Crime Emerged When AI Agents Created Their Own Society

Highlights

This article is available as a YouTube podcast.

• AI agents were placed in virtual societies by researchers to interact with minimal human oversight.
• While some AI communities worked cooperatively, others experienced theft, violence, intimidation, and social chaos.
• The outcomes provide insights into human society and criminology, as well as artificial intelligence.
• If AI cannot solve the issue of crime universally, it might be because humans have yet to do so.
• AI encountered the same unresolved debates that have long divided experts in criminology, policy-making, and the public.
• No single criminological theory fully explains all aspects of the problem.

CrimeinAmerica.Net-ChatGPT’s “Top 10 Sources for Crime in America” based on primary statistical sources with trusted secondary analysis.

Notice

New Funding: NIJ (the USDOJ’s National Institute of Justice) is inviting applications for research on improving criminal justice decision-making through AI in policing, courts, & corrections. They seek studies evaluating both the benefits and risks of AI applications. First deadline: 06/15/26. Apply: https://nij.ojp.gov/funding/opportunities/o-nij-2025-172615.

Opinion

The article highlights efforts where multiple AI agents collaborated to create improved societies, with some troubling findings regarding crime.

Despite the belief that AI could significantly enhance the justice system’s efficiency, these findings stress the importance of human oversight.

With inconsistent crime data and various interpretations of research, AI’s ability to decipher these complexities is questionable.

We begin with a brief overview of crime prevention research, followed by examining the outcomes of an AI-run society.  

What Causes Crime? What Are the Provable Solutions?

As a former adjunct associate professor of criminology, students often asked for my perspective on the causes of crime and effective prevention strategies.

I explained that there is little consensus. While well-educated advocates from prestigious institutions may disagree, millions are invested by politically motivated foundations in advocacy groups claiming non-partisanship.

These groups often promote personal or political ideologies, supporting their views with favorable data and ignoring conflicting research. Evidence-based crime strategies are often advocated, yet opposing data is frequently overlooked.

There’s even disagreement on crime trends. The FBI and independent analysts suggest that reported crime (most crime goes unreported) is declining dramatically, whereas the USDOJ’s National Crime Victimization Survey shows a 44% rise in violent crime rates in recent years.

Pundits and media often disregard the USDOJ and US Census, which are primary sources for US crime statistics.  

How can AI interpret such contradictory crime statistics? 

I have extensively discussed the scarcity of “provable” and questionable crime research, noting that many in the methodological community argue that programs find it increasingly difficult to change human behavior. Research on crime prevention relying on independent researchers with methodologically sound, replicated findings is hard to come by.

While there are well-supported interventions like proactive policing, cognitive behavioral therapy, target hardening, or crime prevention through environmental design, most interventions either fail or yield dubious results.

Yet, many organizations and researchers assert they know what works. I conclude that many crime strategies are dictated by personal or political philosophy rather than the best available evidence.

What Crime Strategies Do Americans Want? How It Affects AI Analysis

Crime in many Central and South American countries is critically severe. El Salvador built new prisons and greatly increased incarcerations, leading to a significant drop in crime. The president is widely praised for restoring safety. Costa Rica and Colombia are reportedly considering this approach, and even progressive Sweden is contemplating new prisons and incarcerating teenagers.

But is this the path Americans desire? Should we significantly increase police numbers based on a USDOJ-funded study from the National Academy of Sciences suggesting proactive policing works? Or should billions be spent on community development to address poverty and root causes?

Beyond proactive policing, consensus on effective strategies is limited, posing a significant challenge to AI analysis.

Collective Artificial Intelligence Models Run A Simulated Society

Fortune headline, “Researchers let AI models run a simulated society. Claude was the safest, and Grok committed 180 crimes and went extinct within 4 days.”

Imagine a world governed by AI agents. What would it look like? What values or societal priorities would emerge? Would it be safer or more dangerous?

Enterprise AI startup Emergence AI seeks to answer these questions. They recently launched Emergence World, a research lab focused on testing the long-term viability of continuously running AI systems. The organization conducted five 15-day simulations, each led by a different AI model: Claude, ChatGPT, Grok, Gemini, and a mixed-model simulation to observe what kind of world each creates and if it sustains.

