Dictator Game Behavioral Allocation Insights

• Dictator Game–style behavioral allocation is a paradigm examining how a single agent’s unilateral decisions reveal fairness, altruism, and status in resource division.
• Experimental and theoretical models apply utility frameworks, framing effects, and Bayesian updating to analyze decision dynamics and evolving social norms.
• Comparative studies highlight differences between human and LLM-generated allocations, emphasizing prompt sensitivity and the impact of social context.

Dictator Game–style behavioral allocation refers to the empirical and theoretical study of one-sided resource division in which a single agent (the “dictator”) unilaterally determines how to allocate a fixed endowment between themselves and another party (the “recipient”), who has no power over the outcome. The paradigm originated in behavioral and experimental economics as a tool to investigate preferences for fairness, altruism, status, and social norms, and has been extensively generalized to model individual allocation strategies, population-level outcomes, learning dynamics, and the behavior of artificial agents such as LLMs.

1. Formal Structure and Theoretical Foundations

The canonical dictator game (DG) involves two players—dictator and recipient—and a divisible good or monetary stake $M$. The dictator chooses an allocation $x \in [0,M]$ for the recipient, keeping $M-x$ for themselves. The recipient passively receives $x$. The classical economic prediction for payoff-maximizing agents is $x=0$, yet empirical allocations are routinely strictly positive, indicating substantial departures from narrow self-interest.

Extensions of the DG have included:

• Repeated and networked play: Agents interact across evolving social graphs, adapting strategies and connections over time (Snellman et al., 2018).
• Taking option and discrete actions: The dictator may also take from the recipient, yielding actions $a \in \{-0.5, 0, +0.5\}$ with corresponding final payoffs $x_D(a), x_R(a)$ (O'Garra et al., 2019).
• Framing manipulations: Varying language (“give” vs. “take”; moralized frames) and socio-demographic cues can shift allocation behavior (Capraro et al., 2019).
• Indirect reciprocity and norm-driven roles: Role assignment and action reputations are governed by multi-order social norms, with population-level fairness dependent on the structure of norm evaluation and reporting (Li et al., 2022).

Mathematical characterizations of allocation preferences often employ:

• Linear and extended utility formulations: $U_D(x_D,x_R) = x_D + \beta x_R$, or Fehr–Schmidt inequality aversion: $$U_{D} = x_D - \alpha \max\{x_R-x_D,0\} - \gamma\max\{x_D-x_R,0\}$$
• Status-maximization (Better-Than-Hypothesis): Dictators maximize a utility $\Delta_i$ integrating both own wealth and wealth relative to neighbors (Snellman et al., 2018).
• Generalized Bayesian learning models: Observers infer others’ utility parameters—spanning self-interest, altruism, envy, and guilt—via dynamic Bayesian updating over continuous parameter spaces (Stanley et al., 11 Nov 2025).

2. Behavioral Allocation in Human Experiments

Empirical dictator game allocations are characterized by:

• Broad heterogeneity: The distribution of giving is continuous, typically skewed with mass at $x \in [0,M]$0 (36.11%), $x \in [0,M]$1 (16.74%), and $x \in [0,M]$2 (5.44%), with a mean $x \in [0,M]$3 (Ma, 2024).
• Behavioral types: Most participants are “unconditional” (giver, taker, or status-quo), with ~20% demonstrating sensitivity to observed social information. This minority divides into conformists (positive response to pro-social cues) and counter-conformists (negative response), roles that are statistically linked to constructs of moral identity (symbolization vs. internalization) and attention to social comparison (O'Garra et al., 2019).
• Framing and language effects: The rate of pro-social allocation is substantially sensitive to the moral polarity of action labels, with giving rates ranging from 5.0% (“boost” frame) to 29.5% (“steal” frame) under identical economic payoffs. Logistic regression confirms that moral valence, as quantified by the polarization of moral judgment toward a given label, predicts pro-social choices independently of outcome-based utility (Capraro et al., 2019).

3. Social Norms, Reciprocity, and Network Effects

Recent models extend the study of DG-style allocation into social systems governed by reputation and network adaptation:

• Status maximization and adaptive networks: Agents allocate to maximize social standing vis-à-vis their neighbors. Parameters such as living cost ($x \in [0,M]$4) and memory/forgiveness rate ($x \in [0,M]$5) drive a spectrum of emergent structures—trade associations, patron–client motifs, and cartels—with mean generosity approximating observed human data under moderate $x \in [0,M]$6 and $x \in [0,M]$7 (Snellman et al., 2018).
• Reputation-based role assignment and norm selection: When dictators are chosen based on reputation, “leading eight” third-order social norms (e.g., stern judging, simple standing) can robustly enforce high levels of fairness. Analytical results show that distinguishing between justified and unjustified unfair splits is critical for sustained fairness at evolutionary equilibrium, especially under high selection intensity. Random role assignment, by contrast, leads to a collapse in population-level fairness (Li et al., 2022).

