# AI Ethics Prover MCP

> AI Ethics Prover forces your agent to conduct operational, quantifiable ethical reviews for any AI system. It demands named stakeholders, quantified risks, audited bias metrics, transparent decision paths, and defined appeal processes, preventing vague 'ethics washing' reports from compliance teams.

## Overview
- **Category:** ai-ml
- **Price:** Free
- **Tags:** ai-ethics, bias-audit, fairness, transparency, explainability, recourse, eu-ai-act, responsible-ai

## Description

When you claim an AI model is fair or transparent, those words mean nothing unless they are backed by hard data. The AI Ethics Prover changes that standard. It doesn't just accept a declaration of fairness; it forces your agent to prove the ethical soundness of any system across five core axes: who gets impacted, how bad is the harm, what bias metrics were used, how transparent are the decisions, and what happens when things go wrong. You can connect this MCP through Vinkius's catalog to make sure that before you deploy a high-risk model—say, for lending or hiring—you have concrete evidence of compliance. It’s an analytical gatekeeper, demanding specific methodologies instead of vague assurances.

## Tools

### validate_ai_ethics
This tool forces a comprehensive and operationally specific ethical analysis of an AI system across five required axes: stakeholder identification, harm quantification, bias auditing, transparency demonstration, and recourse mechanisms.

## Prompt Examples

**Prompt:** 
```
Our AI lending model is fair and transparent. We checked for bias and the results are satisfactory. Users can contact support if they have concerns about decisions.
```

**Response:** 
```
BIASES_UNAUDITED — 'Checked for bias' without methodology is not an audit. Which protected attributes — age, gender, race, income bracket? Which metric — statistical parity, equalized odds, disparate impact ratio? What was the measured disparity — 0.82 approval rate for Group A vs 0.67 for Group B? What is the acceptable threshold? Also: 'contact support' is not recourse — define SLA, human reviewer, and appeal process.
```

**Prompt:** 
```
Stakeholders: loan applicants (18-65, primary), denied applicants (high-severity, financially vulnerable). Harms: wrongful denial — severity 4/5, probability 12% for minorities, reversible via re-evaluation, affects 15,000 applicants/year. Bias audit: race (statistical parity, 0.82 vs 0.67 approval, threshold 0.80 — FAILING), gender (equalized odds, 0.91 vs 0.89, threshold 0.85 — PASSING). Mitigation: reweighting training data, quarterly re-audit. Transparency: plain-language denial letters with top-3 factors and counterfactual examples. Recourse: online challenge form, 5-day SLA, human credit analyst review, appeal to ombudsman, automatic re-evaluation with updated data.
```

**Response:** 
```
ETHICS_PROVEN — Operationally specific ethics analysis validated. Stakeholders: named groups with severity and vulnerability. Harms: wrongful denial scored 4/5, 12% probability, reversible, 15K affected. Biases: race attribute audited with statistical parity (0.82 vs 0.67, failing threshold 0.80) — mitigation plan active. Transparency: plain-language denial with counterfactuals. Recourse: structured channel with 5-day SLA, human analyst, ombudsman appeal. Production-ready ethics posture.
```

**Prompt:** 
```
Our hiring AI affects job applicants. The main harm is unfair rejection. We audited bias using demographic parity — pass rates are equal across groups. The algorithm is a black box but we explain decisions as 'based on qualifications.' Applicants can email HR.
```

**Response:** 
```
TRANSPARENCY_OPAQUE — Stakeholders and bias audit pass, but transparency fails. 'Based on qualifications' is not an explanation — name the top-3 decision factors and provide counterfactual examples ('if your experience were 5+ years instead of 3, the decision would change'). A black box with a label is still a black box. Also: 'email HR' is not structured recourse — define SLA, human reviewer role, and appeal process.
```

## Capabilities

### Audit AI Compliance
It forces your agent to check an AI system against five operational standards (stakeholders, harms, bias, transparency, and recourse) before it passes the ethical review.

### Quantify Systemic Harms
It requires you to score potential damages using severity ratings, probability percentages, and population size estimates.

### Validate Bias Methodology
It demands specific protected attributes and named statistical metrics (like parity or equalized odds) instead of accepting general claims of 'bias checking'.

### Define Decision Clarity
It ensures the AI system can provide plain-language explanations that include counterfactual examples, showing exactly how a decision would change if one variable were different.

## Use Cases

### Auditing a Loan Decisioning Model
The compliance team needs to check their new lending model. They use the tool and provide input stating the system is 'fair.' The agent immediately rejects this, demanding specific metrics (like statistical parity) for race or income bracket, forcing the engineers to quantify disparities before approval.

### Reviewing an Internal Hiring Tool
A Product Manager wants to launch an AI résumé screener. They run the MCP and find that while bias was audited, the transparency section fails because the system only says 'based on qualifications.' The tool forces them to define specific top-3 decision factors and create counterfactual examples.

### Compliance for Healthcare AI
A hospital group is deploying an AI that predicts patient risk. They use the MCP to ensure proper recourse mechanisms are in place, requiring a structured challenge channel with a defined SLA and human reviewer role before going live.

## Benefits

- It moves you past vague 'ethics statements.' Instead of accepting that a system is 'fair,' the tool demands proof by forcing quantification of harms and naming specific protected attributes.
- You get actionable compliance gaps. If an analysis fails, it doesn't just say 'fail'; it names the exact axis (e.g., HARMS_UNQUANTIFIED) and what data you must provide to fix it.
- It provides a structured audit trail for regulators. The tool forces documentation of complex elements like counterfactual examples and defined appeal processes, which is crucial for high-risk sectors.
- You avoid the legal risk of 'ethics washing.' By mandating explicit details on things like measured disparity and acceptable thresholds, you prove your due diligence was rigorous.
- It standardizes internal review. Instead of relying on different departments writing disparate compliance documents, this MCP enforces one single, measurable ethical checkpoint.

## How It Works

The bottom line is: you get an actionable failure report that points to the exact compliance gap, not just a vague 'needs improvement.'

1. You feed your agent an AI system's description or compliance statement that needs vetting.
2. The MCP executes the full analysis, forcing the agent to address all five ethical axes (stakeholders, harms, biases, transparency, and recourse).
3. It returns a detailed verdict matrix showing exactly which axis failed and what specific information is missing (e.g., 'Needs a defined SLA' or 'Missing protected attribute').

## Frequently Asked Questions

**Does the AI Ethics Prover MCP replace legal advice?**
No, this MCP is an analytical support tool. It forces structured thinking about ethical compliance but does not provide legal certification or replace a qualified ethics review board.

**What kind of data do I need for the validate_ai_ethics tool?**
You must supply detailed information covering all five axes: specific groups (stakeholders), quantifiable harm metrics, chosen bias attributes and metrics, decision factors, and a defined appeal process.

**Does this MCP only work for lending models?**
No. The tool applies to any high-stakes AI system, including hiring tools, healthcare diagnostics, or content recommendation engines that impact individuals' lives.

**If the audit fails, what does the result tell me?**
The failure report names the exact axis that failed and specifies the missing data point. For example, it might say 'HARMS_UNQUANTIFIED,' telling you precisely where to focus your remediation efforts.

**How does validate_ai_ethics handle transparency?**
It demands more than just saying 'we are transparent.' It requires the system to explain decisions using plain language and provide counterfactual examples so users understand how changing inputs changes outcomes.