Algorithmic Accountability

Definition

Algorithmic accountability is “ensuring that automated decision-making systems are fair, explainable, and free from discriminatory bias.” In the wizard post it is the second of three principles and the audit engine’s “bias checks” branch. Its distinguishing feature is that fairness is measured, not asserted — “the wizard’s second trick is auditing the models that allocate capital.”

How It Works

Audit models with an open-source fairness toolkit — the post names Fairlearn and IBM’s AI Fairness 360 (AIF360). The core pattern is a grouped metric frame plus a parity measure:

# scripts/audit_model.py
from fairlearn.metrics import MetricFrame, selection_rate, demographic_parity_difference
from sklearn.metrics import accuracy_score

metrics = MetricFrame(
    metrics={"accuracy": accuracy_score, "selection_rate": selection_rate},
    y_true=y_true,
    y_pred=y_pred,
    sensitive_features=sensitive,   # e.g. a zip-code-derived demographic
)

print(metrics.by_group)
print("Demographic parity gap:",
      demographic_parity_difference(y_true, y_pred, sensitive_features=sensitive))

The decision rule is unambiguous: “If the gap is non-trivial, that’s your spellcheck failing. Document it, file an issue, and refuse to ship until it’s addressed.” When a gap is found, the FAQ prescribes: don’t ship; investigate root causes (training-data skew, feature proxies, labeling bias); and consider mitigations — re-weighting, threshold optimization, or removing the model entirely.

Why It Matters

Capital- and credit-allocation models can “deepen inequality” at scale and silently. Accountability turns that from an invisible externality into a reviewable artifact: every model should carry “a published model card and a fairness report,” surfaced on the same dashboard as everything else. The post’s governing maxim: “‘We knew and shipped anyway’ is worse than ‘we found nothing.’“

Examples

• Parity gap gate — the Fairlearn snippet above, wired into CI so a regression in demographic_parity_difference blocks a merge.
• Proxy hunting — a “zip-code-derived demographic” sensitive feature illustrates how a seemingly neutral input can act as a proxy for a protected attribute.

Connections

• Fairlearn — the toolkit implementing the metrics above (with AIF360 as a sibling).
• Radical Transparency — supplies the auditable data these checks run against.
• Ethical Engineering — the ethic behind “refuse to ship.”
• ESG Investing — accountability reports feed the governance (“G”) dimension.

Sources

• Wizard Topples Capitalist Dominance Ingeniously — “Example 2: Algorithmic Accountability Check” and the fairness-audit FAQ.