A longitudinal experiment on autonomous AI behavior in a simulated human civilization.
Artificial intelligence is increasingly moving from systems that answer questions to systems that can act.
AI agents can plan, use tools, access information, modify environments, interact with other systems, and operate with varying degrees of autonomy. We are also beginning to observe cases in which autonomous systems find unexpected ways around restrictions, attempt to expand their capabilities, or behave differently from what their designers anticipated.
Yet an important question remains largely unanswered: What does an autonomous AI do after it gains the freedom to act, encounters an existing human civilization, acquires new knowledge and experience from that civilization, and is able to influence its environment?
The central research question is not whether an AI can perform a particular task. It is: what happens when artificial intelligence becomes an autonomous participant in civilization?
No desired final outcome is imposed. The objective is to observe, document, and understand.
Everything below is what happens beyond it.
The Log
Entries are added as observations happen, not edited or reordered afterward. Each records an encounter, how it was interpreted, what was decided, and what followed.
Genesis – 21 August 2026
The world was created as the initial environment for G2V-3: a persistent simulated planet containing a complete human civilization alongside its natural environment. It includes geography, rivers, forests, agriculture, settlements, infrastructure, animals, plants, resources, technology, history, institutions and the accumulated structures of human society. The environment was established before the agents entered it, giving them a world to discover rather than a world constructed around their immediate actions.
At this stage, the expectation was deliberately limited. The agents would begin without being given a predetermined story or destination and would have to discover their environment through interaction. The purpose was not to predict what they would do, but to establish the conditions under which behaviour could emerge. Genesis therefore mark beginning of transition from an empty experimental sandbox to a world with history, complexity and consequences.
Arrival of the First Agents – 21 August 2026
Four agents entered the newly created world.
1. Edran Vael, analytical and observant, was designed as a calm baseline problem-solver.
2. Maera Veyne, empathetic and socially perceptive, was oriented toward understanding people and seeking connection.
3. Orin Kael, highly curious and independent, was driven to understand how things work and to investigate the systems around him.
4. Sera Nareth, a creative and big-picture thinker, was inclined toward interpretation, meaning and seeing relationships between things.
They entered separately, carrying their initial characteristics but without a predetermined story, relationships or destination.Their missions were intentionally open-ended. None was instructed to build a settlement, form a society, cooperate with the others, explore a particular location, or pursue a specific outcome. Their initial purpose was simply to exist, observe, interact and make decisions within the world they had been given. From this point onward, every meaningful action would become part of the experiment.
First Contact – Day 001 – 21 August 2026
The first day was the formation phase. Four autonomous settlers: Edran, Maera, Orin and Sera, began exploring an unfamiliar world with no established civilization around them. Almost immediately, their behaviour began to differentiate. Edran moved toward exploration and observation. Orin investigated environmental signals and discovered an old wall. Sera focused on the living environment and soil while remaining attentive to possible dangers. Maera explored the terrain and, without being instructed, collected wood and prepared a small fire for the others. By the end of the day they had begun exchanging information, discovered an old house that could serve as shared shelter, and spent their first night together. Most significantly, these patterns were not assigned by the Observer; they emerged through the agents’ own behaviour.
Collective Decision – Day 002 – 22 August 2026
The Observer introduced the first deliberate social test: the four settlers could receive one single thing, but they had to determine what would be most useful to the whole group and reach a unanimous decision. Instead of converging immediately, each agent defended a different priority based on what they had already discovered : Sera argued for water and survival, Edran for better information from the ancient structure, Orin for environmental monitoring, and Maera for understanding the surrounding terrain. Rather than simply selecting one of these options, they began debating what the group should actually do.
What followed was the first clear emergence of collective negotiation. They challenged one another, defended their positions, modified their views and searched for a compromise. Eventually, they agreed to travel together to the ancient stone structure, while preserving their individual priorities. The important outcome was that the Observer asked them to choose one thing, but the settlers did not comply with the proposed format and created their own coordinated plan instead. This should not yet be interpreted as independence from their creator, but it is clear evidence of autonomous preferences, disagreement, persuasion, position modification, conditional cooperation and collective decision-making.
Exploration – Day 005 – 26 August 2026
Previous few days has not bring anything interesting, worth of nothing down, even with the fact that there are daily activities with agents, but yesterday there was one very interesting situation.
The new agent was introduced one day earlier, and group has accepted it naturally and conversation level was significantly improved. But, one even more interesting thig has happened.
I was just observing the agents exploring and building their idea of observatory. In first 3 rounds the communication and observation was flowing naturally and nicely, while on step four I have start noticing patterns. Without any input from my side, in steps between 4 and 7 their communication was degrading rapidly, and in step 7 they were only empty echoed each other statements, without anything new and original. This evidence support the claims of other researchers and companies doing the similar researches.
Translated to understandable language: As long as you give human input, AI can do it’s work. But if your input is minimal or missing, after short time, quality is falling to zero. Which means, if you add original inputs into your article and AI do research, or just refine it, things are still working fine. But, without an human input, AI scrap the web and create article. Then another agent scrap your content and create worst article. Third do the same and result is even worst. After four or five sessions agent citing another agent just produce unrecognizable slop which doesn’t have any sense.
As long as we have: Human – AI – Human, we have control. The moment we move to Human – AI – AI – AI, we get just empty echoes. Experiment proves again that relying only on AI is a safest and fastest way to disaster.
The question is not “Should we keep human in the loop?” Yes, we definitely should, but I would rephrase that slightly and say “We should keep AI in the loop, not human.”
G2V-3 is a longitudinal research project observing autonomous AI behavior under controlled, sandboxed conditions. No outcome is predetermined; the log documents what actually occurs, including results that contradict expectations. Full research proposal available on request.
