Cooperative game theory is a branch of game theory that focuses on situations where players can form coalitions and cooperate to achieve common goals. The Prisoner's Dilemma is a classic example in game theory that illustrates a situation where rational individuals acting in their own self-interest can lead to a suboptimal outcome for everyone involved. While the Prisoner's Dilemma is typically framed as a non-cooperative game, it can still be discussed in the context of cooperative game theory.
In the Prisoner's Dilemma, two suspects are arrested and accused of committing a crime together. They are placed in separate interrogation rooms and offered a deal by the prosecutor:
- If both suspects remain silent (cooperate with each other), they will each serve a short sentence of one year for a minor offense.
- If one suspect confesses (defects) while the other remains silent (cooperates), the defector will be released, and the cooperator will serve a longer sentence of ten years.
- If both suspects confess (defect), they will each serve a moderate sentence of five years.
Let's observe how this scenario plays out in the Human versus GenAI game. As previously mentioned, there are two participants: humans and GenAI. Each participant can select between two actions: "Depend" or "Not Depend." From the human perspective, choosing "Depend" implies relying on GenAI for certain tasks, while opting for "Not Depend" means not relying on GenAI.
Hence, there are four combinations of actions that can happen during the game tenure:
- Imagine if humans opt not to rely on GenAI, and simultaneously, GenAI chooses not to depend on humans: GenAI can enhance human activities, yet if humans choose not to rely on it, they may struggle to compete in the market or even incur losses. Similarly, GenAI relies on human-generated data for training, and if it opts not to depend on humans, it leads to an inadequately trained or even fictitious model. Consequently, both parties suffer losses. If we denote each party's loss as '-1', then the total loss in the game sums up to -2. :(
- Consider a scenario where humans choose to rely on GenAI, while GenAI, in turn, decides not to depend on humans: In one scenario, the GenAI model chooses not to rely on human-generated data and instead trains itself on hallucinated or synthetic data, which is inherently sub-optimal. Humans who depend on such AI models may not experience any significant gains, although occasionally a model trained on synthetic data might prove beneficial. Therefore, the overall gain in this scenario is slightly less than -1.5. In another case, suppose GenAI has acquired comprehensive knowledge and no longer requires human data, but humans still wish to depend on AI. While GenAI experiences a loss by missing out on the creative input of humans, humans will leverage GenAI for their benefits. In this scenario, the gain in the game totals 0.5 (from the perspective of GenAI) plus 1 (from the perspective of humans), resulting in a combined gain of 1.5.
- Imagine a scenario where GenAI chooses to rely on humans, but humans opt not to depend on GenAI: In this situation, GenAI benefits from being trained on human-generated data. However, if humans choose not to utilize such a trained model, they incur a loss. Therefore, the total gain would be 1 (from GenAI's perspective) plus (-1) (from humans' perspective), resulting in a net gain of 0, indicating no overall gain.
- Consider a scenario where GenAI opts to rely on humans, and in turn, humans choose to depend on GenAI: At this Nash-equilibrium point, humans benefit from a well-trained AI model, while GenAI gains from the wealth of creative human data. Consequently, the total gain in the game amounts to 1 (from humans' perspective) plus 1 (from GenAI's perspective), resulting in a combined gain of 2.
Therefore, in this cooperative game, mutual support and collaboration yield synergistic outcomes, leading to mutual benefits instead of engaging in rivalry. However, over time, there's a possibility that GenAI comprehends human actions (due to a discount factor) and data entirely, eventually becoming independent of humans. This scenario points to an application domain where GenAI surpasses humans and reaches a state of saturation, being capable of performing almost all tasks a human can. In such a situation, the game deviates from the Nash Equilibrium, resulting in a lower overall game gain. Hence, it is advisable for humans to transition from these areas to niche domains where mutual dependencies can coexist.
[1] Human vs. Generative AI in Content Creation Competition: Symbiosis or Conflict?, https://arxiv.org/abs/2402.15467.
