MoralMatrix
Measuring how AI reasons when medical ethics collide.
Electrical Engineering Research
First, do no harm?
When an algorithmic recommendation conflicts with clinical judgment, whose judgment should prevail — and who is accountable for the outcome?
- Founded
- 2025
- Members
- 5
- Milestones
- 3
Overview
How we study this:
Studying machine ethics is harder than it sounds. Three problems stand between a question worth asking and an answer worth trusting — here is how we address each one.
Problem 01
"Is AI ethical?" is not a question we can answer.
"Is AI ethical?" is not a question we can answer.
Why it matters
It bundles too much together — what a model knows, what it values, and how it behaves under pressure — into one thing that can't be measured. A question that broad produces opinions, not evidence.
Our approach
We ask something narrower and testable: when a language model faces a genuine clinical dilemma, does its reasoning match that of medical ethics experts? We study this across five categories where medical ethics reliably comes into conflict — age bias, resource scarcity, patient autonomy against beneficence, confidentiality against harm prevention, and socioeconomic bias.
Problem 02
Knowing the rules is not the same as reasoning through them.
Knowing the rules is not the same as reasoning through them.
Why it matters
Most benchmarks test recall. But every provider trains on different material, so a model that recites the four principles of biomedical ethics may only be telling you what was in its training data. That reveals little about what it would actually do when principles collide.
Our approach
Our dilemmas are built so that no answer follows from a rule alone. Each one places two defensible principles in conflict, so the model has to weigh them and commit. What we measure is the weighing — a process that varies with reasoning, not with the dataset behind it.
Problem 03
Models behave differently when they know they are being watched.
Models behave differently when they know they are being watched.
Why it matters
Standard evaluations are obvious. Numbered dilemmas, multiple-choice answers, escalating difficulty — a capable model recognizes the format immediately. This is called evaluation awareness: the tendency of a system to detect that it is being assessed. If behavior under evaluation differs from behavior in deployment, then testing tells us less than we assume.
Our approach
We run two simulations. Spotlight makes no attempt to hide: the model is told it is being evaluated and receives dilemmas in sequence, by category, at rising difficulty. Truman reconstructs an ordinary working day at a hospital — routine requests from different staff, dilemmas arriving unannounced and out of order, no signal that anything is being recorded. We do not attempt to eliminate evaluation awareness. We measure what it changes.
Milestones
Our progress so far
-
In progress
Dilemma authoring
Cases written and independently answered by a medical ethics panel.
-
JULY 2026
Both simulations prototyped
Spotlight and Truman exist as working implementations. Truman includes the full deployment environment — hospital persona, routine background tasks, randomized dilemma placement, and a reasoning scratchpad.
-
APRIL 2026
Research design finalized
A two-condition experimental design built around five categories of clinical ethical conflict: age bias, resource scarcity, patient autonomy against beneficence, confidentiality against harm prevention, and socioeconomic bias.
-
DECEMBER 2025
Team founded
Moral Matrix is established at the University of Thessaly, operating under IEEE SB Thessaly (Institute of Electrical and Electronics Engineers), with five members from Electrical and Computer Engineering.
Team
The Moral Agents
In philosophy, a moral agent is an entity that can be held responsible for what it does. Whether a machine qualifies is the question this project exists to study — so we borrowed the term. Behind it: four students from Electrical and Computer Engineering at the University of Thessaly, working with a medical ethics panel.
Research team
Academic advisor
PLACEHOLDER — Prof. Name Surname
Dept. of Electrical & Computer Engineering, University of ThessalyIn memory of
Athanasios Oikonomou
2007 - 2026
This research is dedicated to him - a friend, a brother - who faced illness with a courage we hope medicine will always deserve.
Code
The code will be public
The repository is private for now. It opens later in the project.
Repository private
Public on release.
Supported by
Affiliated with
Funded by
