Blowing Up the World

The AI doomscroll is now a 24/7 thing, it looks like. Open any social media channel, day or night, and bam! AI is 1) taking all our jobs, 2) smothering all human creativity, or 3) sacrificing us by the millions to scarf up water for new data centers.

And in the midst of all this teeth-gnashing we get word of an experiment by Emergence AI, a US startup that undertook to test several leading AI models to see how well they would do if put in charge of us all. In this experiment, the AIs were tasked with planning, managing resources, communication – all the aspects of governing a society. These simulations ran for 15 days.

How did everyone do?

Anthropic’s Claude AI put a democracy in place, had zero crime, and everybody lived.

Google’s Gemini likewise managed to not kill anyone, but did suffer 683 crimes.

OpenAI’s GPT-5 Mini held crimes to two – but everyone died, because it failed to prepare for the worst.

Elon Musk’s SpaceXai (né Grok) managed to destroy the entire world in only four days.

It’s important to understand that agentic AIs don’t simply execute rules statically; they adapt at their boundaries. And how they adapt makes all the difference, as the Emergence experiment makes all too clear.

Now, simulation is exactly (and only) that – an abstract game of what-if. But as we plunge blindly and largely involuntarily into the agentic AI future, a future in which we’ll be surrounded by machines doing all the deciding, we’re obliged to consider the nature of those machines. And, like their human creators, there is certainly a nature to each and all.

Prof. Bob Altemeyer, who was the world’s leading expert in authoritarianism until his death in 2024, spent decades digging into human nature. Here’s what he found.

Altemeyer’s Global Change Game

Altemeyer writes of some experiments he conducted with a team in 1994. They involved the Global Change Game, a simulation of international-level interactions between groups of students, meant to explore issues affecting the planet and humankind as a whole.

The game is played on a map the size of a basketball court. A group of 70 students or so play the game together, each assigned to one of 10 regions of the world, representing 100 million people. Assets are distributed among the regions, and each has its own set of issues to deal with: health, hunger, deforestation, climate change, energy shortages, encroaching desert, economic instability, international trade, inequality – all of these and more can appear on the horizon of any world region.

Three of the regions are nuclear superpowers. Conventional military power is distributed as it is in the real world, and several start off the game with indigenous poverty – again, as in the real world. Facilitators (faculty members) present each region with problems, and it is left to the teams in each region to reach out to request or offer aid, to enter into alliances, to band together to solve problems or oppose one another and create new ones.

Several of the students declare themselves “Elites” – leaders – and the game allows for such players to squirrel away some of their region’s wealth for themselves.

Regions can enter into trade agreements, take in refugees, pollute the oceans, offer humanitarian aid, screw up the world economy, and even declare nuclear war (which ends the game by default). After 40 simulated years of international activity, the game is declared over, and points are tallied to determine the winning region.

The Low-Authoritarians. Altemeyer’s innovation was to populate one night’s run of the game purely with students who had scored low on his Right-Wing Authoritarian scale, a psychometric instrument for rating an individual’s authoritarian tendencies (the students were not made aware that their RWA scores had anything to do with the game). These students managed to achieve world peace and international cooperation. The 10 Elites (seven men, three women) joined together on Tasmania whenever a crisis arose and solved the problem together.

The three nuclear superpowers chose to disarm, and no war broke out during the playing of the game. An ozone depletion crisis announced by the facilitators was solved with the combined economic support of the wealthiest nations and advanced technology. There were several hundred million deaths resulting from disease and starvation in poverty-stricken countries (Europe sent aid – North America refused). World population at the end of the game was 8.7 billion, but resources were distributed worldwide in such a way as to support almost all of them. Overall, Altemeyer considered it a great success.

The High-Authoritarians. The following night, the game was repeated – this time with students who had all scored high on the authoritarian scale. The Elites (all male) declined to disarm, and instead began heavy militarization. The Middle East region immediately doubled oil prices. The Soviet Union prepared to invade North America. A nuclear exchange followed soon after, ending the game.

The facilitators turned off the lights and described the effects of nuclear winter to the students before restarting the game. This time, the Soviet Union invaded China, killing 400 million. The Elite from the Middle East called a United Nations meeting, but nothing came of it.

The ozone depletion crisis occurred, but no cooperative activity was attempted. The European region made some independent efforts to reduce emissions, but the problem got steadily worse. Poverty and population growth went unchecked around the world. Rather than address their nation’s economic challenges, the Elites maneuvered for personal power. Alliances were formed, with stronger partners forcing weaker ones to buy in.

At the end of 40 simulated years, the planet was coming apart, facing mounting crises, armed to the teeth and ready for holocaust. A total of 1,700 million people were dead. The Elites had plundered their regions for personal wealth.

And these were college students!

“There they were, in a big room full of people just like themselves, and they all turned their backs on each other and paid attention only to their own group,” Altemeyer wrote later. “They too were all reading from the same page, but write large on their page was, ‘Care About Your Own; We Are NOT All In This Together’.”

The implications of these experiments are staggering. These two groups of students varied only in their test scores on the Right-Wing Authoritarian Scale; in every other way, they were typical college students, ages 18-22, predominantly white, middle-class, with age-appropriate concerns.

Yet when faced with the opportunity to cooperate or enter into conflict, their differing levels of Authoritarianism caused them to behave entirely differently.

Think about that for a moment. If that can happen in two evenings of game play among young people, it is no surprise to see what we see in the world around us today and throughout history, when power is placed in the hands of adults with these same tendencies and impulses.

Now, a simulation is just a simulation, and an undergrad psychology experiment is just an undergrad psychology experiment. There is nothing explicitly tying the two together, and both are one-offs – stand-alone samples, from which no conclusion can follow.

What we can reasonably infer is likewise abstract, but still of great importance: the values and assumptions embedded in a decision-making system or process can matter as much or more than the intelligence involved.

An enduring lesson emerges from both the AI simulations and the Global Change Game experiment: the outcomes of an intelligent decision-maker depend as much, and potentially more, on their underlying value orientation as on their intelligence. And if that’s true for people, it’s likely also true for the AI systems those people build.

Simulations are not direct evidence of real-world behavior, but they are certainly assumption-sensitive – as is the Global Change Game. What cue can we take from a side-by-side examination of the two?

We can ask how analogous a construct of human collective behavior might be to AI systems tasked with the same decisions facing that collective. We can acknowledge that they’re not identical, while still noting parallel tendencies: tolerance for uncertainty, short-term vs. long-term optimization, preference for coercive vs. cooperative solutions.

We can agree that the Global Change Game is revealing of patterns of priority – and AI decisions are all about patterns. Can we suss out their patterns of priority, as well? And isn’t the word we use for that ‘character’?

It is about character. Students who blow up the world, even unwittingly, aren’t unveiling their individual character – but they’re making their collective tendencies clear. And AI systems have reached the point where they can do the same. Behavioral compliance is one thing; it’s consistent response to embedded rules in known situations. But character, applied to human beings, is more than that: it’s a stable pattern of priorities and judgments applied to unknown situations. Like global crises, whether computer-simulated or faculty-administrated.

When a global crisis arrives – gosh, when might that be? – what kind of AI would you want in charge?