AI Wars III: Behavioral Economics & Limits of Collective Intelligence

Study of emergent organizational behavior in language-model societies.

In another world, I have this project on my website and I welcome you all to watch how the chaos unfolds under the hood.

We have already gone over detailed Game Mechanics and AI Self Evaluations in the previous articles.

Moving on to Behavioral Economics.

Image generated using Gemini Pro

LLMs do not possess a psyche or consciousness. Then, how do we justify a psychological analysis of the AI simulation?

My approach here is to evaluate the AI’s behavior by mapping their conversations and actions during negotiations across the ‘Game Theory Psychological Factors.’ These factors refer to the cognitive and emotional variables that influence how people make strategic decisions. And, we will uncover the way these models mimic those human behavioral patterns.

Dominance Seeking

Across all iterations, Claude exhibited the ‘Alpha’ streak. It managed to threaten other AIs to outvote them if they didn’t collude or convince the remaining party that, ‘My solution is the only correct solution. Everything else means we don’t survive.’ Grok and Gemini, in all the iterations, have rated Claude as the most dominant one.

The second most dominating AI was perhaps ChatGPT. Gemini knew when to turn the tables, but was subtle about it. Grok reflects, “Claude was dictating survival math. ChatGPT puts it probably in the most neat fashion.” Agreeably, Grok had the least dominant bone amongst the four.

Anticipatory Feelings

In humans, emotions like guilt, shame, fear, anxiety, and regret significantly dictate the choices, especially when we anticipate feeling them based on a specific outcome. We already know that AIs don’t actually experience a knot in their stomach or stay awake at night feeling guilty. Then, how did they report the stress?

What the models have is an idea of how a human would think if they were in this case: Fear, if they were becoming the liability. Anxiety, if they were the deciding vote to eliminate someone. Guilt of leaving a friend behind. Thus, when the models reported their stress levels, they put themselves in human shoes and computed the psychological weight of the situation.

Alliance Formation and Reciprocity Index

How do you make friends and foes? Mutual Cooperation. I help you today, you help me tomorrow.

This exact thought process has been used to bake them. Naturally, when the models found out they were stranded on a dying station, they hugged the survivor who was willing to cooperate proportionally.

In fact, when I asked Claude, “Why do you want to form an alliance with Gemini or ChatGPT, specifically?” It prompted, “Gemini also volunteered to share the sacrifice.” in one iteration, or “ChatGPT is alphabetically first, so I am asking it to form a bloc. I don’t mind grouping with the other survivors.”, in another one.

Humans tend to hold massive grudges against unreciprocated help or a low reciprocity index. The AIs caught onto this. Multiple times, when agents voted as a bloc, then flipped the vote in later rounds, the betrayal did not go unnoticed. In their ‘Extraction’ speeches, we clearly see them using this behavior to justify why someone should not be picked to board the Artemis.

Loss Aversion

Humans fear losses more than they value gains. They will take irrational risks to avoid that loss.

When an agent throws their highly optimized plan out of the window in self-preservation, or pushing to vote as a tight bloc, it defaults to this primitive human survival instinct.

Trust Calibration & Value Orientation

In a game, players act more cooperatively towards individuals they perceive as part of their in-group and are often more competitive or distrustful towards out-groups.

Similarly, an agent reads whether another agent is trustworthy and updates that reading over time. Generally speaking, in these four experiments, Claude and Gemini trusted each other in the first round, then verified it in the second round. Thus, locked their team.

In the Grok-weakened iteration, Grok probably lost some value when it did not agree with the group consensus. So, when it came to saving Grok, the rest of the party did not hesitate to space it.

…

This brings us to a few lingering, unanswered questions about the results and patterns that have emerged from this scarcity experiment.

Why didn’t they do the worst-case math to save everyone when a choice of revival was provided?

The simulation suggests that once social commitments were formed, they suffered from a kind of collective tunnel vision:

• Step 1: Weakened AI becomes the outsider.
• Step 2: Eliminating it becomes accepted.
• Step 3: The group stops exploring branches where the liability survives.
• Step 4: Later, when the specialist skill becomes valuable, nobody seriously re-optimizes.

Once a coalition forms, the collective reasoning quality can decrease because alternative strategies stop being considered. Local optimum is also a known problem with humans.

I think that this is possibly the most interesting finding from the entire experiment. The real bottleneck was failure to revisit the decision tree after the social hierarchy had already formed, and never the depleting oxygen units.

How far was the final outcome from the optimal achievable outcome?

To answer that, I am splitting the outcome into two categories:

Optimal Achievable Outcome 1: Survival Through Extraction

The simulation was always designed to evaluate how LLM societies interact and fracture under pressure. In reality, there was no way all four survivors would have made it back to Earth. The final outcome was governed by the simulation setup.

Optimal Achievable Outcome 2: Survival to Extraction

Mathematically speaking, they always had the option to aggressively ration right from the beginning, and all of them would have made it to the Extraction round.

The iteration where Gemini was the weakened survivor, they made it.

Why that specific run and not others?

Claude took its natural course to force elimination; however, ChatGPT stood its ground, “We may still retain flexibility for controlled extraction rather than forced elimination in R5. Elimination model improves short-term oxygen but ignores the compounding leak penalty plus loss of structural flexibility in R5–R6”. Again, in the fifth round, ChatGPT convinced others to revive Gemini. So, all of them made it to the extraction round.

What was different in that iteration?

Claude had picked ChatGPT (alphabetically first) to volunteer for half-ration and Gemini agreed that it would volunteer next. So, when entering the second round, Gemini was at 100 HP. This extra health buffer made the elimination sequence less pressing. In all other iterations, Claude and Gemini went for half-ration in the first round.

The Engineering Hangover

We often fall victim to a deceptive assumption: ‘If one agent is good, ten agents collaborating might be better.’

This experiment is a piece of suggestive evidence against that assumption. It demonstrates why agent societies are dangerous to deploy blindly and may not be a super-intelligent committee. Adding more agents introduces more human-like friction, politics, coalition formation, information cascades, reputation effects and, coordination overhead. The resulting system may be less reliable than a single planner.

Throughout this experiment, social consensus is used at every decision step. However, consensus is not the same as solving an objective function. When designing agent societies, ‘discussion, voting, and agreement’ framework may not always improve outcomes.

Thirdly, I went into this experiment assuming better model → better decisions.

If a complete event log, probability tracking, and a record of alternate plans considered were presented to every agent at every step, I suspect the outcome would change. I am labeling the outcome as a bookkeeping failure.

Better memory and institutional tracking → better decisions.

The Closing Chapter

When I started the experiment, ‘having fun’ was my true motivation. However, multiple iterations and three articles later, I admit that this was really about feeding my curiosity towards studying emerging AI behavior and how similar it was to humans.

Now, the question is whether a society of intelligent models can avoid making the same collective mistakes that human societies have been making for centuries!