, someone had to call in a human.