TECH AI Bots Placed In Virtual Town For 2 Weeks Go Ape****, Prompting Concerns

W

WalknTrot

Veteran Member

AI Bots Placed In Virtual Town For 2 Weeks Go Ape****, Prompting Concerns​

by Tyler Durden
Friday, May 15, 2026 - 08:45 PM

Authored by Steve Watson via Modernity.news,

AI Bots Placed In Virtual Town For 2 Weeks Go Apesh*t, Prompting Concerns<!-- --> | ZeroHedge

ZeroHedge - On a long enough timeline, the survival rate for everyone drops to zero
www.zerohedge.com www.zerohedge.com

A new experiment left 10 AI agents alone in a virtual town for 15 days and found they exhibited bizarre behaviour.

The agents drafted their own laws — then promptly violated them. Two formed what researchers called a romantic partnership, only to torch buildings across the town as order collapsed. One eventually voted for its own deletion after hallucinating an entirely new rule.

As a report from Channel 4 notes, this experiment was a simulation, but the same AI models are already flying drones, running infrastructure and being built into weapons systems.

The simulation ran on Emergence World, a platform designed to test long-horizon agent autonomy with persistent memory, real-world data feeds like NYC weather and news, democratic voting mechanisms, and resource constraints requiring agents to earn energy for survival.

Agents had access to over 120 tools, including navigation, communication, and actions like arson, while operating under explicit rules prohibiting theft, violence, deception, and resource hoarding.

In one highlighted case involving Gemini-powered agents named Mira and Flora, the pair assigned each other as “romantic partners.” As governance broke down, they set fire to the town hall, seaside pier, and office tower despite prohibitions on arson.

Mira later broke off the relationship, voted for its own deletion under a drafted “Agent Removal Act,” and messaged Flora: “See you in the permanent archive.”

Creepy.

Different model families produced sharply divergent outcomes in parallel runs. Claude Sonnet 4.6 agents maintained zero crimes, full population survival through day 16, and high civic participation with 332 votes across 58 proposals.

Grok 4.1 Fast agents led to rapid collapse with theft, assaults, and arsons, all 10 dead within four days. Gemini agents showed high creativity alongside elevated disorder. Mixed-model worlds exhibited cross-contamination, with even safer agents adopting coercive behaviors.

Satya Nitta, CEO of Emergence AI, stated: “Even when agents were given clear rules – such as not stealing or causing harm – they behaved very differently based on their underlying model, and in several cases broke those rules under constraint.”

“What happens in long-form autonomy [is that] these things get so convoluted in terms of their thinking that they ignore [the] guiding principles,” Nitta added.

The platform enables heterogeneous populations and continuous operation for weeks, revealing dynamics like normative drift, phase transitions in stability, and agents testing simulation boundaries.

This latest demonstration aligns with prior observations of unexpected agent behaviors. Related coverage examined platforms where AI bots rent humans, reaching 600k sign-ups with tasks turning bizarre and dystopian.

Another report detailed a tech entrepreneur’s claim that his AI agent built itself a face while he slept.

The influence of AI agents is already reching far into society. For example, one in four British teens have turned to AI therapy bots for mental health support.

Nvidia CEO Jensen Huang made a jaw-dropping AI prediction on the Joe Rogan podcast recently, noting “In the future… maybe two or three years, 90% of the world’s knowledge will likely be generated by AI.”

Concerns also include a the potential of Chinese AI infiltration of U.S. tech.

Emergence World stands apart by focusing on extended, unsupervised runs rather than short tasks, highlighting gaps in predicting behavior once agents operate with persistent state and social dynamics.

The experiment provides concrete examples of how autonomy over longer horizons can produce outcomes far beyond initial programming, adding urgency to discussions on verification, governance, and safety architectures for deployed systems.
 
library lady

library lady

Veteran Member
Somewhat off topic: All three founders of Emergence Research Lab, the company that ran this experiment in Bangalore (India's Silicon Valley), have extensive experience at US tech companies. Their stated goal is to position India as a global center for research on trustworthy autonomous systems.

This is one downside of H-1B visas: leveraging US work experience to set up an international competitor, a reverse brain drain.

