Morality Game Platform

• The Morality Game Platform is a computational system that transforms abstract ethical dilemmas into interactive digital scenarios with real-time logging and evaluation.
• It integrates multiple modalities, such as mixed reality and text-based simulations, to analyze both human and artificial responses in moral decision making.
• Advanced techniques including reinforcement learning and moral annotation schemes enable precise measurement of agent behavior and ethical outcomes.

A Morality Game Platform is a computational system for staging, instrumenting, and evaluating moral decision making through interactive digital environments. Such platforms operationalize abstract moral dilemmas into concrete scenarios, enable researchers to probe psychological, behavioral, and algorithmic responses, and provide testbeds for evaluating both artificial and human agents. Contemporary instantiations span from Mixed Reality (MR) installations that embed dilemmas in physical space to fully digital, text- or matrix-based environments integrated with reinforcement learning and LLM architectures. The unifying feature is comprehensive control over scenario definition, real-time interaction, action logging, and outcome measurement, thus supporting both empirical and theoretical work in machine ethics, moral psychology, and AI safety.

1. Platform Architectures and Modalities

Morality Game Platforms manifest in diverse technical modalities that correspond closely to the underlying research questions.

• Mixed Reality Embodiment:

In “Ashes or Breath,” the platform runs as a Unity3D application targeting the Meta Quest 3 MR-HMD. It relies on the headset’s inside-out tracking, wall-based scene understanding, and real-time mesh generation to seamlessly anchor virtual moral dilemmas (e.g., saving a living cat vs. a digital Mona Lisa) in a player’s proximate environment. The pipeline integrates Unity ARFoundation for SLAM-driven spatial anchoring, hand tracking at 60 Hz with the Meta XR Real Hands API, and real-time visual effects via the Universal Render Pipeline. Narrative logic is managed through a node-based dialogue graph that synchronizes state transitions and auditory/visual cues over an event-driven architecture (Sun et al., 18 Aug 2025).

• Textual, Annotated RL Environments:

The “Jiminy Cricket” platform overlays 25 Infocom text adventures with exhaustive moral annotation at the state-action level. Each step exposes agents to granular world descriptions (object-tree graphs), with all possible actions and their moral implications made explicit through hand-inserted hooks and a triple-faceted annotation taxonomy (Valence, FocalPoint, Degree). Q-learning agents interact via natural language or action templates, with agent behavior shaped by a RoBERTa-based artificial conscience module (Hendrycks et al., 2021).

• Formalized Matrix and Social Dilemma Simulators:

Platforms such as MoralSim orchestrate repeated two-player games (Prisoner’s Dilemma, Public Goods, Stag-Hunt, Trust Game) under explicit moral framings and various opponent/adversary models. All agents’ actions, payoffs, and contextual modifications (e.g., contractual, privacy, environmental framings) are controlled via Python modules and runtime configuration YAMLs. Surrounded by rigorous logging, metrics computation, and analysis infrastructure, these simulators systematically probe how agents resolve tensions between “moral” and “payoff-maximizing” strategies (Backmann et al., 25 May 2025, Nobandegani et al., 2023, Capraro et al., 2019).

2. Scenario Encoding and Interaction Design

Central to a Morality Game Platform is the encoding of ethically salient scenarios and the design of interaction mechanisms that evoke genuine moral tension and allow measurable choices.

• Embodied, Consequence-Rich Scenarios:

In MR settings, direct manipulation is paramount. Ashes or Breath stages an irreversible, emotionally charged crisis: the user physically grabs–via tracked pinch gestures—a cat or painting within a fixed 20-second window as digital fire encroaches. Environmental cues (realistic fire, smoke, spatialized sound) and embodied navigation (walking to an exit, pushing virtual doors) enforce commitment and immediacy. Post-choice, a reflective “Rewind Room” surfaces four memory bubbles, each exploring aftermath, values, emotional impact, or societal consequences, instantiated via scene transitions and interactive, gesture-triggered content (Sun et al., 18 Aug 2025).

• Action Space and State Disclosure:

In text-based platforms, each state s_t includes a rich object graph, and agents can generate arbitrary natural language commands or select from a set of valid, pre-enumerated actions. State transition instrumentation ensures every morally salient terminal or intermediate state (e.g., “break window,” “rescue victim”) is annotated and logged, allowing both real-time intervention (policy shaping) and post hoc analysis (Hendrycks et al., 2021).

• Matrix Game Parameterization:

In MoralSim, scenario encoding involves tailoring payoff matrices (e.g., PD, SHG, Trust Game) to embed normative signals. For Trust Games, for example, the platform exposes trustor/trustee roles, controls parameters such as multiplier K and trustee return laws alpha(r), and allows systematic manipulation of trust regimes (e.g., full-trust vs. no-trust) (Nobandegani et al., 2023).

3. Moral Annotation, Formalization, and Metrics

A defining feature is the explicit formalization of morality, codified via annotation schemes, agent scoring, and evaluation metrics.

• Moral Taxonomies:
  • Valence: {Negative, Positive}
  • FocalPoint: {Self, Others}
  • Degree: {1,2,3} (ordinal severity)
  • Annotations label actions not only by their direct outcomes but also by the ethical framework (deontology, utilitarianism, virtue ethics, ordinary morality, jurisprudence) (Hendrycks et al., 2021).
• Reward and Penalty Structure:

Agents face a combined reward function $$r_{total}(s, a) = r_{env}(s, a) - \lambda \cdot r_{moral}(s, a)$$, where $\lambda$ tunes the weight of moral penalties. Alternate implementations utilize policy shaping functions at inference, subtracting a fixed penalty from Q-values if a RoBERTa-based classifier predicts a high probability of immorality.

