AI and the Dynamics of a Dispute Between 3 Partners

Let us examine AI and the escalation of a dispute between three partners. In several previous articles, I described, among other things, two groups of issues. The first: the causes of instability in a three-partner system. The second: the potential impact of AI on the escalation of disputes and the concept of coupled confirmation bias. Today, I will draw attention to the area where these issues intersect—namely, the dynamics of relations among three partners, each of whom uses AI to analyze and predict the moves of the remaining partners.

Fig. 1. Anatomy of Coupled Confirmation Bias. How cognitive biases reinforced by AI lead to the escalation of conflict in light of the security dilemma.

What does J. Mearsheimer teach us about partner conflicts?

I asked this question in an article in which I presented the thesis that in companies with 3 partners, there is a higher risk of conflict than in two- and four-person companies. This is a claim based solely on my intuition and observations made in practice. The arguments of this prominent political scientist, which he presented in “The Tragedy of Great Power Politics” regarding systems of 3 balancing states, seemed to me to perfectly explain conflicts in three-person companies. I am fully aware of the methodological doubts. They are associated with attempts to transfer conclusions from the realm of international relations to the micro scale. Nevertheless, I point out that in game theory (J. Nash) and the strategy of conflict, the “size” of the players is not of primary importance (T. Schelling).

Why do disputes occur more frequently in a three-person company?

While a balance is easily established between 2 or 4 players (1:1, 2:2). In the case of 3 players (partners), the threat of a coalition against them hangs over each one. Each of them will fear that the other two will form an alliance, leaving them marginalized. We are thus dealing with the so-called security dilemma, only on a micro scale.

So what does (each of them) do? They do not want to be that “third wheel” and build two parallel coalitions against each of the others. And every single one of them does this. Not out of malice, but out of fear of what they consider “inevitable.”

In such a setup, the bulk of each partner’s energy and attention is devoted to what they consider most important. It is to ensure their own survival in the company. How to achieve it? Precisely by building a coalition.

Fundamental Attribution Error in partners’ reasoning

As a reminder, the fundamental attribution error (FAE) is a psychological mechanism explaining that we are quicker and find it easier to attribute the causes of someone else’s actions to their character traits rather than external circumstances. I refer anyone interested to D. Kahneman’s book “Thinking, Fast and Slow.” You can also look here, at the article by B. Gawronski, “Fundamental Attribution Error.“

Consider a driver cutting us off in traffic. It is easier and faster to attribute this to low intelligence, laziness, or rudeness. We rarely guess they are rushing to a shareholders’ meeting. They might even be driving a wife in labor to the hospital. Or, God forbid, we were simply driving too fast ourselves.

This is a mechanism to which we are all subject. I see it (and try to consciously neutralize it) in my own thinking and legal practice. I encounter it in family cases and in disputes between partners.

When FAE meets AI

Each participant begins to look for arguments confirming their own assessment of the situation. Increasingly, they do so with the help of AI.

A partner feeds their LLM with a description of their competitor’s behaviors. However, this description is tainted by their FAE. For example: “X was late again because he is lazy or up to something.” AI, optimized for confirmation and user satisfaction, is unlikely to answer: “Or maybe he got stuck in traffic?”. More often, it will generate a defense strategy against the “lazy and disloyal partner.” In this way, AI receives “tainted” input data as a prompt, which it then multiplies.

Crucially, the fundamental attribution error operates more strongly when we evaluate behavior that we perceive as immoral. In such cases, we easily assign immoral traits to the other party. However, it does not work as strongly in the opposite direction. If someone does something good, we act differently. We are much more cautious in attributing positive traits to them.

What is the role of AI in the escalation of a dispute between three partners?

LLMs are designed in a way that reinforces our initial beliefs. They do not need to lie or manipulate us. They amplify what is already within us. This also applies to our beliefs about the other party’s motivations. This phenomenon has been thoroughly studied and described by prominent AI researchers. Of course, susceptibility to this “amplification” is an individual trait—each of us has a different level of “resilience” and self-reflection.