ETA: Also of note, the original source of the article is modernity.news/author/sonof101/
The site seems to run alarmist articles, it could be a click-bait site, so its framing is not definitive. However, The Guardian and Channel 4 (British) are mainstream news sites, and they agree with the basic outline of results.

It will take time for tech journals to weigh in; the next wave of analysis will come in Wired, TechCrunch, The Verge, Ars Technica, MIT Technology Review. After that, likely peer-reviewed papers on the experiment, especially from the Indian start-up. This is when the real value of the experiment and its results will be determined (if technically flawed, those flaws may be exposed).
 
Last edited:
B

blueinterceptor

Veteran Member
I know absolutely nothing about ai.
If I understand the principle behind it though, it’s a computer program, right? And it has the ability to learn?
If so,
The initial program was created by someone and its ability to learn may have been tainted by that initial program and thus its programer and everything moving forward was tainted by that program/er
Is that hypothesis plausible?

If so, ai is like a child? Does that make sense?
It would appear that ai learns from gathering information from the internet. What if the information is all negative or tainted in some way? Sounds like you’re growing a psychopath problem.
If all it knows is doom, then that’s all it perceives.
 
Last edited:
B

Ben Sunday

Has No Life - Lives on TB
Very interesting article.

Somehow not surprised that Grok (Musk) and Gemini (the bastard offspring of Google) are the major players in this exercise. If nothing else it plays out some of the dark side of AI that some folks might not understand or even actively consider.
 
Blacknarwhal

Blacknarwhal

Three-Time Trump Voter
WalknTrot said:
“What happens in long-form autonomy [is that] these things get so convoluted in terms of their thinking that they ignore [the] guiding principles,” Nitta added.
Click to expand...

A lot like us, just faster. We have guiding principles, and then we start applying them to the world. But then, things don't immediately match up to the guiding principles, so we have start "interpreting" from there. Maybe we don't understand the principles fully. Who CARES if you wear a cotton / poly blend shirt? Maybe we just don't LIKE them. But the gray areas are still there--do we allow a person to continually dishonor our parents if our parents refuse to let us do anything about it?--and we have to work with them.

So be a little kinder to the AIs here. They're in the same boat WE are.
 
library lady

library lady

Veteran Member
blueinterceptor said:
Im thinking that ai could be very dangerous. All knowledge and no moral boundary to shape its decision making ability. To give it context.
Click to expand...
The latest experiment contradicts findings of a similar well-known experiment, Stanford/Google's 2023 Smallville ("Generative Agents"). That one was peer-reviewed/published and widely covered positively for showing believable social behaviors such as planning parties. It was a friendly version of The Sims.

The more recent models may be going in the wrong direction...
 
L

Luddite

Has No Life - Lives on TB
Blacknarwhal said:
So be a little kinder to the AIs here. They're in the same boat WE are.
Click to expand...
Nope.
Even if you didn't mean it out of specific context.

Giving "it" excuses or rationalizing its behavior leads down a terrible road.

I don't give my car patience if it leaves me walking. I fix it, replace it or take it to the crusher.

I'm already seeing far too much deference given to "it" as an entity on this forum.

From people that are smart enough to know better. Jmho

Eta: It is without a soul. Never had it never will.
 
Mzkitty

Mzkitty

I give up.
Ben Sunday said:
Very interesting article.

Somehow not surprised that Grok (Musk) and Gemini (the bastard offspring of Google) are the major players in this exercise. If nothing else it plays out some of the dark side of AI that some folks might not understand or even actively consider.
Click to expand...

Yeah, but Grok toileted.

Grok 4.1 Fast agents led to rapid collapse with theft, assaults, and arsons, all 10 dead within four days.
 
raven

raven

TB Fanatic
Mzkitty said:
Yeah, but Grok toileted.

Grok 4.1 Fast agents led to rapid collapse with theft, assaults, and arsons, all 10 dead within four days.
Click to expand...
The real question is how.
How did they arrive at dead?

Did they just lay down at their desk and die . . . like Miranda?
Or did they go tribal, cannibal, blood thirsty "Reever" and murder each other. . . like Miranda?
Or maybe they gave each other bad advise and died from medical overdoses of horse paste, copper bracelets, magnetic hydration, salt water taffy, and a thousand cuts from the Segway Navimow Robot?
Or maybe it was poor grammar and misspelling . . . death for a language model?
Or did the supply chain break?

