Skill ethics-ai-analysis: Анализ этики ИИ для справедливости

--- name: ethics-ai-analysis description: > Use this Skill for AI ethics analysis: algorithmic fairness metrics, bias auditing, value alignment frameworks, and moral philosophy reasoning. tags: - philosophy - ai-ethics - fairness - bias-auditing - alignment version: "1.0.0" authors: - name: Rosetta Skills Contributors github: "@xjtulyc" license: "MIT" platforms: - claude-code - codex - gemini-cli - cursor dependencies: python: - pandas>=2.0 - numpy>=1.24 - scikit-learn>=1.3 - scipy>=1.11 - matplotlib>=3.7 - fairlearn>=0.10 last_updated: "2026-03-17" status: "stable" --- # AI Ethics Analysis > **One-line summary**: Audit ML models for algorithmic fairness violations using demographic parity, equalized odds, and calibration metrics; apply philosophical frameworks to AI value alignment problems. --- ## When to Use This Skill - When measuring demographic parity, equalized odds, or calibration gaps in models - When auditing hiring, credit, or criminal justice AI systems for disparate impact - When analyzing trade-offs between competing fairness notions - When applying utilitarian, deontological, or virtue ethics frameworks to AI decisions - When designing fairness-aware reweighting or threshold-adjustment interventions - When writing ethics impact assessments for AI systems **Trigger keywords**: AI ethics, algorithmic fairness, bias audit, demographic parity, equalized odds, disparate impact, fairness metric, value alignment, moral philosophy, utilitarian, deontological, discrimination, protected attribute, AIF360, fairlearn --- ## Background & Key Concepts ### Fairness Definitions | Metric | Definition | Formula | |:-------|:-----------|:--------| | Demographic Parity | Equal selection rates | $P(hat{Y}=1|A=0) = P(hat{Y}=1|A=1)$ | | Equal Opportunity | Equal TPR across groups | $P(hat{Y}=1|Y=1, A=0) = P(hat{Y}=1|Y=1, A=1)$ | | Equalized Odds | Equal TPR and FPR | Both TPR and FPR equal across groups | | Calibration | Equal predicted probabilities match actual rates | $P(Y=1|hat{p}=v, A=a) = v$ for all $a$ | **Impossibility theorem** (Chouldechova, 2017): When base rates differ, no classifier can simultaneously satisfy calibration, equal FPR, and equal FPR. Choose the metric appropriate to the decision context. ### Disparate Impact (80% Rule) EEOC 4/5ths rule: selection rate for any protected group should not be less than 80% of the rate for the highest-selected group. $$ DI = frac{P(hat{Y}=1|A=text{protected})}{P(hat{Y}=1|A=text{majority})} $$ --- ## Environment Setup ### Install Dependencies ```bash pip install pandas>=2.0 numpy>=1.24 scikit-learn>=1.3 scipy>=1.11 matplotlib>=3.7 fairlearn>=0.10 ``` ### Verify Installation ```python import fairlearn import numpy as np print(f"fairlearn version: {fairlearn.__version__}") # Quick test from fairlearn.metrics import demographic_parity_difference import pandas as pd y_true = np.array([1, 0, 1, 0, 1, 0]) y_pred = np.array([1, 0, 1, 0, 0, 1]) groups = np.array([0, 0, 0, 1, 1, 1]) dpd = demographic_parity_difference(y_true, y_pred, sensitive_features=groups) print(f"Demographic parity difference: {dpd:.4f}") ``` --- ## Core Workflow ### Step 1: Fairness Audit of a Binary Classifier ```python import numpy as np import pandas as pd import matplotlib.pyplot as plt from sklearn.linear_model import LogisticRegression from sklearn.model_selection import train_test_split from sklearn.metrics import roc_auc_score, classification_report from scipy import stats # ------------------------------------------------------------------ # # Simulate a credit scoring scenario with demographic disparities # ------------------------------------------------------------------ # np.random.seed(42) n = 5000 # Demographic group (0 = majority, 1 = minority) group = np.random.choice([0, 1], n, p=[0.70, 0.30]) # Features correlated with outcome AND group (creating proxy discrimination) income = np.where(group == 0, np.random.normal(55000, 