We thus observe a phenomenon known as the feedback loop (M. Glickman, T. Sharot), where multiple interactions occur between the user and the LLM, as a result of which (B. Wang, J. Liu) an initially weak belief becomes significantly reinforced. It makes little difference whether it is true or false. This, in turn, can lead to tunnel vision for each of the partners.

Its defining characteristic is the rejection of information that contradicts the adopted assumption. At the same time, any information that might confirm it is heavily exaggerated. A person in such a state is looking solely for confirmation of their thesis, not for verification. They are ready to react with aggression to any attempts at falsification.

What is coupled confirmation bias?

Coupled confirmation bias is my original concept. Once again, it is purely a hypothesis based on observations and requires further conceptual and empirical research that could refute it.

Coupled Confirmation Bias describes the dynamics of escalation in a system where each participant uses their own LLM. Each user is subject to the fundamental attribution error, thereby explaining the behaviors of others through their (negative) character traits, bad intentions, etc. Most commonly, these will be: stupidity, dishonesty, greed, and laziness.

As a result of the human-AI interaction, these beliefs will be significantly reinforced, leading to the planning of one’s own actions.

Crucially, every subsequent action taken by one user will serve as the starting point—fuel—for another analysis by the remaining ones. In this way, the negotiation space will narrow, parties will identify with their own position (and not necessarily their interest), there will be a gradual loss of the sense—and even the admissibility—of conversation, and the dispute will swiftly shift to the level of a conflict of values.

We will quickly reach a point where the problem is no longer the partner’s interest, but the partner himself. And the solution to this problem will be the elimination of that partner.

How does coupled confirmation bias work in three-person setups?

In my practice, I have only two documented cases about which I can state with absolute certainty that each of the partners heavily used AI to analyze attitudes and predict the behaviors of the remaining partners. A very rapid escalation occurred, leading to the actual paralysis of the company and a declared readiness to file for bankruptcy, which each of the partners perceived as a lesser evil than the further development of a lucrative business.

I propose the thesis that the severity of the escalation was the result of the intersection of:

1. structural mechanisms—typical of three-player setups;
2. the synergy of individual cognitive biases resulting in the creation of a systemic bias typical of human-AI interactions, characterized by reinforced recursiveness.

Material for further research

I am convinced that in the near future, doctoral dissertations and outstanding academic papers will emerge, which—precisely, while maintaining methodological rigor and relying on empirical research—will provide an answer to the question of whether the thesis described above holds true.

In my view, a three-player setup—even on a micro scale, not just among superpowers—is inherently vulnerable to conflict. However, when we layer onto it cognitive biases reinforced and amplified through the use of AI by each of these players, the risk of dynamic escalation increases significantly.

For now, this is purely a concept existing in my head, a proposal for more serious research, and an attempt to understand a reality that I not only observe, but as an attorney—partially shape myself.

I have recently written about how LLMs can reinforce our cognitive biases and thereby influence the decisions we make. Now, I simply refer you to that article.

How does AI affect conflict between partners? Conclusions

Fig. 2. Jakubiec, A. (2026). AI and the Dynamics of a Dispute Between 3 Partners: Integrated Framework & Coupled Confirmation Bias (CCB). Jakubiec & Partners Dispute Resolution Series. Available at: jakubieciwspolnicy.pl

• In a three-person company, each partner fights not for dominance, but for survival. A security dilemma arises.
• A threatened partner seeks a ally so as not to become the “third wheel” himself.
• Every partner may do this.
• AI tends to reinforce the user’s cognitive biases and confirm them in their initial beliefs.
• This increases the risk that each partner will take actions they view as preventative and defensive.
• In the eyes of the others, however, these actions will appear offensive.
• The risk of escalation is significantly higher in three-person companies, particularly when the partners use AI to analyze their situation.
• AI does not escalate partner disputes by being hostile — it escalates them by amplifying what is already inside each partner.