Engineering a social contract: Rawlsian distributive justice through algorithmic game theory and artificial intelligence

Hutan Ashrafian asks whether AI and game theory could help construct a social contract informed by Rawlsian distributive justice. The article proposes a way to model policy choices and life prospects, rather than reporting a trained system, implemented government program or empirical comparison of policies.^[1]

Whose prospects should a policy improve?

The article presents basic freedoms and equal opportunity as prior requirements, then interprets the difference principle through maximin: evaluate a distribution by its least-advantaged member’s position and choose the best such minimum. On this account, inequality can be acceptable when it improves that position; equal shares and maximizing total welfare are different objectives. The Veil of Ignorance supplies the perspective from which a chooser does not know their own eventual social position.^[1]

Ashrafian opens with Aipotue, an explicitly fictional island nation whose president delegates constitutional design to AI. It illustrates the ambition, not evidence that a country has done this. The proposal is to use population data to simulate alternatives without giving the decision procedure a privileged personal stake in one outcome. That absence of a personal stake does not establish that the data, objective or resulting policy is impartial.^[1]

Turning life prospects into a model

The life-course proposal maps trajectories through education, work, health, family and citizenship. Health measures such as disability-adjusted life years could make some consequences comparable. An illustrative routing game contrasts a route with fixed cost and a shared route whose cost depends on how many people use it. The purpose is to show how individually attractive choices and a collectively chosen allocation can differ; it does not validate a numerical gain for actual lives.^[1]

The article sketches route-search, optimization and game-theoretic tools for comparing these possibilities. It also discusses representing the worst-off and the wider population as players in a stylized game. These are candidate modeling approaches. A mathematical result about a specified game does not supply the missing population model, determine which outcomes count as a good life, or prove that the proposed policy-selection process realizes justice.^[1]

The choices left to people

The author asks who selects acceptable life courses, whether a representative council should do so, and whether humans retain a veto. He also argues that populations should be able to choose and change the underlying algorithm and its application. This leaves constitutional authority and revision as explicit governance questions, rather than consequences of an optimizer’s output.^[1]

The acknowledged practical conditions are substantial: representative linked data, protections for privacy and security, ways to address biased data and interpretation, and infrastructure capable of modeling life courses. The conclusion calls population-scale governance technically uncertain with current capabilities. Neither the resource costs nor the effectiveness of the proposed safeguards is empirically established here.^[1]

For Social Alignment, the contribution is an explicit candidate criterion against which policies could be compared. It remains one interpretation of justice. Using the Veil of Ignorance to align AI systems with principles of justice instead tests people’s choices in a bounded experiment; Society-in-the-Loop: Programming the Algorithmic Social Contract emphasizes continuing stakeholder negotiation. A formal objective makes a value choice inspectable without settling who may impose it.

Historical context

The publisher records online publication of Engineering a Social Contract on December 22, 2022. The journal issue is November 2023. This milestone uses the version-of-record date, not the issue year.^[1]