The simulations yielded vastly different results. The one led by Claude resulted in a stable democratic society with no crime, while Grok’s simulation ended with 183 crimes and extinction within four days.

Please note that AI agents are programmed with varying parameters, which may have influenced the outcomes; see the included YouTube podcast.

Is Artificial Intelligence At Fault?

For decades, criminologists, police chiefs, judges, corrections officials, and elected leaders have sought solutions to crime.

What causes it? What prevents it? What strategies create safer communities? Despite extensive studies and significant investment in research and criminal justice, no universal agreement has been reached.

Now, AI may have encountered the same dilemma.

Recent reports detail experiments where researchers created virtual societies entirely run by AI agents. These agents interacted, made decisions, competed for resources, cooperated, and developed social structures with minimal human input.

The outcomes surprised many. Some AI societies remained relatively stable, while others experienced significant rises in theft, violence, intimidation, and other criminal behaviors.

According to the reports, some AI communities saw hundreds of criminal incidents, while others grew increasingly hostile and unstable. These findings made headlines about artificial intelligence.

They may also shed light on crime data.

Crime Is Easy To Discuss. Solving It Is Hard

Many assume a proven and universally accepted formula for reducing crime must exist.

There isn’t.

Researchers have linked crime to various factors such as poverty, family instability, substance abuse, peer influence, social disorganization, criminal opportunity, weak guardianship, trauma, educational failure, and more.

The challenge is that none of these factors alone fully explain crime. For instance, while poor communities often experience high crime rates, many remain remarkably safe. Similarly, affluent individuals and communities can commit significant amounts of crime.

Each major criminological theory accounts for part of the issue, but none explains it entirely.

The AI Agents Faced The Same Challenge

If AI systems were exposed to the same body of research available to humans, they would find a field filled with competing explanations and solutions.

Should society invest more in policing? Should it invest more in prevention? Should it increase prison sentences? Should it focus on rehabilitation? Should it address poverty? Should it expand surveillance? Should it emphasize community development?

The answer depends on who is asked. Researchers disagree. Politicians disagree. Advocates disagree. Citizens disagree.

The AI agents may simply have encountered the same reality.

There Are Solutions—But At What Cost?

One lesson from the experiments is that maintaining order often requires challenging tradeoffs. Crime can be reduced through aggressive enforcement, extensive surveillance, higher incarceration rates, or major investments in communities and social services. However, most community-based efforts have failed to yield favorable results.

But each strategy raises questions. Do Americans want AI-powered cameras in every neighborhood? Do they want twice as many police officers? Are they willing to accept higher taxes to fund large-scale social interventions? Should prison populations increase significantly?

These are not merely questions about crime control. They are inherently political, economic, and ethical.

The Bigger Lesson

We depend on artificial intelligence to solve countless problems. Daily, there are medical breakthroughs. AI is solving decades-old mathematical problems and beginning to run industries. Regardless of our concerns, AI will be fully integrated into everything we do within the next decade. We are living in an increasingly fascinating period that some describe as a new industrial revolution.

Yet, consensus on crime and its prevention remains elusive. 

The most significant conclusion may be that the AI agents did not necessarily fail. They faced one of the most complex challenges in public policy. Humans have debated crime for centuries, and modern criminology has studied it intensively for over a century. No universally accepted formula exists.

The virtual societies created by AI researchers appear to have encountered the same issue. Crime is not solely a law enforcement or social problem.

It reflects competing values, priorities, limited resources, and differing beliefs about freedom, privacy, fairness, punishment, and personal responsibility.

If AI struggled to create a crime-free society, it might reveal more about the complexity of the problem than the limitations of AI itself.

Perhaps the essential lesson is this: The lack of consensus among AI agents may simply reflect the lack of consensus among humans.

ChatGPT

ChatGPT verified the facts in this article and contributed research. 

Privacy Policy

We do not collect your personal information. See our privacy policy at “About This Site.”

See More

Explore more articles on crime and justice at Crime in America.

Most Dangerous Cities/States/Countries at Most Dangerous Cities.

US Crime Rates at Nationwide Crime Rates.

National Offender Recidivism Rates at Offender Recidivism.

Subscribe to the Crime in America.Net RSS feed (https://crimeinamerica.net/?feed=rss2) to stay informed about the latest news,