Three depths of interaction

Interfaces have taken various visual and experiential forms, from command lines to skeuomorphism, expressive minimalism, voice, gesture control, and beyond (more on this in “The GUI moment”). However, they still often get in the way rather than enabling direct access to what we came for.

GUIs in particular tend to surface many functions, often regardless of importance. Layouts have stayed linked to books, magazines, and the publishing industry since the early days of graphical interfaces. Common elements like icons or buttons have retained their forms and positions, and common menus have kept their categories, letting repetition turn position into memory until an interface feels obvious. However, this is shifting as interfaces become more adaptable (DAS), contextual (ChatGPT, Claude …), and generated. While these developments can make interaction paths more efficient, they also risk removing the familiar cues that made these interfaces easy to use. Carrying familiar cues forward within evolving systems helps us stay confident and in control as we navigate simple or complex, adaptive interfaces.

The Three-Depth Principle tackles these interaction stages in three steps:

If you remove access, context disappears. Strip away orientation, and action becomes a guessing game. Finally, removing resolution makes the interaction seem aimless.

Depth is the noticeable stage we experience as we move from intention to outcome. However, it’s neither linear nor one-dimensional, since more than one action can occur within a single depth, and a single action can shift depths as understanding develops. The three stages illustrate our understanding of interaction as its path branches, reconnects, or unfolds within physical space. Having three depths doesn’t necessarily mean you’ll have three clicks, three screens, or three actions. The question is whether the shift from context to possibility to consequence is clear.

Reducing the number of clicks remains a widely adopted technique, and a simple rule (the three-click rule) suggests that key content should be reachable within three clicks.[2]

However, a User Interface Engineering study reported by Joshua Porter, in which 44 people attempted 620 tasks with over 8,000 clicks, found no correlation between clicks and task success. In Porter’s words: “there wasn’t any more likelihood of a user quitting after three clicks than after 12 clicks,” and “hardly anybody gave up after three clicks.” Satisfaction showed no clear relationship with the click count either: the unsatisfied share ranged from 46% to 61% across different path lengths.

Reducing the number of clicks doesn’t necessarily improve interaction.

Within the same experience, one click can open a clearly labeled destination, while another click can engage an unclear icon, and you won’t know what it does until after you click. This is a blind commitment, which happens when people act without understanding what the action means, where it leads, what will change, or whether they can reverse it (see Study 01).

Clearer commitments help people better understand what their actions mean and where they lead. Clicks capture movement, while commitments capture decisions.

Complex systems can have substantial depth. But let’s start with a simple, familiar example: When changing a setting on an iPhone, should I count the depths from unlocking the device or from opening Settings?

The answer begins with intent, and intent naturally precedes depths, setting the scope of the passage, even when it is partial or shifts along the way.

The Settings example demonstrates a simple, interactive frame. A frame can span an entire device or a software system, then narrow to a single function, control, and resulting state. This is why this principle operates with recursive frames. In the iOS Settings example, when a person lands in Settings, that resolves the device frame but not the broader intent to change a setting. In other words, Settings is open, but nothing has changed yet.

The Settings frame narrows to a specific (sub)feature, which then transitions to a new state. If the journey starts in Settings, the device frame is already set. However, if it begins in the pocket and results in a permission change, the analysis starts earlier, and every frame belongs to the analysis.

A screen or page change alone doesn’t create a new frame, but semantic resolution can. A new loop isn’t created simply because the interface opens another panel or screen. Recursion starts when the previous sequence produces a clear resolution and sets up a new, noticeable context with its own set of possibilities.

A system can also feel complex even when it exposes controls. A common example is an airplane cockpit, but let’s explore a more illustrative one: modular Eurorack systems (read: geeky synthesizers). These systems usually expose many controls. Even with years of training, the cognitive load can go through the roof. So despite the exposure of options, or (sometimes) “what you see is what you get,” one might be paralyzed by options, since every option can be combined with other options from the same set.

In research presented at CHI 2008, Gajos and colleagues tested an adaptive toolbar with a small group of 23 people. Increasing predictability and accuracy both led to significantly improved satisfaction. Contrary to the authors’ expectations, accuracy improvements affected performance, utilization, and some satisfaction ratings more than predictability. Under the random rule, eleven of the 23 called the toolbar “random,” “confusing,” or otherwise unpredictable, without prompting.[3]

Seventeen years later, in November 2025, Google Research showed a model generating the whole interface for a prompt, not only the content. Its evaluations showed that, with generation speed set aside, human raters strongly favored generated interfaces over standard LLM outputs, though pages created by human experts still came first.[4]

Another Google study, conducted with a group of 72 people who each used two versions of “a digital banking prototype,” found, in the authors’ words, that “the Adaptive generative version achieved a System Usability Scale (SUS) score of 84.38 (‘Excellent’), significantly outperforming the Deterministic version’s score of 53.96 (‘Poor’).” The authors credit the gap to a lower “navigation tax” and, in their reading, that is what improved perceived usability.[5]

Microsoft Research has turned the same question into a fellowship challenge, a brief for researchers to work on “AI-powered collaborative dynamic experiences that adapt and morph to users’ tasks and context.” The brief already says it: “Realizing this vision will require new computing abstractions, principles, and core interaction patterns, which empower users to work with these capabilities rather than be overwhelmed by them.” The stated aim is “the balance of dynamism with consistent and shared mental models.”[6]

Adaptation and generation can significantly enhance usability. In the toolbar study example, both predictability and accuracy improvements increased overall satisfaction, and improved accuracy also led to faster task completion and more frequent use of the adaptive toolbar. That supports adaptive and ambient interfaces (ambient meaning more than ubiquitous computing under a new name), with a design requirement: however the surface changes, the path must remain clear enough to follow.

At this point, designers and design thinkers have covered (almost) everything about the design process; let’s respectfully call it the legacy design process, which is valuable as we build on it and continue.

I’ll simplify it into a 2D analogy. Consider most of what we have designed so far moving from 2D to 3D. With that, we are adding a Z-axis that represents input and orientation across time, context, modalities, and their states. Besides deciding what appears, where, and how, designing for the dynamic includes multiple dimensions and modalities.

I used to see traditional design systems as responsive graphics standards manuals. However, that’s changing again. We can now confidently say design systems are becoming adaptive and evolving. This applies to all design aspects, including branding, colors, tones, experience, interactions, interactive elements, functionality, and how they’re expressed across different conditions. You can clearly see it’s not about individual segments but design as a whole.

The focus shifts from designing fixed components to establishing the rules that generate them. We will design fewer fixed interfaces and more rules by which interfaces are created. The designer becomes an author of conditions.

With a recursive structure of access, orientation, and resolution, The Three-Depth Principle joins a broader conversation about the future of interaction. Norman’s gulfs of execution and evaluation, Hierarchical Task Analysis, Information Foraging Theory, feedforward, cognitive walkthroughs, and human-AI interaction guidelines each illuminate part of this territory.[7][8][9][10][11][12]

The future of interaction design is built on design memory and depth, enabling digital environments to grow in complexity while meeting human interaction needs.

Interfaces are moving toward personalization, adapting to context and reshaping their appearance in real time, but the main goal remains the same: to translate complexity into a digestible form. To get there, we have to keep shaping the paths between intent and outcome.