The Human in the Channel

Most AI governance frameworks share a common architectural assumption: the AI system is the variable. The human is the control. The AI gets constrained, audited, evaluated. The human operator reviews outputs, sets policy, and decides what to act on.

This assumption is wrong in a specific, structural way.

The human is not at the end of the channel. The human is in the channel. And humans drift.

What Changes Over Time

The question to ask — following Feynman’s practice of looking at what things do rather than what they’re called — is not “is the human operator reliable?” It is: what does the human do differently at hour one versus hour twelve?

Three things change predictably. They have names.

1. Context Drift

At the start of a session, the operator has sharp situational awareness. They remember what was established at the beginning: the constraints they set, the scope they defined, the commitments the system made early in the conversation. This state is active and accessible.

By hour twelve — after dozens of turns, extended use, parallel tasks — the early session state has been cognitively buried. The human is no longer cross-referencing new outputs against the initial framing. Early commitments that should still be binding have become invisible to the operator even when they remain formally active.

Context drift is not inattention. It is a predictable property of working memory under load. The intervention is architectural: session anchoring. At structured intervals, the system surfaces a canonical summary of active constraints, scope boundaries, and prior commitments. This is a channel re-sync event — an explicit mechanism to realign the human’s working model of the session state with the system’s structural record.

2. Delegation Drift

Early in a deployment, operators verify carefully. They check outputs against independent sources. They confirm generated code against known requirements.

After a run of correct, useful outputs, the verification threshold drops. The system has earned trust. This is also rational behaviour. The problem is that earned trust tends not to stay bounded to the task class where it was earned. It transfers. The operator begins delegating tasks that are adjacent to — but structurally different from — what was verified. The trust that was warranted for task class A gets applied to task class B.

The structural intervention is explicit scope contracts: a formal agreement at session start specifying what the system is validated for in this context. When an operator query falls outside the contracted scope, the system signals this explicitly. The signal is in the architecture, not in the operator’s attention.

3. Verification Collapse

The third change is the most insidious. When an operator has verified twenty consecutive outputs and all twenty were correct, the checking behaviour changes. This is not negligence — it is Bayesian. Past performance is evidence. The operator updates their model of the system as reliable and reduces the overhead allocated to verification.

The failure arrives when the system’s reliability in a specific context changes — due to a different input distribution, a topic at the edge of the training data — and the reduced verification overhead means the first incorrect output isn’t caught.

The structural intervention is graduated autonomy: the system earns reduced-verification authorisation for specific task classes through a validated track record, drawn from audit trail evidence. The authorisation is specific to the task class. It does not transfer.

Why “Human-in-the-Loop” Is Insufficient

The phrase “human-in-the-loop” appears frequently in AI governance policy as a safety property. It describes a deployment where a human reviews AI output before it is acted on.

This is better than no human review. It is not channel engineering.

Adding a human reviewer to the end of an ungoverned channel does not address context drift, delegation drift, or verification collapse. It adds a human to the output layer and assumes that human will maintain consistent, drift-free, accurate review behaviour across the deployment lifetime. Under real operating conditions, this assumption fails in the three predictable ways described above.

A human-in-loop claim without structural support for the human layer is governing the AI side of the channel and leaving the human side unarchitected. The channel includes the human. Governing half the channel is not equivalent to governing the channel.

A Builder Checklist

For any AI deployment that includes human operators in the signal path, the governance architecture requires:

• Session anchoring: Structured re-sync events that surface active constraints, scope, and prior commitments at defined intervals.
• Explicit scope contracts: Formal validation scope agreed at session or deployment configuration time; out-of-scope queries flagged at the system level, not at the operator’s discretion.
• Graduated autonomy: Task-class-specific authorisation for reduced verification overhead, based on audit trail evidence — not subjective trust. Authorisation does not transfer across task classes.
• Drift detection signals: Indicators in the interface that surface when operator behaviour has moved significantly from session-start patterns.

None of these are features for careful users. They are properties of a channel architecture that remains reliable at the human layer across the full deployment lifetime.

Next: SIG·6 — Signal at Scale. Engineering human drift out of the channel leads directly to scale. What does channel integrity look like for one million simultaneous transmission paths?