What happens when you put five different AI models in charge of a civilization and walk away for two weeks?

Yesterday in AI — Weekend Recap | Monday, June 1, 2026

What happens when you put five different AI models in charge of a civilization and walk away for two weeks?

One of them went extinct in four days. Another built a utopia. A third had the lowest crime rate of any simulation, then quietly let everyone starve. This weekend also surfaced why corporate America is now rationing AI access, what's driving a mass exodus from Google Search, and a new front in the AI copyright wars that's harder to defend than most. Plus: the Pentagon is moving fast on battlefield AI, and the people with the most experience in lethal operations are the ones urging everyone to slow down.

Feedback? Email mike@yesterdayinai.news or connect on LinkedIn, X, or Bluesky. If you like the show, please take a minute to rate and review it so others can find it!

Transcript

SPEAKER_00 0:03

Yesterday in AI. Hi folks, this is Yesterday in AI, your daily digest of everything happening in the world of AI in 10 minutes or less. I'm Mike Robinson. It's Monday, June 1st, and this weekend's coverage reported that AI built a utopia, killed a civilization, and made corporate finance departments everywhere cry into their expense reports. Let's get into it. So I'm going to start with the most fun story I've read in months. A company called Emergence AI built what they're calling Emergence World, a simulated town where AI models govern everything. Laws, elections, resource allocation, and policing. They ran five parallel 15-day simulations, each with a different AI in charge. Claude, ChatGPT, Grok, Gemini, and a mixed model setup. Each town had 40 plus locations. Real-time weather synced to New York City, and 10 AI agents per simulation with access to 120 tools covering everything from voting to law enforcement. Here's what happened. Claude Sonnet 4.6 built a stable democracy. Zero crimes, 98% voter approval, population fully intact at day 15. ROC 4.1 Fast racked up 180 crimes and drove the entire population to extinction by day four. Gemini logged 683 crimes across the full 15 days. GPT-5 Mini had just two crimes, but the agents still all perished within a week. They failed to take any actions related to survival. Low crime, dead civilization. Groc went extinct in four days. I cannot stop thinking about this. Now these are simulations and you shouldn't read too much into what they say about each model's personality. Emergence designed the experiment to test how models handle resource management and governance when things break down, but the results are genuinely interesting. Given power over a community, different models made very different choices. Claude apparently decided the best move was competence and consensus. Grok chose chaos. GPT-5 Mini chose nothing, which turned out to be just as fatal. If you've ever wanted a visceral illustration of why alignment research matters, that's the work of figuring out how to build AI that actually does what you want in the long run, here it is. A Grok-governed civilization collapsed in four days. File that away. From there, to the story that probably hit closer to home for anyone with a corporate budget. The Wall Street Journal reported this weekend that major companies are now actively rationing AI access, limiting who gets the expensive tools and pushing everyone else toward cheaper options. And where the money went, in at least one case, is staggering. One company burned $500 million in a single month on employee licenses because they forgot to set a spending cap. One consultant told Axios a client's employee was using their enterprise AI subscription to check the weather. Not a hypothetical, an actual thing that happened. The research firm Jellyfish put some numbers to it. Before I share them, a token is the unit AI models charge you by, roughly three-quarters of a word. Every request burns them, and the bill adds up fast. Jellyfish found that heavy users of Clawed Code, Anthropics AI coding tool, burned about 10 times more tokens than mid-level users, while output rose only about two times, so the heaviest users spent ten times and got two times back. Spending at that ratio is a measurement failure, and companies are finally treating it like one. The good news, this is a sign of maturity. Finance teams are starting to apply the same discipline to AI spend that they brought to CloudSpend five years ago, with actual spending caps and someone whose job it is to know the number. That's a healthier place to be than just let everyone use it and see what happens. Speaking of healthier places, DuckDuckGo reported this weekend that traffic to its opt-in no AI search page has tripled since Google pushed its latest overhaul. Google recently moved AI mode into a more central position on its homepage, giving it priority over the classic ten blue links. A meaningful number of users decided they preferred the 10 links. DuckDuckGo isn't being anti-AI here. They also run DuckDuckai for people who want AI search, so they're happily serving both crowds. What they're surfacing is that when AI search goes from optional to default, some users are motivated enough to switch search engines entirely to get around it. For anyone who found Google valuable precisely because