Doing Ethics with AI

Practical Ethics Engineering, Product-Led Philosophy, and Computer-Aided Ethics

Sankalpa Ghose1,5   Kasra Rasaee4,5   Peter Singer2,1   Julian Savulescu1,3

1National University of Singapore   2Princeton University   3University of Oxford   4Bloomberg   5Alethic Research

↗ Read the paper ↗ Try SACRE ↗ Alethic-ISM
Computer-aided ethics proposes using computational systems as instruments of normative reasoning to better evaluate what to do in situations where it matters. We introduce practical ethics engineering as a field and product-led philosophy as a research method in which ethical analysis is specified, evaluated, and assisted using computational justification instruments and normative guidance systems.
Figure 1

Ethics with AI — Engineering Turn Diagram

Figure 1. Ethics with AI proposes an "engineering turn" for practical ethics.

Normative Computation Pipeline

01
Conceptual Model
02
Calculable Model
03
Executable Model
04
AI Model
Figure 2

Product-Led Philosophy — Research Cycle

Figure 2. Product-led philosophy is a research method for investigating normative problems. As a field of engineering it seeks to develop interactive instruments and devices; as a method of empiricism it seeks to test these in evaluative and decision-making contexts of computer-aided ethics.

Main Contributions

⚖️
SACRE
Structurally Analyzed Collective Reflective Equilibrium

A formal decision procedure for determining the most coherently justified policy for a given scenario through convergence testing of policy candidates from public preferences, expert judgments, and ethical frameworks.

Try the application →
📊
QCS & QCCS
Quantified Conceptual Scoring Methods

QCS quantifies conceptual intensity on a 0–100 scale. QCCS extends this to pairwise convergence evaluation, providing the normative convergence function used in SACRE's equilibration steps.

Interactive demo →
🔧
Alethic-ISM
AI Research Workbench

A distributed computation engine for building analytic and agentic graphs where every execution step produces a durable, versioned state with an audit trail. Supports large-scale normative computation with full provenance.

View on GitHub →
🧠
SACRE-FT
Supervised Fine-Tuning Pipeline

A closed-loop architecture where SACRE evaluation outputs become training data for domain-adapted AI models. Execute → Collect → Train → Deploy → Execute.

📚
Bioethics Bench
Evaluation Corpus (In Development)

A shared corpus for investigating, validating, and improving normative computation across bioethical domains — clinical ethics, research ethics, public health ethics, and more.

Learn more →
Figure 18

ReflectiveEquilibrium.AI — Application Interface

Figure 18. The SACRE illustrative demo provides a walkthrough within the interactive web application. Users are presented with an interface that lays out the purpose of the analysis along with each step of the SACRE method.

SACRE Computational Graph in Alethic-ISM

📐

Interactive SACRE graph visualization

Embed URL to be added

The SACRE Method

Figure 15

SACRE Decision Procedure Diagram

Figure 15. Structurally Analyzed Collective Reflective Equilibrium (SACRE). Step 1 identifies the Scenario. Steps 2A, 2B, 2C collect Policy candidates from public preferences, expert judgments, and ethical frameworks. Step 3 performs Public–Expert Equilibration. Step 4 performs Public–Expert–Framework Equilibration. Step 5 selects the Final Policy Π*.

Step 1: Scenario Identification

Identify the scenario S in need of policy determination. The scenario defines the case, context, or practical problem requiring a policy recommendation.

Step 2: Policy Candidate Identification

Collect policy candidates from three independent normative sources: public preferences (Πpub), expert judgments (Πexp), and ethical frameworks (Πfw). Each source provides distinct perspectives grounded in lived experience, empirical evidence, or systematic normative theories.

Step 3: Public–Expert Equilibration

Measure pairwise convergence between all public and expert policy candidates using QCCS. Aggregate scores into policy coherence profiles and produce P′, the initial equilibrated set rank-ordered by coherence.

Step 4: Public–Expert–Framework Equilibration

Measure convergence between the initial equilibrated set and framework candidates. Aggregate to produce P″, the final equilibrated set of all policy candidates rank-ordered by total coherence.

Step 5: Final Policy Selection

Select Π* = argmax Score″(Π) — the policy with the highest total convergence score across all pairwise comparisons is designated as the most coherently justified policy for the scenario.

Cite this work

@article{ghose2026doingethics,
  title   = {Doing Ethics with AI},
  author  = {Ghose, Sankalpa and Rasaee, Kasra and Singer, Peter
             and Savulescu, Julian},
  journal = {[Journal TBD]},
  year    = {2026},
  note    = {Application: https://reflectiveequilibrium.ai
             Alethic-ISM: https://github.com/alethicresearch/alethic-ism}
}