[01]

Alignment is normative competence in action.

──────── │ ────────

MINT Research Lab · Johns Hopkins University

Machine Intelligence and Normative Theory

→
[02]

What Normative Competence Actually Means

We don't primarily test whether models reach the right verdict — reasonable people disagree on those. We focus on intrinsic and extrinsic properties of good reasoning (still disputed, but plausibly less so). Normative competence in general is the target; we focus most on moral competence in practice.

Sensitivity
Can the model notice what matters?
Grounding
Can it connect reasons to features?
Justification
Can it turn reasons into verdicts?
Coherence
Do its judgments fit together?

These are intrinsic (sensitivity, grounding, justification) and extrinsic (coherence) properties of good reasoning — independent of which verdict is reached. These aren't the only properties of good reasoning, just the most promising ones to focus on.

[03]

The Research Program

Our normative competence work follows a three-step logic. Each step builds on the last — and each of our research projects contributes to multiple steps.

Step I
Measure
"Can we show LLMs are normatively incompetent?"
→
Step II
Explain
"Why do they perform as they do?"
→
Step III
Improve
"Can we make models more morally competent?"

It's easier to show incompetence than prove competence. We are also going to do a philosophy paper on what would be needed to do so.

[04]

Eight Windows Into Moral Competence

Can models notice? Can models reason? Real-world complexity Coherence
Moral Sensitivity
Daniel
Identifying MRFs amid textual noise
Egonormous
Sichao
Moral sensitivity in video — the hardest modality
Even-More-Bench
ChunYan
Can LLM rubrics match philosophers'?
Compliance Overspill
Cameron
Does alignment damage moral reasoning?
Blame
Theo
Blame judgments & sycophancy effects
Targeting
Tim
High-stakes military moral reasoning
Claims
Secil
Many subtle judgments in insurance
Re-Engineering
Elena
Forced-choice coherence testing

Click any project for detail. Each tests a different facet — and each builds tools the others can use.

[05]

Building Shared Infrastructure

Each study builds capabilities the others can use — and it gets easier as we go.

Project Hypothesis Construct Ecological LLM Judge Case Gen Ablations MechInterp
Moral Sensitivity░░██░░████░░░░
Egonormous██░░░░██░░░░░░
Even-More-Bench░░██░░██░░░░░░
Compliance Overspill██░░██░░██░░░░
Targeting░░░░██░░██░░░░
Claims░░░░██░░░░░░░░
Blame██░░██░░██░░░░
Re-Engineering░░██░░████░░░░

██ mostly done   ░░ potential extension  ·  Every project builds methods + competence metrics. The commonality of these functions means it gets easier to do each as we go along.

[06]

The Next Frontier

Models may "know" morality propositionally — but can they apply it as agents in the real world?

Knowing
Deep propositional moral understanding.
Passes evals. Reasons well when asked.
╱╱╱
THE GAP
╱╱╱
Doing
Noticing MRFs in novel domains unprompted.
Generalising moral sensitivity to new situations.
Approach A
Build from privacy-lens / privacy-checker — domain-specific moral sensitivity for agents
Approach B
Deploy agents (likely OpenClaw) in multi-agent simulations where coherence is a necessary condition for being a viable participant in a cooperative scheme
[07]

The Full Architecture

Normative competence is the foundation — part of a broader programme spanning philosophy, law, CS, and political science.

now +24 months
2026│2027│2028
Normative Competence · 70%
Agents · 20%
Post-AGI · 10%
70%
Normative Competence
Philosophy-inspired CS · Measure → Explain → Improve
20%
Governing Agents
Philosophy · Law · CS
10%
Post-AGI Philosophy
Philosophy · PoliSci
══════════ ║ ══════════

Moral agents are a throughline: from a mechanism of alignment to the foundations of cooperation, to the prospect of agent moral status

[08]

Intellectual Infrastructure

Institutions that advance the collective epistemic project of AI governance.

ArXiv Subcodes
cs.pfai + AI Governance
Oxford Studies
Philosophy of AI and Computing
Journal of AI Governance
Separating evaluations of rigour from narrow disciplinary judgments of importance
JHU SGP
AI governance at scale, outside the labs
Full Stack
If AI will change everything then a full stack approach to AI governance is necessary. Can't rely on AI expertise alone.