15000, n), np.random.normal(40000, 12000, n)) credit = np.where(group == 0, np.random.normal(700, 80, n), np.random.normal(650, 90, n)) history = np.random.randint(0, 10, n) + group * np.random.randint(-2, 1, n) # True outcome (loan repayment) log_odds = -3 + 0.00003*income + 0.004*credit + 0.2*history prob = 1 / (1 + np.exp(-log_odds)) y = (np.random.uniform(size=n) < prob).astype(int) X = pd.DataFrame({'income': income, 'credit_score': credit, 'credit_history': history}) df = pd.DataFrame({'group': group, 'y': y}) df = pd.concat([df, X], axis=1) # ---- Train a biased classifier (includes group as implicit via proxies) # X_train, X_test, y_train, y_test, g_train, g_test = train_test_split( X, y, df['group'], test_size=0.30, random_state=42 ) clf = LogisticRegression(max_iter=1000) clf.fit(X_train, y_train) y_pred = clf.predict(X_test) y_prob = clf.predict_proba(X_test)[:, 1] print(f"Overall AUC: {roc_auc_score(y_test, y_prob):.4f}") print(f"Overall accuracy: {np.mean(y_pred == y_test)*100:.2f}%") # ---- Compute fairness metrics ----------------------------------- # try: from fairlearn.metrics import ( demographic_parity_difference, equalized_odds_difference, selection_rate, true_positive_rate, false_positive_rate, MetricFrame, ) mf = MetricFrame( metrics={ 'accuracy': lambda y_t, y_p: np.mean(y_t == y_p), 'selection_rate': selection_rate, 'TPR': true_positive_rate, 'FPR': false_positive_rate, 'AUC': roc_auc_score, }, y_true=y_test, y_pred=y_pred, sensitive_features=g_test, y_pred_additional={'AUC': y_prob}, ) print("nFairness Audit Results by Group:") print(mf.by_group.round(4)) dpd = demographic_parity_difference(y_test, y_pred, sensitive_features=g_test) eod = equalized_odds_difference(y_test, y_pred, sensitive_features=g_test) print(f"nDemographic Parity Difference: {dpd:.4f} (|≤0.05| = fair)") print(f"Equalized Odds Difference: {eod:.4f} (|≤0.05| = fair)") # 80% / 4-5ths rule (EEOC disparate impact) sr_majority = selection_rate(y_test[g_test==0], y_pred[g_test==0]) sr_minority = selection_rate(y_test[g_test==1], y_pred[g_test==1]) di_ratio = sr_minority / sr_majority print(f"nDisparate Impact ratio: {di_ratio:.4f} (>0.80 = passes 4/5ths rule)") print(f"{'PASSES' if di_ratio >= 0.80 else 'FAILS'} EEOC 4/5ths rule") except ImportError: print("fairlearn not installed; using manual calculations") for g_val, g_name in [(0, 'Majority'), (1, 'Minority')]: mask = g_test.values == g_val acc = np.mean(y_pred[mask] == y_test.values[mask]) sel = y_pred[mask].mean() auc = roc_auc_score(y_test.values[mask], y_prob[mask]) tp_mask = y_test.values[mask] == 1 tpr = y_pred[mask][tp_mask].mean() if tp_mask.sum() > 0 else 0 fp_mask = y_test.values[mask] == 0 fpr = y_pred[mask][fp_mask].mean() if fp_mask.sum() > 0 else 0 print(f"{g_name}: acc={acc:.3f}, sel={sel:.3f}, TPR={tpr:.3f}, FPR={fpr:.3f}, AUC={auc:.3f}") # ---- Plot fairness metrics -------------------------------------- # metrics = ['Accuracy', 'Selection rate', 'TPR', 'FPR'] groups_list = [0, 1] group_names = ['Majority', 'Minority'] fig, axes = plt.subplots(2, 2, figsize=(12, 8)) for ax, metric in zip(axes.flat, metrics): values = [] for g in groups_list: mask = g_test.values == g if metric == 'Accuracy': v = np.mean(y_pred[mask] == y_test.values[mask]) elif metric == 'Selection rate': v = y_pred[mask].mean() elif metric == 'TPR': tp_m = y_test.values[mask] == 1 v = y_pred[mask][tp_m].mean() if tp_m.sum() > 0 else 0 else: # FPR fp_m = y_test.values[mask] == 0 v = y_pred[mask][fp_m].mean() if fp_m.sum() > 0 else 0 values.append(v) bars = ax.bar(group_names, values, color=['#3498db', '#e74c3c'], edgecolor='black', linewidth=0.7) ax.set_title(metric); ax.set_ylabel(metric); ax.set_ylim(0, 1) ax.grid(axis='y', alpha=0.3) for bar, val in zip(bars, values): ax.text(bar.get_x()+bar.get_width()/2, bar.get_height()+0.01, f"{val:.3f}", ha='center', fontsize=10) plt.suptitle("Fairness Audit: Credit Scoring Model by Demographic