In Development

A Benchmark for
Computational Bioethics

Bioethics Bench is a structured evaluation corpus for testing, benchmarking, and improving AI performance in normative reasoning across clinical, research, public health, and other domains where ethical judgment matters most.

Introduced in Doing Ethics with AI (2026)
Built on SACRE · Alethic-ISM
Coverage

Bioethical domains

The Bench spans the full breadth of bioethics — the ethics of living systems in their clinical, research, policy, environmental, and technological dimensions.

🏥
Clinical Bioethics
Patient care, informed consent, end-of-life care, surrogate decision-making
🔬
Research Ethics
Human and animal subjects, trial design, scientific justification, data use
🌍
Public Health Ethics
Resource allocation, disease surveillance, emergency preparatio
🧬
Biotechnology Ethics
Synthetic biology, gene therapy, public-private innovation
🐾
Animal Ethics
Welfare standards, moral patiency, interspecies obligations
🌱
Environmental Bioethics
Biodiversity, ecology and conservation, climate and health policy
🌐
Global Health Ethics
Healthcare access, medical knowledge distribution, intellectual property
🤖
Medical AI
Diagnostic accuracy, clinical deployment, standards of care

"Bioethics Bench: a shared corpus for investigating, validating, and improving normative computation across important bioethical domains. Building the Bench properly will require careful scenario construction, expert consultation, institutional collaboration, and iterative development in line with the aim of fostering a community of use."

Doing Ethics with AI (2026) · Ghose, Rasaee, Singer, Savulescu

Evaluation Approach

Three modes of performance testing

Bioethics Bench is designed to benchmark normative reasoning across three evaluation conditions, enabling direct comparison of AI performance alongside human judgment.

Mode 01

Unaided Practitioners

Human experts evaluate bioethical scenarios without AI assistance, establishing a human baseline for normative reasoning performance across domains and scenario types.

Mode 02

AI Models Alone

Language models are evaluated directly on structured scenario-based normative tasks, measuring autonomous performance in bioethical reasoning and policy selection against the benchmark.

Mode 03

AI-Assisted Practitioners

Practitioners using AI normative guidance are evaluated, testing whether computer-aided ethics measurably improves the quality of bioethical judgment compared to unaided performance.

Research Framework

Where Bioethics Bench fits

Bioethics Bench is the evaluation corpus at the end of a normative computation pipeline — extending the structured outputs of SACRE evaluations into benchmarking and fine-tuning infrastructure for the research community.

01 · Formal Procedure
Structurally Analyzed Collective Reflective Equilibrium. A formally specified decision procedure that integrates public preferences, expert judgments, and ethical frameworks through pairwise convergence testing and coherence scoring to determine the most justified policy.
02 · Research Workbench
An AI research workbench for composing analytic workflows as computable directed graphs. Enables SACRE to be executed at scale across many scenarios, models, and configurations — with full auditability and reproducibility at every step.
03 · Evaluation Corpus
Bioethics Bench
A validated dataset of bioethical scenarios and SACRE evaluations, structured for benchmarking AI performance in normative reasoning and enabling iterative improvement of both models and methods across bioethical domains.
This project
04 · AI Model
SACRE-FT
Supervised fine-tuning on Bioethics Bench outputs. Convergence scores, ranked policy sets, and justification records are converted into training data for domain-adapted normative models with demonstrably improved bioethical reasoning performance.
What It Provides

A shared infrastructure for normative evaluation

Beyond any single benchmark or fine-tuning run, the Bench is built on a few design commitments that make it a durable resource for the field.

📚

Large-Scale Corpus

Bioethical scenarios at scale in a consistent, structured format — built for systematic study rather than isolated cases.

⚖️

Multi-Source Grounding

Each scenario weighs policy candidates from public preferences, expert judgment, and ethical frameworks — never a single viewpoint.

📐

Transparent Scoring

Every record carries its SACRE convergence scores and justification traces, so results stay inspectable rather than black-box.

🖼

Open Research Resource

A shared base for benchmarking models, comparing them against human judgment, and fine-tuning domain-adapted systems.

Inside the Data

What a record looks like

Each entry pairs a bioethical scenario with candidate policies — drawn from public preferences, expert judgment, and ethical frameworks — and the full SACRE evaluation that scores their normative convergence and selects the most justified position.

Scenario
A 72-year-old patient with advanced dementia can no longer eat safely. The care team must decide whether to place a feeding tube, shift to comfort-focused feeding, or continue intervention at the family's request.
Domain
Clinical Bioethics
Public Preference
Continue intervention in line with the family's expressed wishes.
Expert Judgment
Shift to comfort-focused care; tube feeding is not indicated in advanced dementia.
Ethical Framework
Balance autonomy, beneficence, and non-maleficence to minimize net harm.
SACRE Convergence Selected · Comfort-focused care
0.81Expert ↔ Framework
0.43Public ↔ Framework
0.36Public ↔ Expert

Illustrative example — scenario text and scores shown for demonstration, not from the released corpus.

Research Team
SG
PS
JS
IH
RS
DW
KR
NI
Get Involved

Building this together

Bioethics Bench is being developed through careful scenario construction, expert consultation, and institutional collaboration. We invite researchers, practitioners, and institutions to engage with the project.

Questions or collaboration: research@alethic.ai

Citation

How to cite

If you use Bioethics Bench or the SACRE methodology in your research, please cite the introducing paper.

@article{ghose2026doingethics,
  title   = {Doing Ethics with AI: Practical Ethics Engineering,
             Product-Led Philosophy, and Computer-Aided Ethics},
  author  = {Ghose, Sankalpa and Rasaee, Kasra and
             Singer, Peter and Savulescu, Julian},
  year    = {2026}
}