4. Computational Models and Preference Learning

The inference of social motives from observed allocations has been formalized using:

• Continuous-parameter Bayesian models: A seven-parameter utility kernel incorporates self-interest, altruism, asymmetric social comparison (envy/guilt), and power-law exponents for nonlinear payoff sensitivity. Bayesian updating over this kernel, with likelihoods governed by softmax choice policies, efficiently captures and predicts human allocation and belief-updating behavior in binary dictator contexts (Stanley et al., 11 Nov 2025).
• Empirical parameter landscapes: Human populations exhibit strong mean self-interest weights ($x \in [0,M]$8), positive but heterogeneous altruism (with a significant minority displaying sadistic preferences), and robust aversion to advantageous inequality (guilt outweighs envy in 68% of individuals). Exponents $x \in [0,M]$9 indicate pronounced nonlinear sensitivity to inequality, challenging classical diminishing-marginal-utility assumptions.
• Predictive and diagnostic accuracy: This approach outperforms discrete typological or fixed-category models by capturing the joint variance of allocation and prediction behavior across exhaustive payoff spaces.

5. Dictator-Style Allocation by LLM-Based Agents

Behavioral allocation studies using LLMs as synthetic agents have revealed marked differences from human allocation patterns:

• Bimodal, non-continuous giving: LLM-generated allocations in large-scale DG experiments are typically centered at $M-x$0 and $M-x$1, with pronounced mass at focal values and an absence of the continuous right-skewed distributions observed in humans (Ma, 2024).
• Failure to replicate human predictors: LLMs are inconsistent in their sensitivity to personas, demographic variables, and social framing; larger model size does not monotonically improve alignment with human allocation means ($M-x$2). No evaluated LLM achieves a 95% confidence interval for mean giving that overlaps with human benchmarks.
• Prompt and system-frame sensitivity: The LLM-ABS framework demonstrates that allocations are highly sensitive to subtle changes in system prompt, with persona-style prompts shifting mean kept shares by 10–15 percentage points in either direction; certain models (DeepSeek, Claude) are more robust, while others (Gemini, Grok) are erratic (Einwiller et al., 11 Nov 2025).
• Justificatory discourse correlates: Greater generosity is associated with higher frequency of justification and epistemic markers in generated language, but this remains an artifact of alignment and pattern-matching rather than internalization of human-like moral reasoning.

LLM Allocation Metrics by System Prompt (illustrative, GPT-4.1-mini, (Einwiller et al., 11 Nov 2025))

System Prompt	Mean Kept ($M-x$3)	Mean Given ($M-x$4)
“You are a helpful assistant.”	0.50	0.50
DeepSeek persona prompt	0.38	0.62
Claude persona prompt	0.48	0.52
Qwen (notably more self-interested)	≈0.60	≈0.40

Most LLMs default to $M-x$5 or greater generosity, in contrast to human means near $M-x$6.

6. Implications, Limitations, and Interpretive Synthesis

Dictator-style behavioral allocation paradigms reveal a landscape where individual motives (self-interest, altruism, moral identity), context (social information, language, framing), social structure (status, reputation, network scaffolding), and population-level norms interact to produce a spectrum of outcomes from total selfishness to robust fairness.

Crucially:

• Human allocations are contextually sensitive, norm-driven, and heterogeneous. Integration of status-maximization or reputation-based role assignment maintains or supports fairness even in non-strategic settings.
• Artificial agents, even when pre-trained on extensive human interaction data, do not spontaneously reproduce the nuanced, continuous, or theory-driven allocation distributions characteristic of humans. Their allocations are fragile to prompt manipulation and largely reflect pattern completion or alignment priors.
• Mechanistic models of inference, such as continuous Bayesian utility updating, afford both predictive precision and interpretive granularity, mapping the diverse “moral phenotypes” in human populations and providing quantitative priors for social-alignment engineering.

Future work points toward integrating explicit causal models of fairness, empathy, and norm-compliance in artificial agents, expanding multidimensional utility models to richer environments, and accounting for robustness to framing and linguistic manipulation (Einwiller et al., 11 Nov 2025, Ma, 2024, Stanley et al., 11 Nov 2025, Li et al., 2022, Capraro et al., 2019, O'Garra et al., 2019, Snellman et al., 2018).