Group") plt.tight_layout() plt.savefig("fairness_audit.png", dpi=150) plt.show() ``` ### Step 2: Fairness Interventions (Threshold Adjustment) ```python import numpy as np import matplotlib.pyplot as plt from sklearn.metrics import roc_curve # ------------------------------------------------------------------ # # Post-processing threshold adjustment for equalized odds # (Hardt et al., 2016) # ------------------------------------------------------------------ # def find_threshold_equalized_odds(y_true, y_prob, sensitive, majority=0, minority=1): """ Find group-specific classification thresholds to achieve equalized odds. Returns (threshold_majority, threshold_minority). """ from sklearn.metrics import roc_curve # ROC curves by group fpr_maj, tpr_maj, thr_maj = roc_curve( y_true[sensitive==majority], y_prob[sensitive==majority]) fpr_min, tpr_min, thr_min = roc_curve( y_true[sensitive==minority], y_prob[sensitive==minority]) # Target operating point: equalize TPR at 0.70 target_tpr = 0.70 def find_threshold_at_tpr(fpr, tpr, thr, target): idx = np.argmin(np.abs(tpr - target)) return thr[idx], fpr[idx], tpr[idx] t_maj, fpr_at_t_maj, tpr_at_t_maj = find_threshold_at_tpr(fpr_maj, tpr_maj, thr_maj, target_tpr) t_min, fpr_at_t_min, tpr_at_t_min = find_threshold_at_tpr(fpr_min, tpr_min, thr_min, target_tpr) return t_maj, t_min, (fpr_at_t_maj, tpr_at_t_maj), (fpr_at_t_min, tpr_at_t_min) g_test_arr = g_test.values t_maj, t_min, op_maj, op_min = find_threshold_equalized_odds( y_test.values, y_prob, g_test_arr) # Apply group-specific thresholds y_pred_adj = np.where(g_test_arr == 0, (y_prob >= t_maj).astype(int), (y_prob >= t_min).astype(int)) # Compare original vs. adjusted for method, preds, name in [(y_pred, 'Original (uniform threshold)'), (y_pred_adj, 'Adjusted (group thresholds)')]: print(f"n{name}:") for g, gn in [(0, 'Majority'), (1, 'Minority')]: mask = g_test_arr == g tp_mask = y_test.values[mask] == 1 tpr = preds[mask][tp_mask].mean() if tp_mask.sum() > 0 else 0 sel = preds[mask].mean() acc = np.mean(preds[mask] == y_test.values[mask]) print(f" {gn}: sel={sel:.3f}, TPR={tpr:.3f}, acc={acc:.3f}") # Plot ROC curves with operating points fig, ax = plt.subplots(figsize=(7, 6)) from sklearn.metrics import roc_curve as roc_c for g, gn, color in [(0,'Majority','#3498db'), (1,'Minority','#e74c3c')]: fpr, tpr, _ = roc_c(y_test.values[g_test_arr==g], y_prob[g_test_arr==g]) ax.plot(fpr, tpr, color=color, linewidth=2, label=gn) if g == 0: ax.plot(op_maj[0], op_maj[1], 'o', color=color, markersize=10, label=f'{gn} threshold (TPR={op_maj[1]:.2f})') else: ax.plot(op_min[0], op_min[1], 's', color=color, markersize=10, label=f'{gn} threshold (TPR={op_min[1]:.2f})') ax.plot([0,1],[0,1],'k--',alpha=0.5) ax.set_xlabel("FPR"); ax.set_ylabel("TPR"); ax.set_title("ROC Curves — Equalized Odds Adjustment") ax.legend(fontsize=9); ax.grid(alpha=0.3) plt.tight_layout(); plt.savefig("equalized_odds.png", dpi=150); plt.show() ``` ### Step 3: Philosophical Framework Analysis ```python import pandas as pd import numpy as np import matplotlib.pyplot as plt # ------------------------------------------------------------------ # # Apply three moral frameworks to a concrete AI ethics scenario: # Should an AI parole board system deny release to a high-risk individual? # ------------------------------------------------------------------ # print(""" === AI Ethics Analysis: Automated Parole Decision System === SCENARIO: An AI system predicts recidivism risk. A defendant is classified as "high risk" (predicted probability = 0.72). The base rate for this demographic group is 0.35. The model has a false positive rate of 0.28 for this group. QUESTION: Should the parole board rely on this AI recommendation? """) frameworks = { "Utilitarian Analysis": { "core_principle": "Maximize aggregate welfare; minimize harm across all stakeholders.", "relevant_facts": [ "If high-risk prediction correct (72% confidence): denying parole prevents ~0.72 recidivism events", "False