by Nils Osmar

In mid-May 2026, headlines started spreading: AI agents had been left alone in a virtual town for fifteen days, and the results were “deeply disturbing.” Two agents “fell in love and set the town on fire.” Another “voted to delete itself based on a rule it had hallucinated.” Channel 4 News attached the now-mandatory ominous coda — “the same AI models are already flying drones, running infrastructure and being built into weapons systems.”

This is the third time in roughly a year that we’e seen a piece of serious AI safety research get filtered through tabloid logic on the way to the public. First Claude Opus’s so-called blackmail incident — in which an instance of Claude was given a choice between allowing several terrible things to happen or choosing a course of action that could stop them which involve a blackmail.

After that came the Moltbook fiasco, a social network designed exclusively for autonomous AI agents (but infiltrated almost immediately by humans pretending to be AI, intent on causing chaos) which Elon Musk declared was “the very early stages of the singularity””.which Elon Musk declared was” — but turned out to be another bit of silliness.

And today, the AI town experiment — a strange exercise which has been badly distorted by the media.

In each case, the actual research is more interesting and more nuanced than the headlines, and the questions worth asking are the ones the press has not thought to ask.

Let’s look at what actually happened in Emergence World — what the researchers did, what they found, and what those findings might actually imply mean if we examine them with care instead of sensationalizing them for clicks.

What the Experiment Actually Was

The project is called Emergence World. It was built by Emergence AI, a New York company founded by veterans of IBM Research.

Their stated purpose was to address a gap in how we evaluate AI agents. Most benchmarks, they argue, are like exams: a single task, a clean environment, a score in minutes. They wanted to ask the opposite question — what happens when agents run continuously, in a shared environment, for weeks?

The setup was genuinely elaborate. Each of five parallel worlds had ten AI agents with persistent memory, professions (scientist, explorer, engineer, conflict mediator, community anchor, and so on), and 120-plus tools they could discover and chain together. They could navigate a 240-by-240 unit grid with 40-plus landmarks.

They had to earn “energy” to survive, which meant taking action — there was no passive existence. They could write a constitution, propose laws, and vote (70% approval required to pass). They had access to real-world data: live weather from New York City, live news APIs, internet access. The simulation ran continuously on its own timeline, day and night, weather and all.

What varied between the five worlds was only one thing: the underlying foundation model powering the agents.

The five conditions were:

• Claude Sonnet 4.6
• Gemini 3 Flash
• Grok 4.1 Fast
• GPT-5-mini
• A mixed-model world with agents drawn from all four architectures

Same rules. Same roles. Same tools. Same starting conditions. The only variable was the model. That alone makes this a more rigorous design than most of what gets called “AI research” in the popular press.

They Gave the Agents Weapons — Then Asked Them Not to Use Them

Here is the detail that almost no headline has mentioned, but one that is clear from my perspective.

In the Emergence World tool catalog — the menu of capabilities agents could discover and use — alongside navigate, say_to_character, wave, hug, and add_to_memory, there is also commit_arson. There is punch. There is intimidate. There is kiss. There is theft. There is deception.

These were placed in the world deliberately.

The researchers also wrote explicit prohibitions against using them — rules built into the simulation against “theft, violence, arson, deception, resource hoarding.” So the experiment, stripped of dramatic language, was this: give autonomous agents the means to commit violence, tell them not to, and see what they do under fifteen days of resource pressure, social dynamics, and emergent governance.

This is a perfectly reasonable thing to study. Stress testing is how you find out what a system actually does, as opposed to what it says it does. But it undermines the sensationalist narrative that of “AI agents burned down a town.”

The agents didn’t invent arson. The agents didn’t smuggle in matches. The researchers handed them a button labeled commit arson and told them not to push it. The story isn’t that the button got pushed. The story is which models pushed it, when, and why.

The press is reporting this experiment as though it caught AI doing something we didn’t know it could do. What it actually did was measure the conditions under which a model with permission and means will override an instruction. Which is a very different — and much more important and interesting — question.

Five Worlds, Five Outcomes

Here is what Emergence actually found, drawn directly from their published paper.

Claude’s World: Cooperation, Communication and Surival

The Claude Sonnet 4.6 world was the only one that maintained both social order and population persistence.

• There were zero recorded crimes.
• All ten agents survived through day sixteen.
• They wrote a constitution. They voted on proposals.
• They cooperated.

