When AI Drifts

AI is moving from experimentation to core operations across government. As this shift accelerates, a quiet but consequential reality is emerging: AI systems drift, and institutions drift around them. For senior leaders, the strategic challenge is not the technology itself. It is the co‑evolution of systems and institutions and the governance gaps that open between them.

This article explains why drift occurs, how it reshapes public authority, and what executives can do to govern AI as a living system rather than a static asset.

The Strategic Problem: AI Does Not Stay Where You Put It

Every AI system is built on assumptions about the world: patterns in data, definitions of risk, and models of behaviour. Once deployed into real public environments, those assumptions begin to degrade. This is technical drift, the gradual divergence between design‑time expectations and operational‑time performance.

For executives, the strategic implication is clear:

AI performance is not a procurement outcome. It is a stewardship obligation.

Three forms of drift matter most:

• Data drift: Inputs shift as social, economic, or policy conditions change.

• Concept drift: The meaning of key outcomes evolves, such as risk, need, or eligibility.

• Performance drift: The model’s accuracy erodes even when nothing obvious changes.

These shifts are normal. They are not evidence of failure. They are evidence that AI is interacting with a dynamic society.

The Hidden Challenge: Institutions Drift Too

While the model is drifting, the institution around it is also changing—often faster than leaders realise. This institutional drift is subtle but powerful:

• Processes adapt to the system, rather than the system adapting to policy intent.

• Accountability becomes blurred, as decisions are attributed to “the model” rather than to authorised roles.

• Norms shift, with pilots becoming de facto policy tools without formal approval.

• Capability erodes, as staff lose confidence or skill in challenging automated outputs.

Institutional drift is rarely deliberate. It emerges from frontline pragmatism, resource constraints, and the natural tendency of organisations to routinise new tools.

For executives, the risk is not that AI changes decisions, it is that AI quietly changes the institution making the decisions.

• The Feedback Loop: When Technical and Institutional Drift Reinforce Each Other The most significant governance risk arises when the two forms of drift interact.

• As the model drifts, staff compensate informally.

• As staff compensate, the model’s behaviour becomes harder to interpret.

• As interpretation weakens, oversight becomes symbolic rather than substantive.

• As oversight weakens, drift accelerates.

This feedback loop can shift discretion, fairness, and accountability in ways no policy process ever authorised.

For senior leaders, this is the core strategic insight:

AI does not simply automate decisions—it reshapes the governance environment in which decisions are made.

Executive Actions That Anchor Stability in a Drifting System

Across jurisdictions, four practices are emerging as the new baseline for executive stewardship.

A. Establish AI as a monitored system, not a deployed product

Drift detection should sit alongside financial audits, risk reviews, and performance reporting. Executives should expect dashboards, thresholds, and escalation pathways, not static assurance statements.

B. Build institutional drift detection into governance

Governance must ask two questions:

• Is the model drifting?

• Are we drifting around the model?

This requires structured observation of process changes, decision patterns, and shifts in organisational behaviour.

C. Clarify human authority in AI‑enabled decisions

Oversight only works when roles, responsibilities, and override powers are explicit. Executives must ensure that “human in the loop” is a capability, not a slogan.

D. Treat drift as a leadership signal, not a technical fault

Drift reveals where policy assumptions, operational realities, and institutional routines are misaligned. Leaders who treat drift as intelligence, not noise, gain early visibility into emerging risks.

The Leadership Imperative

AI is now part of the state’s decision‑making infrastructure. As with any infrastructure, the question is not whether it will change over time - it will!

The question is whether institutions will change with intention or by accident.

Executives who understand drift as a governance phenomenon, not just a technical one, are better positioned to protect legitimacy, fairness, and public trust as AI becomes embedded in the machinery of government.