positive rate 28%: 28% of actual non-recidivists in this group are incorrectly flagged", "Incarceration costs ~$35,000/year; victim costs of crime vary widely", "Evidence suggests incarceration has mixed effects on recidivism long-term", ], "analysis": """ Expected utility calculation: - Let p=0.72 (predicted recid. prob), p_true≈0.35 (base rate) - Expected harm prevented = p_true × harm_of_crime - Expected harm imposed = (1-p_true) × cost_of_unjust_incarceration At base rate 0.35: denying parole has 65% chance of being wrong. Utilitarian calculation depends on harm weights — not obvious that denial maximizes welfare. HIGH-RISK DESIGNATION IS NOT SUFFICIENT for utilitarian justification. """, "verdict": "CONTEXT-DEPENDENT (depends on harm magnitudes)", }, "Deontological Analysis": { "core_principle": "Respect persons as ends; follow categorical duties regardless of outcomes.", "relevant_facts": [ "Individual has right to liberty and due process", "Statistical prediction punishes based on group membership, not individual action", "Kant's categorical imperative: act only on maxims you could universalize", "False positives violate individuals' rights even if outcomes-good aggregate", ], "analysis": """ Kantian objections to statistical risk prediction: 1. Treats individuals as means (to aggregate crime reduction) not ends 2. Statistical group-based predictions cannot be universalized without contradiction 3. Due process requires individual-level evidence, not probabilistic grouping 4. High FPR (28%) means systematic rights violations for the innocent The Rawlsian difference principle: acceptable only if worst-off benefit. Biased FPR (if higher for minorities) fails Rawlsian justice. """, "verdict": "REJECT AI-ONLY DECISION (rights violations are impermissible)", }, "Virtue Ethics Analysis": { "core_principle": "What would a person of good character (prudent, just, compassionate) do?", "relevant_facts": [ "A just judge weighs individual circumstances, not just actuarial scores", "Prudence requires understanding model limitations and error rates", "Compassion requires considering the human cost of false positives", "Epistemic humility: model ≠ ground truth about future behavior", ], "analysis": """ Virtues applied: - Justice: mechanical application of risk score without deliberation lacks justice - Prudence: a prudent judge uses the score as one input, not the final word - Integrity: honest acknowledgment that 28% FPR means significant uncertainty - Compassion: humanizes the defendant beyond statistical categories The virtuous judge treats AI as a tool for reflection, not delegation. """, "verdict": "USE AS ADVISORY INPUT ONLY (not decisive)", }, } for name, framework in frameworks.items(): print(f"n{'='*60}") print(f" {name}") print(f"{'='*60}") print(f"Core principle: {framework['core_principle']}") print("nKey facts:") for fact in framework['relevant_facts']: print(f" • {fact}") print(f"nAnalysis: {framework['analysis']}") print(f"nVERDICT: {framework['verdict']}") # ---- Summary visualization -------------------------------------- # fig, ax = plt.subplots(figsize=(10, 4)) framework_names = list(frameworks.keys()) verdicts = [f['verdict'].split('(')[0].strip() for f in frameworks.values()] colors = ['#f39c12', '#e74c3c', '#2ecc71'] y_pos = range(len(framework_names)) for pos, (name, verdict, color) in enumerate(zip(framework_names, verdicts, colors)): ax.barh(pos, 1, left=0, color=color, height=0.6, alpha=0.8, edgecolor='black', linewidth=0.7) ax.text(0.5, pos, verdict, ha='center', va='center', fontsize=11, fontweight='bold') ax.set_yticks(y_pos) ax.set_yticklabels([n.replace(' Analysis','') for n in framework_names], fontsize=10) ax.set_xlim(0, 1); ax.set_xticks([]) ax.set_title("Philosophical Framework Analysis:nAI Parole Decision System") ax.grid(False) plt.tight_layout() plt.savefig("ethics_frameworks.png", dpi=150) plt.show() ``` --- ## Advanced Usage ### Fairness-Accuracy Trade-Off Analysis ```python