Possibility Ethics: Measuring Moral Value as the Expansion of Safe Future Options in Sociotechnical Systems

Technical Summary: Possibility Ethics

Problem Statement
Sociotechnical systems (including AI, digital platforms, and institutional workflows) increasingly shape not only immediate outcomes but the entire range of viable futures available to individuals and communities. While existing ethical frameworks articulate principles such as autonomy, fairness, and capability, they generally lack a transparent, computational method to compare alternative system designs based on how they preserve or expand ethically acceptable future opportunities. Current approaches often fail to operationalize these principles into a routine audit pipeline that accounts for feasibility, harm, distributional equity, and the structural nature of "reachable futures."

Methodology: The Possibility Ethics Framework
The paper introduces Possibility Ethics, a conceptual-computational framework designed to evaluate sociotechnical systems by measuring their effect on distinct, viable, and ethically admissible future option classes. The methodology is implemented in the Omega-PE software toolkit and follows a deterministic audit pipeline:

1. Formal Definitions:

   • Viability: A trajectory is viable if it meets hard prerequisites and a feasibility score ($F$) exceeds a threshold ($\tau$). This filters out nominal options that are practically impossible (e.g., due to cost or lack of support).
   • Harm Guardrail: A trajectory is ethically admissible only if its predicted risk ($R$) to the agent, third parties, or collective goods remains below a tolerance ($\gamma$).
   • Deterministic Clustering: Admissible trajectories are mapped to outcome classes using a similarity rule and threshold ($\theta$). This prevents "menu inflation" by treating duplicate or near-identical outcomes as a single option class.
   • Option Entropy ($\Omega$): The core metric is a logarithmic opportunity index defined as:
     $$\Omega^d_i = \log(\kappa + K^d_i) - \log \kappa = \log\left(1 + \frac{K^d_i}{\kappa}\right)$$
     Where $K^d_i$ is the count of distinct admissible outcome classes for cohort $i$ under design $d$, and $\kappa$ is a smoothing parameter. This logarithmic form ensures diminishing marginal returns (the value of the 101st option is less than the 2nd) and compresses scale for cross-domain comparison.

2. Aggregation and Safeguards:

   • Vulnerability-Sensitive Aggregation: Designs are scored using a weighted sum ($L(d)$) where vulnerable cohorts receive higher weights ($V_i$).
   • Group Floors: A minimum threshold ($\phi_g$) is set for protected groups to ensure aggregate gains do not mask severe losses to the least-advantaged.
   • Cohort-Level Non-Regression: A design must not reduce a cohort's option entropy below a baseline minus a tolerance ($\epsilon_i$). Violations trigger Option Repair, a procedural requirement to identify barriers and evaluate restoration or compensation.

3. Mathematical Properties:
   The framework establishes guarded monotonicity (adding a new admissible class always increases the score), duplication invariance (adding a duplicate class does not change the score), and diminishing marginal contribution. A characterization theorem proves that under an independent-composition axiom, the logarithmic form is the unique continuous, strictly increasing measure of effective opportunity multiplicity.

Key Contributions

• Integration of Seven Elements: The paper uniquely integrates computable opportunity measurement, feasibility/viability filtering, harm guardrails, deterministic clustering, vulnerability-sensitive aggregation, non-regression safeguards, and a procedural repair mechanism into a single audit pipeline.
• Formal Characterization: It provides a rigorous mathematical foundation for "Option Entropy," distinguishing it from Shannon entropy by focusing on the combinatorial size of the state space rather than probabilistic distributions.
• Reproducible Implementation: The framework is released as the Omega-PE software artifact, including scripts for synthetic benchmarks and real-data audits, ensuring full reproducibility of results.

Results and Evaluation
The framework was evaluated using two distinct approaches:

1. Synthetic Benchmark: A controlled comparison of three designs (Baseline, Exit/Offline, Personalized Assistant) across four cohorts.

   • Finding: The "Exit/Offline" design ranked highest because it created distinct admissible pathways for cohorts with fewer baseline options. The "Personalized Assistant" ranked lower, sometimes reducing options for specific vulnerable groups despite aggregate improvements.
   • Sensitivity Analysis: The preferred design remained first across a grid of viability ($\tau$) and harm ($\gamma$) thresholds, though secondary rankings shifted, demonstrating the diagnostic utility of the framework in identifying assumption-dependent outcomes.

2. Real-Data Audit (OULAD): An observational audit of the Open University Learning Analytics Dataset (7,909 student records across four presentations).

   • Finding: The 2014J presentation showed the highest aggregate opportunity level ($L(d) = 7.773$). However, the framework detected cohort-level regression: the "disability" cohort experienced a drop in option entropy in 2014J, and the "repeat attempt" cohort regressed in 2014B.
   • Implication: These regressions triggered the Option Repair protocol, illustrating that aggregate improvements can conceal specific losses to vulnerable groups—a nuance hidden by standard aggregate metrics.

Significance and Claims
The paper positions Possibility Ethics not as a replacement for moral theory, political reasoning, or comprehensive justice frameworks, but as a transparent, reproducible, and auditable decision-support tool.

• Scope: It addresses a specific, often implicit question: "Given two candidate sociotechnical designs, which one preserves or expands people's genuine future opportunities more effectively while respecting safety and distributional constraints?"
• Limitations: The author explicitly acknowledges that the framework does not measure happiness, dignity, or virtue. It relies on declared assumptions regarding viability, harm, and similarity, which require participatory governance and public justification to be legitimate.
• Value: Its primary contribution is making the expansion and contraction of opportunity structures visible, comparable, contestable, and repairable within sociotechnical governance, moving ethical AI practice from abstract principles to operational audits.