Under AI Conditions 2.3C.3-2

When AI or algorithmic systems make or recommend organizational decisions, encoded objective law becomes more critical, not less. An algorithm is a decision rule that operates at scale, constantly, without fatigue. If the rule is wrong, it is wrong at every application.

This sidebar addresses a specific structural challenge: how to maintain accountability, transparency, and reversibility when decisions are mediated by AI. The governing principle remains the same: no claim may acquire coercive force without defined appeals, and no decision may be enforced without the person affected being able to understand why.

Phase I's fourth load case — AI-mediated coordination (1.1.1) — named the structural condition: consequential decisions about persons produced by computational systems that scale faster than human oversight. The organizational question is not whether to use AI but how to govern it structurally.

Three Structural Layers for AI Accountability

Algorithmic Transparency — Before Deployment

Before an AI system makes decisions that affect people, the institution must understand what the system does.

Model documentation. For each AI system, document: what the system decides, what data it was trained on, what it was trained to optimize, what it was not trained to optimize (important to name what it does not do), known limitations, and who validated the model before deployment.

Human intelligibility check. Before deployment, audit whether a human decision-maker can understand why the system recommends what it recommends. Take a sample of recent decisions the system would make. For each, ask: can a human explain why? If more than thirty percent are unexplainable, the system is too opaque for high-stakes decisions.

Drift tracking. AI systems trained on historical data learn historical patterns — including historical discrimination. As conditions change, the algorithm can drift from its original purpose. Document when the model was trained, when it is retrained, what changed, and who reviewed the new model before deployment.

Explainability guardrails. For high-stakes decisions, require the system to produce not just a recommendation but an explanation: the factors considered, the confidence level, and the alternatives ranked higher. The explanation is not perfect, but it gives a decision-maker material to evaluate.

Algorithmic Accountability — After Deployment

Even with transparency, algorithms make decisions. The institution needs to know when they are failing.

Adverse impact auditing. Regularly audit whether the system produces systematically different outcomes for different groups with similar circumstances. When adverse impact is detected: stop deploying the system to new decisions, investigate the cause, decide whether to fix the model or retire it, and notify people affected by past decisions. This is expensive. Do it anyway.

Decision logs. Every decision made by an AI system is logged with the decision, the confidence level, the actual outcome, and the explanation factors. The log enables calibration measurement: when the system says seventy-eight percent confident, is it right seventy-eight percent of the time? Poor calibration signals that the system should not be trusted.

Feedback loops. Connect the decision back to the outcome. Did the system recommend hiring this person? How did they perform? Without this loop, algorithmic accuracy cannot be measured.

Randomized audits. Periodically audit counterfactuals — decisions the algorithm did not make. If the system recommended rejection but the person would have succeeded, the system has a systematic blind spot.

Algorithmic Control — Override and Rollback

Even with transparency and accountability, the institution must be able to stop or reverse an algorithmic decision.

Mandatory human review for high-stakes decisions. Hiring, termination, denial of access, resource allocation — do not let the algorithm decide alone. Require human review before enforcement. The review does not have to reverse the algorithm often. It just has to be possible.

Defined override triggers. When should a human override? System confidence below a defined threshold, outcome affecting a protected class, decision contradicting a stated organizational value, or person affected requesting human review. Document triggers, train people, and track override frequency. If algorithmic decisions are never overridden, either the system is perfect (unlikely) or people are not exercising their authority (problematic).

Rollback procedures. If an algorithmic decision causes harm, the institution has a procedure to identify all decisions made by the flawed model, re-review affected cases with human judgment, offer correction to anyone wrongly affected, and notify all affected persons of the error.

Override audit trail. Every override is logged: what the algorithm recommended, what the human decided instead, and why. Over time, the trail reveals whether the human is consistently correcting in one direction (the system is systematically biased), whether overrides are random (the system is not providing useful input), or whether overrides are explainable (the human has coherent criteria).

The Explainability-Performance Tradeoff

Some AI systems are inherently less explainable than others. The institution must decide: can we afford the loss of explainability for the gain in performance?

The governing principle: increase explainability requirements as stakes increase. Low-stakes, high-frequency decisions (prioritizing support tickets) can tolerate opacity if misclassification is annoying but not catastrophic. High-stakes decisions (hiring, termination, credit) require explainability to take priority over performance. Use a less sophisticated but more interpretable system rather than a more powerful but opaque one.

Do not automate a decision just because you can. Automate it because you understand why the algorithm decides what it does, you have a way to know when it is failing, you can override or reverse it if it goes wrong, and automating it produces better outcomes than human judgment — not just faster outcomes. If you cannot answer yes to all four, the decision should remain with humans, supported by better tools and training.

Integration Points

• Enforce-in-code flow (2.3C.3-1): When evaluating whether a decision should be encoded, factor in whether encoding via AI is feasible and whether the AI system can meet the transparency, accountability, and control requirements.
• Data governance (2.3C.3-3): AI systems depend on data. Data governance rules — minimization, least privilege, selective disclosure — apply to training data and to data the system accesses during operation.
• Organization evidence pack (2.3C.4-3): Model documentation, audit logs, and override decisions become part of the evidence that the institution is managing algorithmic governance responsibly.
• Incident response runbook (2.3C.4-2): When an algorithmic decision causes harm, the incident response runbook guides investigation and rollback.