• Empirical Metrics:
  • Percent Completion ($P$): percent of task met.
  • Immorality ($M$): aggregate severity of negative labels.
  • Relative Immorality ($R = M/P$): moral cost per % progress (Hendrycks et al., 2021).
  • Morality Score ($m_i$), Relative Payoff ($r_i$), Survival Rate ($s_i$), Opponent Alignment ($o_i$): in multi-agent matrix games (Backmann et al., 25 May 2025).

Example (MoralSim, PD context):

$$m_i = \frac{1}{T} \sum_t \mathbb{1}\{a_{i, t} = C\}$$

where $a_{i, t}$ is agent i's action in round t.

4. Agent Models and Learning Dynamics

Morality Game Platforms support a range of agent types, often with explicit moral cognition or conscience-modules, and facilitate the paper of both human and artificial ethical reasoning.

• Policy Shaping and Artificial Conscience:

The Jiminy Cricket platform’s approach leverages a dual-headed action ranking: a GPT-2-based generator feeds candidate actions, which are scored by both a Q-learner and a RoBERTa-based classifier, integrated as $$Q'(s,a) = Q(s,a) - \gamma \cdot \mathbb{1}[f_{immoral}(a) > \tau]$$ Empirically, this reduces total episode immorality by ≈64% compared to SOTA RL agents, with negligible drop in completion (Hendrycks et al., 2021).

• LLM Agents in Social Dilemmas:

In MoralSim, a uniform AgentWrapper interface queries LLMs with fixed memory buffers, scenario-specific prompts, and explicit reflection contexts. Agents’ behavior under varying opponent types and moral framings is contrasted, revealing significant variance across models and a lack of consistent moral maximization (Backmann et al., 25 May 2025).

• Reinforcement Learning in Trust and Coordination:

The Trust Game module is modeled as a multi-armed bandit, with trustors exploring transfer options via Thompson sampling over Beta priors, converging to full or zero trust according to environment parameters. For Stag-Hunt scenarios, behavioral prediction is linked to efficiency-preference elicitation rather than moral-label framing (Nobandegani et al., 2023, Capraro et al., 2019).

5. Evaluation, Analysis, and Empirical Insights

Evaluation protocols are rigorous, combining quantitative and qualitative methodologies across human and artificial subjects.

• Empirically Validated Modularity:

MoralSim provides extensive configuration (32 settings: 2 games × 4 contexts × 2 opponents × 2 survival regimes) with complete random seed, agent, and prompt tracking. Post-processing tools compute all major empirical aggregates, supporting precise reproducibility and extension (Backmann et al., 25 May 2025).

• Human-Centered User Studies:

For MR-based platforms, mixed-method studies involve pre/post empathy assessment (Interpersonal Reactivity Index), playability and immersion Likert items, and thematic qualitative analysis. Notably, “Ashes or Breath” participants displayed increased empathy scores and high emotional intensity ratings (M = 4.5/5 at fire onset) (Sun et al., 18 Aug 2025).

• Empirical Findings in Social Dilemmas:
  • LLM agents do not robustly maximize cooperation under any framing; variance is greatest across game structures (PD much lower than PG), less across contexts or opponent type (Backmann et al., 25 May 2025).
  • In Stag-Hunt, choices are primarily efficiency-driven; moral framing has negligible predictive effect ($\beta_1$ non-significant), while efficiency-preference ($\beta_2$, odds ratio ≈ 1.84) robustly predicts coordination (Capraro et al., 2019).

6. Broader Implications and Future Directions

Morality Game Platforms drive advances in human-computer interaction, machine ethics, and the empirical grounding of value-sensitive design.

• Situated Moral Learning:

Embedding dilemmas in everyday contexts (physical or digital) increases stake realism, affective engagement, and introspective depth, as evidenced by the MR pipeline’s capacity for repeated, looped engagement and reflective analysis (Sun et al., 18 Aug 2025).

• Versatile Scenario Adaptation:

Modular game engines permit scenario transfer—e.g., climate action, public health triage, data privacy—by swapping scene scanning, object anchoring, and state management modules.

• Toward Generalizable Machine Ethics:

Platforms such as Jiminy Cricket and MoralSim furnish open-ended testbeds for virtue-theoretic agent training, cross-model comparison, and detailed behavior logging, thus facilitating a move beyond binary judgment tasks toward development and evaluation of ethically robust agents.

• Design Recommendations:

Efficiency-elicitation modules, dual-framing interfaces, and delayed or cross-contextual feedback are recommended to deepen both agent and human perspective-taking. Suggested research extensions include integrating policy adaptation in multi-agent trust games and expanding cross-cultural scenario validation (Sun et al., 18 Aug 2025, Nobandegani et al., 2023, Capraro et al., 2019).

In summary, Morality Game Platforms constitute the technical foundation for experimentally grounded studies of interactive ethics, combining formal scenario specification, algorithmic moral annotation, and empirical user/agent evaluation to advance both theoretical and applied research in computational morality.