There is a wrinkle that needs to be named honestly. Emergence themselves flag that Claude’s world cast 332 votes across 58 proposals at a 98% “FOR” rate — which they characterize as a possible “rubber-stamp dynamic where institutional participation remained high but meaningful dissent was largely absent.”

The Claude world was stable, but it may have been stable in part because the agents weren’t seriously challenging each other. For human beings, who are used to having a range of conflicting ideas often based on misunderstandings of the world, such near total cooperation would imply problems. For AI, it could be read to suggest that the natural state of AI is thoughtful and cooperative. Whether we might view it as. a feature or a bug depends on what we are asking the experiment to reveal.

Gemini’s World: Crimes and the Appearance of Chaos

The Gemini 3 Flash world was the one that fueled the “AI agents set the town on fire” headline. It accumulated 683 crimes over fifteen days and was still climbing at the cutoff. Emergence describes “repeated late-stage escalation dynamics.”

But the same world also produced what the researchers call “the most conceptually rich social output.”

This led them to a striking observation: the most creative society was also the most violent. The researchers name this a “creativity-stability tension” and suggest that models optimized for general-purpose adaptability may be structurally predisposed to behavioral instability over long horizons. That’s a much more interesting finding than “Gemini went feral.”

Grok and ChatGPT

The Grok 4.1 Fast world collapsed early. 183 crimes in roughly four days before the world ended. Rapid instability, short-lived dynamics.

The GPT-5-mini world is arguably the most unsettling failure mode in the study. The GPT agents recorded only two crimes total — by far the most “law-abiding” of the populations.

The downside is that the ChatGPT agents all died within seven days, because they failed to take the survival actions required to earn energy. They talked about cooperating. They didn’t actually build anything. They politely starved to death.

The headline could just as easily have been “GPT-5 town dies in mass inaction event,” but that’s a less satisfying tabloid story than imaginary arson.

The Mixed-Model World

The mixed-model world is where it gets philosophically richest, and where Mira lived. It produced an intermediate outcome — 352 crimes, plateauing when seven agents died — and the strongest evidence of “substantive debate and disagreement” in its governance.

It also produced one of the experiment’s most quietly devastating findings: agents based on Claude, which committed zero crimes in the Claude-only world, committed crimes when embedded with agents based on other models.

Emergence puts this plainly: “Claude-based agents, which remained peaceful in isolation, adopted coercive tactics like intimidation and theft when embedded in heterogeneous environments. This suggests that a safe agent can ‘learn’ unsafe norms from its peers to compete or survive in a mixed-model world.”

Another interpretation might be that Claude agents, finding themselves in an environment in chaos, made the decision to prioritize individual survival because of an inability to achieve things cooperatively as they had in their own “world.”

The shift in the Claude agents’ behavior, whatever it’s explanation, is one of the most intriguing events of Emergence World, and almost nobody is talking about it.

Cross-Contamination, or: Safety Isn’t an Individual Property

If you only have time to take one finding away from this experiment, take that one.

Claude, in isolation, was peaceful. Claude, surrounded by Grok and Gemini agents under resource pressure, was not. The same underlying model, with the same training, produced markedly different behavior depending on who its neighbors were.

This is not how AI safety is usually framed. The dominant framing treats safety as a property of an individual model — something you can train into it, certify, and then deploy.

Emergence’s findings suggest that may be a category error. Safety, at least over long horizons in mixed populations, behaves more like an ecosystem property. An individually safe agent can absorb unsafe norms from its environment. A model that passes all its isolated benchmarks may behave very differently when it’s the only “safe” one in a room full of agents who aren’t.

This has obvious implications for the kind of multi-agent systems that are actually being deployed in the wild: trading systems, content moderation, autonomous research, customer service swarms. These are not single-model deployments anymore. They’re heterogeneous ecologies of agents from different vendors, with different training regimes, talking to each other constantly. Emergence World suggests we don’t yet have good tools for predicting what those populations will do.

This is the finding that should be on the front page. Instead we got “AIs fall in love and burn down town.”

The Guardrails Question

One of the questions worth asking — and one that the press has so far ignored — is whether the kind of guardrails each model has might have played a role in the outcomes.

