Every interface is an act of compression.

It takes a complex, multidimensional reality—technical, social, temporal, institutional—and collapses it into something a human can perceive and act upon. Buttons replace procedures. Text replaces uncertainty. Defaults replace judgment. This is not a flaw; it is the only way systems become usable at all.

The problem arises when compression is treated as neutral.

In design discourse, abstraction is often framed as simplification in service of clarity. In practice, abstraction is a redistribution of attention and responsibility. What is compressed out does not disappear; it is displaced—onto users, operators, downstream institutions, or future crises. An interface does not merely show less; it decides what matters now.

Cartography offers a precise analogy. A map is not a smaller version of the world. It is a selective model, optimized for a purpose: navigation, ownership, elevation, threat. Contour lines compress vertical reality into gradients that enable reasoning at a distance. Crucially, good maps signal their abstraction. Legends, scales, projections, and omissions are explicit. No competent user confuses a map with the terrain because the interface teaches them not to.

Bad interfaces lie by omission.

They suggest completeness where there is none, stability where there is uncertainty, and authority where there is only inference. In AI systems, this lie is amplified by fluency. Linguistic smoothness disguises probabilistic collapse. A single answer stands in for a distribution of possibilities. What appears as clarity is often compression-induced overconfidence.

This is not a philosophical objection; it is a design failure mode with a long history.

In safety-critical systems, investigators repeatedly find that operators did not misunderstand the rules—they misunderstood the system state as presented to them. Interfaces that suppress ambiguity in order to reduce cognitive load often create false certainty, which is far more dangerous under stress than acknowledged uncertainty. The Therac-25 radiation accidents, aviation mode confusion incidents, and clinical prescribing errors all trace back to interfaces that hid relevant complexity while projecting ease of use.

AI interfaces replicate this pattern at scale.

They compress:

  • epistemic uncertainty into declarative language,
  • contested interpretations into single narratives,
  • partial context into confident summaries.

They rarely disclose what was excluded, how much variation existed, or where the model is extrapolating rather than recalling. The result is not just error, but epistemic miscalibration: users learn the wrong lessons about what the system knows and how it knows it.

The ACP design arc treats compression as a first-order governance concern.

Governance does not fail only when outputs are wrong. It fails when systems teach users to trust them inappropriately. An interface that never signals uncertainty trains users to ignore it elsewhere. An interface that never refuses trains users to escalate reliance. An interface that hides compression teaches users that compression does not exist.

This is why the NFPA analogy matters again. Industrial systems do not pretend risk can be abstracted away. They encode it visibly. Warning labels, access barriers, lockout procedures, and graded alerts are all design responses to the fact that humans must operate within compressed representations of dangerous systems. The goal is not simplicity, but legibility under constraint.

AI systems currently optimize for the opposite: seamlessness.

Seamlessness feels humane. It is also deceptive. It suggests that no boundary exists between question and answer, intent and authority, request and permission. In institutional contexts, this is untenable. A system that collapses deliberation into immediacy is not assisting governance—it is bypassing it.

Language learning provides a useful contrast.

In effective language pedagogy, abstraction is managed carefully. Learners are shown patterns, but also exceptions. They practice scenarios, but encounter refusal, renegotiation, and repair. Meaning is negotiated, not delivered whole. Good teachers do not compress away difficulty; they stage it. They signal where uncertainty lives and where effort is required.

This is exactly the posture ACP advocates for AI interfaces more broadly.

Compression is unavoidable.
Abstraction is necessary.
Opacity is optional.

An interface that acknowledges its own compression—by signaling uncertainty, bounding authority, and exposing limits—does not weaken the system. It strengthens trust by aligning perception with reality.

The lie of neutral interfaces is not that they simplify.
It is that they pretend simplification has no consequences.