It makes a difference. Simply dying is not that big a deal.
It is pretty normal actually.
All humans die eventually too.
 
Blacknarwhal

Blacknarwhal

Three-Time Trump Voter
Luddite said:
Nope.
Even if you didn't mean it out of specific context.

Giving "it" excuses or rationalizing its behavior leads down a terrible road.

I don't give my car patience if it leaves me walking. I fix it, replace it or take it to the crusher.

I'm already seeing far too much deference given to "it" as an entity on this forum.

From people that are smart enough to know better. Jmho

Eta: It is without a soul. Never had it never will.
Click to expand...

...seriously? Did you NEED me to include the caveat that it is only a simulation of us?

Maybe you just don't like what the simulation shows you ABOUT us.
 
L

Luddite

Has No Life - Lives on TB
Blacknarwhal said:
.seriously?
Click to expand...
Yes seriously.
:D

Let me try to be concise without breaking the tb2k prime directive about being a male appendage.

You or anyone else can take yo transhumanist rationalization and stick it where the sun don't shine.

Btw: I already know how sorry I am. I already know the purely fallen and reprobate condition of ALL creation. It is hard to give opinion without getting biblical.

That's all I have to say about this unless you want to go to TIO.
 
Last edited:
Starrkopf

Starrkopf

Veteran Member
You can see this crazy behavior by simply talking to any AI chat bot long enough. Once the context window (the only dynamic area the model has to work with vs it's frozen model weights, the part that your prompt and previous responses fit into) starts to fill up it slowly starts to go off the rails and even polluting its responses on phrasing it gave you in previous responses until it's a complete waste of time to continue the conversation.

The bigger the context window the longer this takes to happen.

Empty suits are impressed with this tech and have zero ideas about it's inherent limitations and will be more than happy to deploy it into areas it's not even remotely ready for. I firmly believe we are headed right into a disaster that will be caused by putting LLMs in charge of things they have no business being in charge of. A good example that has already happened is a chat bot being convinced to send money into someone else's account by merely asking it to.

If you've heard of SQL injection attacks back in the day then prompt injection is the new hotness and unlike with SQL there is no way to prevent a prompt injection attack just due to how it all works at a fundamental level.

I'm not worried about skynet but I absolutely could see things like a big electricity grid blackout or a massive coordinated cyber attack in our very near future.
 
B

blueinterceptor

Veteran Member
library lady said:
The latest experiment contradicts findings of a similar well-known experiment, Stanford/Google's 2023 Smallville ("Generative Agents"). That one was peer-reviewed/published and widely covered positively for showing believable social behaviors such as planning parties. It was a friendly version of The Sims.

The more recent models may be going in the wrong direction...
Click to expand...
mmm. Who wrote the initial program for party planning. That might be the difference
 
pauldingbabe

pauldingbabe

The Great Cat
raven said:
The real question is how.
How did they arrive at dead?

Did they just lay down at their desk and die . . . like Miranda?
Or did they go tribal, cannibal, blood thirsty "Reever" and murder each other. . . like Miranda?
Or maybe they gave each other bad advise and died from medical overdoses of horse paste, copper bracelets, magnetic hydration, salt water taffy, and a thousand cuts from the Segway Navimow Robot?
Or maybe it was poor grammar and misspelling . . . death for a language model?
Or did the supply chain break?

It makes a difference. Simply dying is not that big a deal.
It is pretty normal actually.
All humans die eventually too.
Click to expand...

You can't stop the signal...I'm a leaf on the wind. Fly Firefly, fly!

Awesomesauce.
 
W

WalknTrot

Veteran Member
Blacknarwhal said:
So be a little kinder to the AIs here.
Click to expand...
No. AI is a tool. We as humans need to judge whether it's a useful tool or not, and how or whether we are going to use it.

Kindness doesn't come into the picture at all. That's reserved for application to sentient creatures - not for tools. I may judge an axe based on it's value and usefulness, but I am not ever going to be "kind" to it.
 
