Emily Campbell: The layers of AI experience

Emily Campbell is a VP of Design and AI UX advisor who has spent years studying how design practice needs to change for AI-native products. In this essay, she proposes a six-layer model for thinking about AI experience — a framework intended to replace the idea that design is primarily about controlling what users see on screen.

The six layers she identifies move from the surface down: user interface (the direction and oversight mechanisms through which users guide AI), context (the information about users and their environment that shapes system behavior), operational structure (what the model can access and execute), model (the training, capabilities, and behavioral characteristics of the underlying AI), governance (the rules that define what the AI is permitted to do), and emergence (the unpredictable behaviors that appear only through actual use at scale).

The key argument is that design decisions made at any one of these layers ripple through the others in ways that surface at the interface. A designer who only thinks about the interface layer — what text to show, where to place controls — is working with an incomplete picture of the system they’re shaping. Campbell traces this through several practical examples: how context architecture becomes critical to output quality, how governance constraints change what interaction patterns are even possible, and how emergence requires ongoing observation rather than up-front specification.

This represents a significant shift from the design methods that James Garrett and others codified in the early 2000s, when interfaces were deterministic. Campbell calls the contemporary alternative “probabilistic design” — an approach that accepts that the system can’t be fully specified in advance and focuses instead on identifying the points in a system where design decisions produce the largest downstream effects.

One implication she draws out is that interfaces should change as the relationship between a user and a system matures. Early interactions require more explicit direction because the system has little user context. As context accumulates, interface complexity can reduce — a principle she calls progressive autonomy. Designing for this trajectory requires thinking beyond individual screen states toward how a system should behave across a timeline of use.

The framework is dense in places and assumes readers who are already working on AI products. But for product designers, UX researchers, and design leads trying to develop a more rigorous vocabulary for AI-specific challenges, the layered model gives a structure that most practitioner discussions lack. It’s particularly useful for teams whose design reviews still treat AI outputs as fixed assets rather than as behavior shaped by interconnected system decisions.