Mathematical Modelling of Ethical AI Use in Higher Education: A Coordination Game Framework for Future-Facing Learning

Synthesis: Ogbo, Song, Ghareeb, and Han (2026) provide a formal, mechanism-level account of why policy statements alone fail to change student AI-use behavior. Reframing student AI use in assessments as a coordination problem shaped by peer expectations and assessment design rather than individual compliance, they develop an evolutionary game-theoretic framework capturing learning value, effort, perceived fairness, and transparency, with institutional AI governance modeled implicitly through reflective assessment incentives. Using analytical results and finite-population simulations, they reveal threshold-driven behavioral transitions: small, well-calibrated changes to reflective assessment incentives can trigger rapid cohort-wide shifts toward responsible AI use, while weak or misaligned incentives allow opportunistic practices to persist.

AI Use as a Coordination Problem

Rather than treating student AI use as a matter of individual compliance, this framework models it as a collective norm-formation process. Each assessment task is a one-shot decision context in which a student chooses how to engage with AI tools, but those choices interact with peer expectations to generate shared norms within a cohort. Institutional influence enters implicitly through assessment design — the signals assessment tasks convey about acceptable AI use. This norm-based perspective is grounded in evidence that assessment design strongly shapes learner behavior and that responsible AI engagement should be framed in terms of learning and ethics rather than narrow compliance.

The Evolutionary Game Model

The framework uses evolutionary game theory in finite populations to analyze how peer interaction, assessment incentives, and social learning jointly shape the emergence and stability of AI-use norms over time. The model captures four key factors:

• Learning value — the additional learning benefit of engaging responsibly with AI versus relying on it opportunistically.
• Effort — the cognitive and time cost of responsible engagement (e.g., reflecting on one's AI interactions).
• Perceived fairness — whether students see assessment structures and incentives as fair and legitimate.
• Transparency — the clarity of norms and expectations around AI use.

Institutional AI governance is modeled implicitly through reflective assessment incentives — rewards for activities like requiring students to reflect on their AI interactions.

Key Findings: Threshold-Driven Norm Transitions

Across the model results, a consistent pattern emerges: responsible AI use does not increase smoothly with incentives, but arises through threshold-driven transitions.

1. Reflection must be rewarded enough. Embedding reflection as a meaningful, rewarded component of assessment can displace opportunistic AI use, but only once a critical reward level is reached. Below this threshold, opportunistic behavior remains dominant — reflecting strategic responses to assessment structures rather than a lack of ethical intent.
2. Peer sensitivity drives cascade speed. When students are highly responsive to peer practices, modest changes in assessment incentives can trigger rapid norm cascades; weak peer sensitivity leads to gradual, incomplete transitions. This explains why perceptions of peer behavior strongly influence the legitimacy of AI use.
3. Proportionality matters. Responsible AI use emerges most robustly when the reward for reflection is proportionate to the effort required. When reflective tasks impose high cognitive or time demands without sufficient assessment value, opportunistic AI use persists even under generous incentives — a formal explanation for why well-intentioned but overly burdensome assessment interventions fail.

What this means for practice

• Instructors. Make reflection a scored component of the assessment rather than an optional add-on: in the model, responsible AI use stays rare until the reflection reward crosses a critical threshold near r ≈ 1.5, after which it rapidly dominates the cohort.
• Instructors. Calibrate the reward to the effort. When reflective work costs more than it earns in assessment value, opportunistic AI use remains the stable outcome even with generous incentives attached.
• Instructors. Assess reflection for substance, not presence: symbolic or minimally embedded reflection stays at consistently low frequency across the whole incentive range, so superficial submissions should not earn the reward.
• Administrators. Expect threshold effects rather than proportional ones. Peer sensitivity determines the speed of change — a sharp transition at β = 0.5 versus a gradual, incomplete shift at β = 0.01 — so interventions that shift visible peer practice can accelerate norm change far faster than policy statements.
• Administrators. Prefer assessment redesign over surveillance when justifying institutional responses, since the model treats student behavior as a strategic response to assessment structures and perceived fairness rather than to stated rules.

Limitations

• The results are analytical and simulation-based: conclusions rest on a finite-population model with N = 100 under Fermi imitation updating, and the authors state that empirical validation through surveys, behavioral experiments, or interactive assessment settings is still needed to ground the predictions.
• Parameters are illustrative rather than estimated from data: payoffs a = 1, b = 0, c = 1, d = 2, reflection effort cost κ = 1, superficial reflection factor σ = 0.4, legitimacy cost δ = 1, and misuse penalty τ = 1, with peer sensitivity β = 0.1 in the baseline (probed at 0.01 and 0.5). Quantitative thresholds such as r ≈ 1.5 are therefore model-relative, not measured classroom values.
• The model compresses student AI use into four factors — learning value, effort, perceived fairness, and transparency — plus one reflection mechanism, so it cannot test other plausible drivers such as disciplinary conventions or detection pressure.
• Peer influence is represented by a single sensitivity parameter instead of measured student networks, so the cascade dynamics are not calibrated against observed cohort structure.

Citation

Ogbo, N. B., Song, Z., Ghareeb, S., & Han, T. A. (2026). Mathematical Modelling of Ethical AI Use in Higher Education: A Coordination Game Framework for Future-Facing Learning. arXiv preprint.