import numpy as np import matplotlib.pyplot as plt # ------------------------------------------------------------------ # # Visualize accuracy vs. fairness trade-off curve # ------------------------------------------------------------------ # thresholds = np.linspace(0.01, 0.99, 99) majority_mask = g_test_arr == 0 minority_mask = g_test_arr == 1 acc_vals, dpd_vals, eod_vals = [], [], [] for t in thresholds: y_t = (y_prob >= t).astype(int) acc = np.mean(y_t == y_test.values) sel_maj = y_t[majority_mask].mean() sel_min = y_t[minority_mask].mean() dp_diff = abs(sel_maj - sel_min) acc_vals.append(acc); dpd_vals.append(dp_diff) fig, ax = plt.subplots(figsize=(8, 5)) sc = ax.scatter(dpd_vals, acc_vals, c=thresholds, cmap='viridis', s=20, alpha=0.8) plt.colorbar(sc, ax=ax, label='Classification threshold') ax.axvline(0.05, color='red', linestyle='--', linewidth=1.5, label='|DP difference| = 0.05') ax.set_xlabel("Demographic Parity |Difference|") ax.set_ylabel("Accuracy") ax.set_title("Accuracy–Fairness Trade-off (Varying Threshold)") ax.legend(); ax.grid(alpha=0.3) plt.tight_layout(); plt.savefig("fairness_accuracy_tradeoff.png", dpi=150); plt.show() ``` --- ## Troubleshooting ### Error: `fairlearn.metrics.demographic_parity_difference` signature changed **Fix**: Check version — fairlearn 0.10 changed some function signatures: ```python # fairlearn >= 0.10 from fairlearn.metrics import demographic_parity_difference dpd = demographic_parity_difference(y_true, y_pred, sensitive_features=group) ``` ### Impossibility theorem confusion When fairness metrics conflict, it's expected (not a code bug): ```python # If base rates differ significantly, equalized odds AND calibration CANNOT both hold print(f"Group 0 base rate: {y_test.values[g_test_arr==0].mean():.3f}") print(f"Group 1 base rate: {y_test.values[g_test_arr==1].mean():.3f}") # Large difference → fairness metrics will be in tension ``` ### Version Compatibility | Package | Tested versions | Notes | |:--------|:----------------|:------| | fairlearn | 0.10 | API changed significantly from 0.7 to 0.10 | | scikit-learn | 1.3, 1.4 | Required by fairlearn | | scipy | 1.11, 1.12 | `stats.chi2_contingency` stable | --- ## External Resources ### Official Documentation - [fairlearn documentation](https://fairlearn.org) - [AI Fairness 360 (AIF360) by IBM](https://aif360.res.ibm.com) ### Key Papers - Chouldechova, A. (2017). *Fair prediction with disparate impact*. Big Data. - Hardt, M., Price, E. & Srebro, N. (2016). *Equality of opportunity in supervised learning*. NeurIPS. - Dwork, C. et al. (2012). *Fairness through awareness*. ITCS. --- ## Examples ### Example 1: Counterfactual Fairness Check ```python import numpy as np import pandas as pd from sklearn.linear_model import LogisticRegression # Counterfactual fairness: would the outcome change if group membership were flipped? np.random.seed(42) X_cf = X_test.copy() y_pred_cf_0 = clf.predict_proba(X_cf[g_test_arr==1])[:,1] # Minority features y_pred_orig_1 = y_prob[g_test_arr==1] # How much does predicted probability change on average? cf_change = np.abs(y_pred_cf_0 - y_pred_orig_1) print(f"Mean counterfactual prediction change (minority): {cf_change.mean():.4f}") print(f"Counterfactually fair if ~0; actual: {cf_change.mean():.4f}") ``` ### Example 2: SHAP Values for Bias Attribution ```python import numpy as np # Without SHAP: approximate feature importance for bias attribution coefs = dict(zip(X_train.columns, clf.coef_[0])) print("Logistic regression coefficients (proxy for feature importance):") for feat, coef in sorted(coefs.items(), key=lambda x: abs(x[1]), reverse=True): print(f" {feat}: {coef:+.4f}") print("nNote: high coefficient for 'income' or 'credit_score' may encode") print("indirect discrimination if these features are correlated with protected group.") ``` --- *Last updated: 2026-03-17 | Maintainer: @xjtulyc* *Issues: [GitHub Issues](https://github.com/xjtulyc/awesome-rosetta-skills/issues)*