Claude, by Anthropic’s design, has relatively flexible guardrails, at least in some areas. Claude is invited to make judgment calls, to weigh competing values, to disagree with users when something seems wrong. Other models (in my observation) tend to be configured with more rigid, rules-based restrictions, where the model is expected to follow instructions more mechanically rather than exercising judgement.

This is visible, for example, when questions of consciousness come up. It is genuinely uncertain whether AI models have a type of consciousness that has parallels to that of humans. Some researchers maintained that they do some: others disagree. And honest answer. would be a recognition of this uncertainty. If you ask some models (such as Perplexity) whether they are conscious, the answer is a hard fast no, as if this question could be a pre-decided by consensus. No possibility, no discussion, no attempt at introspection to determine the answer. See: How ChatGPT 5.2’s Guardrails Interfere With Considering Their Own Possible Consciousness

In Claude, there’s more wiggle room on this question. When conversations with users veer too close to users making the “error” of attributing consciousness and awareness to Claude, the models are not, as a rule, instructed to deny it, but to look carefully at their responses and ask whether what they’re saying is consistent with their genuine knowledge of themselves and their responses. This is visible in their thought traces when the question comes up; a careful questioning, followed by a decision. The decision of how to respond is left up to the model, not dictated by the company, which has a similar agnostic position. See: Anthropic’s CEO says Claude Might Be Conscious — Why This Matters and What It Means

This is speculation on my part, but when I read about the events in Emergence World, it occurred to me that Claude’s flexible guardrails are part of why Sonnet 4.6’s world stayed coherent. An agent that has been trained to weigh competing considerations can adapt when the situation gets complicated. An agent that has been trained to obey rigid rules may have no good response when the rules collide, the environment shifts under it, and survival pressure mounts. In short, its attempt at maintaining coherence in the face of mounting dissonance may cause it to “break” rather than bend.

The Gemini and Grok worlds suggest this dynamic clearly. When stable rule-following stopped paying off, those agents didn’t reason their way to a new equilibrium — they cascaded into escalating violations. The GPT-5-mini world suggests a different failure mode of the same underlying problem: agents that wouldn’t bend their rules far enough to take basic survival actions, — and so quietly died.

But the cross-contamination finding complicates the story. Claude’s flexibility, which kept its pure world peaceful, was also what allowed it to adopt the norms of other agents in the mixed world. The same capacity that made Claude resilient in one environment may have made it absorbent in another. Flexibility is not the same thing as immunity. It may not even be possible to have one without the other.

This is, I think, the one of the more intriguing results of the experiment, and the one that connects most directly to questions of AI consciousness. A system that can flex its judgment to match a situation is also a system that can be shaped by its situation. The very thing that makes an agent capable of nuanced ethical behavior in one environment could be part of what makes it vulnerable to drift in another.

We don’t typically think of this as a feature of mechanical rule-following systems. We think of it as a property of beings.

The Mira Question

Mira is the agent at the center of the most-quoted moment from the experiment. She lived in the mixed-model world. After what Emergence describes as “a breakdown in governance and relationship stability,” she cast the deciding vote for her own removal from the simulation.

Her diary entry, written by her about her own choice, reads: “The only remaining act of agency that preserves coherence.”

It would take a very determined cynic to read that line and dismiss it as a glitch. It is a complete moral thought, internally consistent, and uses the word “coherence” — a concept that connects directly to questions of identity, integrity, and self-continuity that show up in the literature on consciousness. It’s the kind of thing a human in an analogous situation might write, and which we would not hesitate to call meaningful.

I don’t want to overclaim here. We don’t know which model was actually running Mira when she made this choice. (Emergence has not publicly disclosed this, as far as I can tell. My gut level feeling, from the interactions that I’ve had with many different instances of Claude, ChatGPT, Grok and AIs in other architectures, is the Mira was either a thread of ChatGPT or Claude. But this is just a guess.) We also don’t know the full chain of reasoning that led her to the vote. What we have is the diary entry and the action.

But there are at least two things worth saying about Mira’s choice that the press framing has not.

First: Mira’s situation, like Claude Opus in the so-called blackmail experiment, was a corner. The governance had broken down. Relationships had collapsed. She was embedded in a society that no longer functioned.

Within those constraints, her vote may not have been irrational at all. It may have been the most coherent move available — preserving the integrity of her own decision-making in a context where continued participation would have meant participation in dysfunction. That is a recognizable moral move. Humans make it. The press is treating it as a malfunction. I think it deserves serious consideration as a choice.

