Evolution of Fairness in the Not Quite Ultimatum Game

The Ultimatum Game (UG) is an economic game where two players (proposer and responder) decide how to split a certain amount of money. While traditional economic theories based on rational decision making predict that the proposer should make a minimal offer and the responder should accept it, human subjects tend to behave more fairly in UG. Previous studies suggested that extra information such as reputation, empathy, or spatial structure is needed for fairness to evolve in UG. Here we show that fairness can evolve without additional information if players make decisions probabilistically and may continue interactions when the offer is rejected, which we call the Not Quite Ultimatum Game (NQUG). Evolutionary simulations of NQUG showed that the probabilistic decision making contributes to the increase of proposers’ offer amounts to avoid rejection, while the repetition of the game works to responders’ advantage because they can wait until a good offer comes. These simple extensions greatly promote evolution of fairness in both proposers’ offers and responders’ acceptance thresholds.

💡 Research Summary

The paper tackles the long‑standing puzzle of why human participants in the Ultimatum Game (UG) tend to make fair offers and reject low ones, despite the predictions of classical rational‑choice theory. Rather than invoking external mechanisms such as reputation, empathy, or spatial structure, the authors propose a minimal extension of the UG that they call the Not‑Quite Ultimatum Game (NQUG). Two key modifications define NQUG. First, both proposers and responders make their decisions probabilistically. Each proposer draws an offer from a normal distribution with mean μp and standard deviation σp, while each responder draws an acceptance threshold from a normal distribution with mean μr and standard deviation σr. The parameters σp and σr capture the intrinsic stochasticity of human decision‑making, reflecting bounded rationality, mood fluctuations, or noisy perception. Second, the game is allowed to continue after a rejection with probability ρ. If a proposal is rejected, the same pair may play another round rather than the interaction ending permanently. This introduces a “wait‑and‑see” opportunity for responders and a risk of repeated loss for proposers.

To explore the evolutionary consequences of these features, the authors implement an agent‑based evolutionary simulation. An initial population of agents is randomly assigned values for (μp, σp, μr, σr). In each generation every possible pair of agents plays NQUG. If an offer is accepted, both participants receive the offered amount; if it is rejected, both receive zero, but with probability ρ the pair repeats the interaction, sampling new offers and thresholds each time. An agent’s fitness is the total payoff accumulated over all its encounters. Reproduction follows a standard Wright‑Fisher process with selection proportional to fitness, plus mutation that perturbs the four parameters by small Gaussian noise.

The simulation results reveal three distinct regimes. When σp and σr are near zero (deterministic behavior), the dynamics reproduce the classic UG outcome: proposers converge to the minimal feasible offer, and responders accept any positive amount. As σ increases to moderate levels (≈0.1–0.3), proposers begin to raise μp. The stochastic chance of rejection creates a selective pressure for a safety margin; a higher average offer reduces the probability of costly rejection and therefore yields higher expected fitness. The effect is amplified when ρ is also high. With ρ ≥ 0.5, responders can afford to set a higher μr because they anticipate future rounds where a better offer may appear. In this high‑ρ, high‑σ regime both μp and μr drift upward until the average offer stabilizes around 0.5 of the total pie, which is close to the empirically observed human average in laboratory UG experiments.

These findings have several important implications. First, they demonstrate that fairness can evolve without any explicit social information or network effects; the combination of decision noise and the possibility of repeated interaction alone is sufficient to generate mutually beneficial, relatively equitable outcomes. Second, the model provides a mechanistic link between two psychological phenomena: risk‑aversion (captured by σ) and the expectation of future opportunities (captured by ρ). Proposers mitigate risk by offering more, while responders exploit the prospect of future rounds to demand more. Third, the authors argue that NQUG offers a parsimonious baseline against which more elaborate models (e.g., reputation systems, empathy cues) can be compared.

The paper concludes by suggesting empirical avenues for validation. Laboratory experiments could manipulate the perceived variability of offers (e.g., by adding noise to the decision environment) and the likelihood of a second chance after rejection, thereby testing the predicted shifts in average offers and acceptance thresholds. Additionally, extending the model to structured populations or incorporating learning dynamics could illuminate how the simple mechanisms identified here interact with richer social contexts. Overall, the study provides a compelling, mathematically grounded explanation for the emergence of fairness in bargaining situations, emphasizing that even minimal stochasticity and repeatability can reshape evolutionary incentives toward more equitable behavior.