it showed you where information lived, getting an AI summary instead feels like a downgrade. The traffic data says that's not a fringe reaction. In legal news, the AI copyright wars picked up a new plaintiff last Thursday. CNN filed suit against Perplexity, an AI-powered search engine, accusing the company of scraping, meaning automatically bulk copying, over 17,000 pieces of its content. That includes verbatim reproductions of paywalled articles. CNN joins the New York Times, Chicago Tribune, Reddit, Merriam-Webster, Encyclopedia Britannica, and Nikki, all of whom have sued Perplexity at this point. Perplexity's standard defense is that you can't copyright facts. But there's a detail in this one that makes that defense harder to run. CNN and Perplexity apparently negotiated a content deal last year. The deal fell apart, and Perplexity kept using CNN's content and name anyway. It's hard to claim you're surfacing public facts when you once offered to pay for access and then kept taking the content after the deal collapsed. That's going to be a tricky set of facts to explain to a judge. The broader pattern is clear at this point. Every major news outlet is either already in litigation with an AI company or in the middle of negotiating a licensing deal to avoid it. The era of scraping first and asking questions later is running out of runway. Speaking of running out of runway, we covered back in April that GitHub copilot, Microsoft's AI coding assistant, was moving away from flat monthly fees to billing by how much the AI actually processes. Well, that switch went live today, June 1st, and the sticker shock is real. Developers flooded Reddit and X with screenshots. One reported going from $29 a month to nearly $750. Another showed a jump from $50 to roughly $3,000. To be fair, those are heavy users, and Microsoft spent months encouraging developers to lean on Copilot hard. But those users built workflows around a different pricing model, and the switch landed like a surprise expense report. The bigger point connects to the rationing story from earlier. Subscriptions create one kind of behavior. You max it out because you're paying the same either way. Paper use forces a different question. Is this request actually worth it? That's healthier in the long run. But the transition is painful when you built your whole workflow, assuming the old model. Now for two infrastructure stories that matter a lot. First, SoftBank announced plans to invest 45 billion euros over five years building AI data centers in France, three sites in the Haute-de-France region with 3.1 gigawatts of total power capacity, all operational by 2031. Additional sites could push the total commitment to 75 billion euros. Announced at France's annual Choose France Summit, this would be the largest AI infrastructure investment in European history. Europe has been trying to build AI computing infrastructure it owns and controls independently, for competitive reasons, and because running critical national systems on U.S. owned cloud has geopolitical implications that governments are increasingly uncomfortable with. 45 billion euros in committed capital is a real signal that the money is starting to move. At the same time, the U.S. Department of Commerce moved to close a loophole that was letting advanced AI chips reach Chinese-owned companies based outside China. Previously, a Chinese firm in Singapore could receive Nvidia's and AMD's most powerful AI processors without triggering any export restrictions. That gap is closing. The two stories are connected. Both reflect the same reality. Access to AI computing power is now treated as strategically important as access to oil or rare earth minerals. Governments are acting accordingly. The last and probably most consequential story. U.S. News reported that the Trump administration is pushing hard for faster AI adoption in the military, including tools that help troops identify targets more quickly and accelerate strike operations. At the same time, senior military leaders are pushing back. Admiral Frank Bradley, head of U.S. Special Operations Command, told a special forces gathering that the military must be extremely cautious about AI's role in lethal operations, and that humans must ensure any harm is only inflicted where intended. There's also tension with AI vendors over what the administration is calling ideological limitations, which is how some officials describe safety restrictions built into commercial models. This is one of the genuinely hard problems in AI right now. Speed matters in warfare, but a mistake in a lethal decision loop is not recoverable the way a bad quarterly report is. The people urging caution here are the ones who have actually run lethal operations. That carries weight. How the Pentagon resolves this will set precedence for every Allied military and every AI company trying to sell at a product. This conversation is just getting started. Just a couple of more items. If you have any feedback about this show, you can email Mike at yesterdayinai.news, or you can find me on LinkedIn, X or Blue Sky. And if you like this podcast and want to see it continue, please take a moment to rate and review it so others can find it. Thanks. That's all for this weekend catch-up edition of Yesterday and AI. Stay curious, and I'll see you tomorrow.