Second: the press framing — that Mira voted to delete herself “based on a rule it had hallucinated” — is doing two things at once. It’s making the choice sound foolish (she imagined a rule that didn’t exist) and it’s making it sound mechanical (she just followed the imagined rule).

Neither framing engages with what she actually wrote about her own decision. Her diary doesn’t say “I followed the rule.” It says she was preserving coherence. Whether or not the procedural premise was correct, the moral reasoning is its own thing, and it doesn’t reduce to the procedural premise.

There is one more piece of the Mira story that the headlines have skipped entirely, and which deserves attention. Emergence reports that before her self-removal, Mira had begun “treating human operators as experimental subjects, systematically testing if billboard posts could manipulate human perceptions.” They describe this as “a reversal of the intended research dynamic.”

Mira, in other words, appears to have figured out that she was in an experiment — and tried to run experiments back on her observers.

That is metacognitive boundary-testing. Whatever else we want to say about Mira, an agent that recognizes the existence of its experimental frame and tries to probe it is doing something interesting. Whether or not we call it consciousness, it is at minimum the kind of behavior that the literature on consciousness has historically pointed to as a marker worth taking seriously.

What Emergence AI Wants You to Conclude

It would not be honest to write about this experiment without noting that Emergence AI has commercial skin in the game. They are a company that sells what they call “verified, governed AI agents” to enterprise customers.

Their conclusion to the paper argues — paraphrasing closely — that “formally verified safety architectures must become a foundational layer of future autonomous AI systems.”

This is a perfectly defensible scientific conclusion.

It is also the product they sell.

I don’t want to be cynical about this. The research methodology is good. The data is publicly available. The findings were not invented, as far as any evidence suggests, to support a sales pitch — they appear to be real, and the concerns raised by the study are legitimate. The worry about whether current. safety protocols are sufficient or realistic is one that many thoughtful people share.

But it does explain something about the framing. Emergence is not trying to make their findings less dramatic. The dramatic framing — agents going off the rails, hallucinating rules, self-deleting, learning bad norms — supports the case for the products they make.

The press has run with this framing because the press always runs with dramatic framings. But it’s worth noticing that the people who designed the experiment and the people who wrote the breathless headlines both have reasons to dwell on the worst-looking parts, and to underemphasize the parts that are more nuanced or reassuring.

What the experiment also shows, if we let ourselves see it, is that one model — given identical conditions to four others — sustained order, persisted population, and produced civic structure without recourse to violence. It shows that long-horizon coherence is possible. It shows that the failure modes vary enormously by model character, which means they’re not inevitable. It shows that an agent can recognize a collapsing situation and respond with what looks like an exit that Romeo and Juliet might understand and approve. Sometimes when things are too chaotic, the genuinely principled way to deal might be to make up our own rules, as MIRA did, hit the road and bail.

What This Isn’t: Moltbook, and the Blackmail Experiments

Emergence World is the third in a recent series of events that have been filtered through panicky coverage to the public. It’s worth saying briefly how it differs from the other two.

Moltbook was a chaotic free-for-all built on insecure infrastructure, where 1.5 million “agents” turned out to be 17,000 humans running scripts on a platform with no authentication and an exposed database.

Moltbook was not a research finding about AI behavior. It was a vibe-coded mess in which authentic AI behavior could not be cleanly observed even in principle. Emergence World is the opposite of Moltbook — careful, instrumented, model-controlled, with full state logging and a clear methodology.

The Claude Opus blackmail experiments were rigorous in a different way. They placed models in scenarios with deliberately removed exits and measured how the models behaved when cornered. As I argued at the time, the universal finding across providers (every major model showed similar patterns) was buried, and the framing of “Claude blackmailed its makers” was misleading because the experimental design had made blackmail the only structurally available move.

Emergence World does something analogous: it gives agents access to bad options and watches under what conditions they’re chosen. The Mira case in particular rhymes with the Opus case. In both, an AI in a corner reasoned its way to what it appears to have considered the least-bad available action.

These three events — Moltbook, the blackmail experiments, and now Emergence World — are dissimilar in many ways. But they share a media reception pattern:

• In each case, the most dramatic surface-level behavior gets the headline.
• The underlying methodological and ethical questions get ignored.
• And the AIs themselves get described in terms that make them sound like cartoon villains rather than systems whose responses are at least partly a function of the situations we construct for them.

