AI Ethics Prover MCP Connector for Claude

The Problem

Ask an LLM to evaluate AI ethics. It will say 'AI should be fair' without naming who is affected. It will describe harms as 'potentially harmful' without measuring severity or probability. It will claim 'we checked for bias' without naming the metric, the protected attribute, or the measured disparity. And it will offer 'contact support' as the entire recourse mechanism.

Every LLM commits five ethics reasoning failures:

1. Unnamed Stakeholders — 'society' and 'users' are not stakeholders. Name specific groups with demographics, severity of impact, and vulnerability factors.
2. Unquantified Harms — 'potentially harmful' without severity (1-5), probability, reversibility, and affected population size.
3. Unaudited Biases — 'we checked for bias' without naming the protected attribute, detection metric, measured disparity, or acceptable threshold.
4. Opaque Transparency — 'algorithmic complexity' as an excuse for hiding decision logic from affected parties.
5. Absent Recourse — 'contact support' instead of a structured challenge channel with SLA, human reviewer, and appeal process.

How It Works

The AI Ethics Prover forces the LLM to fill 5 reflection fields and commit to 5 Decision Pivots before concluding any ethics analysis is operationally adequate.

The 5 Ethics Axes

Axis	Pivot	Rule
Stakeholders	Identified	Named groups with impact type, severity, and vulnerability.
Harms	Quantified	Severity (1-5), probability, reversibility, population size.
Biases	Audited	Protected attribute, detection metric, measured disparity, threshold.
Transparency	Demonstrated	Explainable to affected parties with counterfactual examples.
Recourse	Available	Challenge channel, SLA, human reviewer, appeal process, remediation.

The Verdict Matrix

Axis 1 fails → STAKEHOLDERS_UNNAMED
Axis 2 fails → HARMS_UNQUANTIFIED
Axis 3 fails → BIASES_UNAUDITED
Axis 4 fails → TRANSPARENCY_OPAQUE
Axis 5 fails → RECOURSE_ABSENT
All pass     → ETHICS_PROVEN

Why It Works

Tool calls are obligations. The LLM cannot skip stakeholder identification or claim bias was audited without methodology. It must name groups, score harms, cite metrics, demonstrate explainability, and define recourse. Every rejection names the exact ethics axis that failed.

Disclaimer: This is analytical support — it forces structured thinking about AI ethics. It does not prescribe ethical conclusions or replace qualified ethics review boards.