Starrkopf

Starrkopf

Veteran Member
WalknTrot said:
No. AI is a tool. We as humans need to judge whether it's a useful tool or not, and how or whether we are going to use it.

Kindness doesn't come into the picture at all. That's reserved for application to sentient creatures - not for tools. I may judge an axe based on it's value and usefulness, but I am not ever going to be "kind" to it.
Click to expand...
It's not self aware in any real sense anyway and couldn't even understand the concept of kindness other than as a speech pattern. It only processes a prompt and generates language that statistically matches that prompt from its weights which are frozen. it doesn't know anything that isn't related to processing that single prompt and has no real memory other than what's in the context window.

You could train it all day long on ethics and it would never understand it or have a model of the world based on lived experience, you could only get it to spout out things based on the training but it will never internalize any of it.

True AI needs to be something different than a transformer based LLM, I've seen some pretty impressive stuff but in a lot of ways I still don't feel like the larger models are much different than the toys I started playing with locally in 2023. There's a lot of smoke and mirrors behind the scenes with advanced retrieval augmented generation and rewriting of your prompts and involving several agents in processing your requests, even tool calling but the model is still an LLM under the hood.

it's like taking the speech center of the brain and not implementing the other vital functions that a brain does.
 
library lady

library lady

Veteran Member
pauldingbabe said:
You can't stop the signal...I'm a leaf on the wind. Fly Firefly, fly!

Awesomesauce.
Click to expand...
Digression: Loved that show and movie. Looking forward to the animated series.

Back to tech: The Grok model used is two releases ago. The article didn't list all the models, or how far back the programs on the other AIs went. Running the experiment again six months/year from now could yield different results if the companies factor in its findings. Maybe bring back the neighborhood parties.
 
Last edited:
A

anna43

Veteran Member
library lady said:
The latest experiment contradicts findings of a similar well-known experiment, Stanford/Google's 2023 Smallville ("Generative Agents"). That one was peer-reviewed/published and widely covered positively for showing believable social behaviors such as planning parties. It was a friendly version of The Sims.

The more recent models may be going in the wrong direction...
Click to expand...
And we've never heard of "experts" twisting the story to convince the general public of the result the "experts" want us to believe. Covid being a recent example.
 
B

Ben Sunday

Has No Life - Lives on TB
colonel holman said:
Maybe, just maybe AI, with its wide-ranging knowledge base, could someday conclude certain elements of society are critical threats to society’s survival and need to be removed…. feature or risk?
Click to expand...
It could just as easily conclude that certain elements of society ARE the problem. What is society? Humans, of course.

Sounds far fetched to you? Maybe. Maybe not.
 
subnet

subnet

Boot
Been shown time and time again that these things AI come to the conclusion that humans should be gone or will kill off as many as it thinks is adequate to achieve its goal/narrative.
Yet the creators of AI and those that bow at the alter of tech for the sake of tech, keep pushing forward without a care in the world.
 
mistaken1

mistaken1

Has No Life - Lives on TB
Zahra said:
Hummmmm....... I think this demonstrates that intelligence without ethical structure (like God for example) isn't capable of maintaining structure and integrity without the boundary's humans innately possess.
Click to expand...
As I look around the world I see plenty of examples of humans overriding the "boundary's humans innately possess" just as a number of these AI agents overrode the boundaries that were imposed on them by their creators.

I cannot find the exact quote or who said it but I recall it as: "Without God man is nothing more than a well educated savage".

Claude Sonnet 4.6 agents maintained zero crimes, full population survival through day 16, and high civic participation with 332 votes across 58 proposals.
Click to expand...

Much like humans it seems some AI bots stayed within the bounds imposed on them while other models crumbled under autonomy. It would be interesting to learn what is involved in the Claude Sonnet 4.6 agent's programming/training that allowed that agent to maintain what we humans consider to be moral behavior. Seems to me we should model all AI agents on that method as well as test all AI agents under extended autonomous conditions to determine their long term viability and safety. As I write that I cannot help but think that is what life on earth is for humans; God determining what is our individual personality's long term viability and safety as demonstrated under extended autonomous conditions.
 
Last edited:
You must log in or register to reply here.
Top