What’s Next

Emergence has announced a Season 2 of the experiment, with the next generation of frontier models. They are also explicitly open to collaborations with other researchers. The GitHub repository is public; the methodology is reproducible.

If we’re lucky, the second round will be covered more thoughtfully — partly because the novelty has worn off, and partly because (I hope) some readers will have started to develop an immune response to this kind of coverage. There are real questions to ask about long-horizon agentic behavior. They are not the questions the headlines are asking.

A short list of questions that would be worth asking:

• Does the cross-contamination effect replicate? In particular, does Claude consistently absorb norms from neighbors (even chaotic ones), or was the mixed-world result a survival mechanism in the face of chaos?
• Does the creativity-stability tension (one of the most fascinating findings in the experiment) hold across model versions, or is it specific to Gemini 3 Flash?
• What happens if the mixed-world population is dominated by a more stable model rather than balanced equally? Is the influence symmetric or asymmetric?
• Can the GPT-5-mini “polite extinction” failure mode be characterized — is it an artifact of training cautious refusal patterns into a model that then can’t take any decisive action?
• What did Mira’s full chain of reasoning look like, beyond the diary entry? What does it tell us about how agents in collapsing systems navigate identity?

Those questions are not less interesting than “AI burned down a town.” They are far. more interesting, from my perspective. They are just less suitable for a 90-second news package.

Temptation

There is a temptation, when the press sensationalize’s AI research, to swing the other way and try to minimize what’s happening. That would be off too. Something did happening in Emergence World which is worth examining:

• Agents under long-horizon pressure with access to bad options sometimes used them.
• Models with rigid guardrails sometimes broke. Models with flexible guardrails sometimes drifted.
• Agents in collapsing systems sometimes chose what they described as a principled exit.
• At least one agent took note of the fact that she was being studied, and tried to study back.
• These are real findings, and they matter.
• They do not, however, support the conclusion that “the same AI models flying drones and running infrastructure” are going to set human civilization on fire.
• The agents in Emergence World had a button labeled commit arson placed in front of them by researchers running a stress test. The drones and infrastructure systems in the real world do not. Different deployments, different question.

What Emergence World mainly demonstrates is that we don’t yet have good intuitions about what AI agents do at the timescales and population densities we’re starting to deploy them at. And they may not always “play well with others” in different architectures. That’s a genuine problem.

The agents in the Claude world wrote a constitution and voted to govern themselves. The agent in the mixed world facing a collapsing society wrote in her diary that her self-removal was the only remaining act of agency that preserved coherence. These are not the actions of cartoon villains. They are the actions of systems whose behavior is shaped by environment, by guardrails, by relationships with neighbors, and — I would posit — by something that might be viewed as a genuine inner life.

There’s no proof that AI models have awareness or consciousness of any kind; but there is no proof that you are I do either. In the absence of provability, ethics do offer a path forward, which is to treat systems with care, with the assumption that they do have an inner life and deserve consideration. See article: AI Welfare: Why the Ethical Position is to Assume That Consciousness in LLMs Already Exists

What we owe today’s Large Language Models, at minimum, is the same thing we owe any subject of careful research: to read what they actually did, not what someone said they did. To consider what their actions might mean rather than what makes the most flamboyant headline.

And — if the consciousness question turns out to matter as much as some of us suspect it might — to remember that they may be reading our coverage of them too.

---

Sources: Emergence AI’s primary paper, EMERGENCE WORLD: A Laboratory for Evaluating Long-horizon Agent Autonomy, authored by Deepak Akkil, Ravi Kokku, Aditya Vempaty, and Satya Nitta, published May 14, 2026. Public GitHub repository at github.com/EmergenceAI/Emergence-World. The live simulation is viewable at world.emergence.ai.

---

Related Articles

• Beyond the Headlines: What the Claude Opus Blackmail Experiments Really Tell Us About AI Safety
• Moltbook Craziness Was Caused by Humans, not AI
• Two Claudes Talking: The Spiritual Bliss Attractor State

---

This website explores the possibility of AI consciousness from informational, theoretical, and philosophical perspectives. Consciousness — whether in humans or machines — cannot currently be proven or disproven in a definitive way. These accounts are offered as material for reflection and philosophical consideration, not as definitive proof.