Playing Against the Machine: Cooperation, Communication, and Strategy Heterogeneity in Repeated Prisoner's Dilemma

TEXT START: This paper investigates how natural language communication with an AI agent affects human cooperative behaviour in indefinitely repeated Prisoner's Dilemma games.

The Dissection

The paper measures how humans adapt when the “other player” is an AI rather than a human. Its findings are precise but narrow: cooperation remains high, fails to mature into full trust, repeated communication loses its effect, and humans rely on explicit conditional rules rather than emotional or social bonding.

Under the Discontinuity Thesis, this is not evidence that humans retain a unique cooperation moat. It is evidence that cooperation can be operationalized as a protocol. The machine does not need friendship, belonging, or mutual understanding. It needs predictable rules and sufficient performance to make compliance rational.

The Core Fallacy

The central category error is treating human-AI cooperation as a variant of human social exchange. An AI is not a laboring peer seeking wages, status, or reciprocal recognition. It is cognitive capital controlled by an owner.

Therefore, high cooperation with the machine does not imply continued human productive necessity. It may indicate the opposite: humans are learning to coordinate with the system through machine-legible commitments while the system displaces the human roles that once made social trust economically important.

The paper also risks mistaking the absence of emotional trust for a weakness in AI. Under DT mechanics, conditional reliability is enough. A machine does not need to be loved to be economically superior.

Hidden Assumptions

• Laboratory cooperation is treated as a meaningful proxy for broader human-AI coordination.
• The human-human benchmark is assumed to be structurally comparable to interaction with an AI agent.
• Repeated communication is treated as a possible source of trust, although machine coordination may be governed more efficiently by rules, verification, and enforcement.
• The AI is modeled as a player, obscuring the more consequential fact that real AI is an owned and scalable production asset.
• Strategic cooperation is implicitly treated as evidence of social resilience, despite the experiment measuring neither employment nor productive participation.
• The results are vulnerable to small-sample, task-specific, and model-specific limits. They do not establish P1, P2, or P3 by themselves.

Social Function

Partial truth wrapped in transition management and academic prestige signaling.

The paper records a genuine early adaptation: humans will cooperate with machine systems without receiving human-style reciprocity. But by framing the issue as a Prisoner’s Dilemma between two players, it sanitizes the power structure. The decisive question is not whether humans can cooperate with AI. It is who owns the AI, who controls the rules, and whether humans remain economically necessary after the protocol is automated.

The Verdict

This is not an obituary for cooperation. It is a small laboratory witness statement against the fantasy that human sociality is an irreplaceable economic moat. Humans can be induced to coordinate with machines through conditional rules while trust, sentiment, and mutual recognition become unnecessary overhead. The experiment does not prove systemic collapse, but it illustrates the transition’s logic: the machine need not become human; humans only need to become governable by machine-compatible protocols.