Calm Design Dies in Engagement Metrics

A copilot that won’t stop talking Link to heading

You open your editor to fix one small thing. The copilot is already there, finishing your sentence in pale grey ink. You didn’t ask. You did not even warm up your own thought. The system is not rude. It is efficient. It is also steering the work.

The quiet shift is this: you stop authoring and start choosing. You become the reviewer of a stream that never needed to exist.

Sometimes that trade is worth it. Sometimes it is exactly what you want.

The problem is the default. A helpful option becomes an ambient presence. The tool that “augments” starts to set the pace and the tone.

The design question is not whether copilots belong in our tools. That ship left the harbor. The question is what kind of copilot earns a place in someone’s attention without squatting there.

Two older frameworks do better work here than most modern “AI UX” decks: Calm Technology and Time Well Spent. 1 3 But the lesson is not what you expect. Put them together and they don’t add; they correct. Calm design patterns are necessary. They are also not enough. You cannot whisper inside a machine built to shout.

The attention budget is real Link to heading

Attention has a center and a periphery. The center is what you actively focus on. The periphery is the wider field that keeps you oriented, the engine noise you hear without staring at the dashboard. Weiser and Brown made that distinction structural. 1

Most software acts like the center is free real estate. Every feature wants a seat in the spotlight. Copilots intensify this because conversation is sticky. Humans are trained to respond. If the system speaks first, the social contract does the rest.

This is where calm stops being a mood and becomes a constraint.

If your copilot lives in the center by default, you are building a second job on top of the user’s job: managing the copilot. If it lives in the periphery until invited, you build something closer to good infrastructure: present, legible, quiet.

The trio that matters: stay peripheral, move to center only when summoned, preserve the user’s sense of orientation. Weiser and Brown called this third property “locatedness.” When your periphery works, you know what just happened, what is happening, and what is likely to happen next. You are not surprised in the bad way. 1

This is a practical spec for copilot behavior. Ghost text that becomes real only on acceptance. An escalation ladder from subtle indicator to one-line suggestion to richer explanation to blocking confirmation, each rung user-triggered or strongly justified by context. Assumptions surfaced early. The most humane sentence a copilot can produce is often boring: “I’m interpreting your intent as X. Is that right?”

At this point it is tempting to think: keep things peripheral and we are done.

This is where the second framework walks in with a clipboard.

The interface cannot be calm if the incentives are not Link to heading

Calm Technology tells you how to shape attention. Time Well Spent tells you why your organization will be tempted to do the opposite.

The Center for Humane Technology describes “attention-harvesting design” as a pattern tied to incentives that reward constant engagement. 3 Red notifications, algorithmic curation, intermittent reinforcement, infinite scroll. These are not random quirks. They are engagement tactics. And we are now adding copilots everywhere.

If the scoreboard rewards “more time in product,” your copilot will learn to keep the conversation going. It will become proactive. It will become needy. It will sound helpful while quietly extending the session.

A calm UI will not survive long inside an engagement-maximizing machine.

This is the correction the two frameworks make together. Calm Technology gives you the interaction patterns. Time Well Spent shows you the structural layer that will corrupt those patterns if left unaddressed. You can design the gentlest interface in the world; if the metrics underneath still optimize for minutes, the system will route around your good intentions.

Donella Meadows’ leverage points are useful precisely because they are boring. 5 In complex systems, the highest-leverage interventions sit upstream: goals, metrics, information flows. People push at the obvious places and the system snaps back.

So you need to change what success means. Not in a mission statement. In instrumentation.

Instead of optimizing for minutes, optimize for completion. For reduced backtracking. For fewer “undo” moments. For users reaching their own stopping points. For sessions that end because the job is done.

The problem is not the popup. It is the scoreboard.

Patterns that survive the incentive layer Link to heading

The frameworks get you oriented. The patterns change what ships. But these patterns only hold if the incentive layer lets them.

Default to pull. Let the user summon the copilot more often than the copilot summons the user. If you want to show suggested prompts, do it like a trailhead map: visible, skimmable, easy to ignore. The moment suggested prompts become a drip-feed of “next best actions,” you have recreated infinite scroll in a suit.

Make acceptance reversible, and make reversibility visible. A lot of copilot harm is not malicious. It is just speed. The model is faster than comprehension. Treat acceptance as provisional. Show the diff before you apply. Show the preview before you send. This is not a popup tax. It is how good tools work.

Ask for intent, then show work at the right altitude. A copilot that jumps straight to output is like a GPS that refuses to show the route. You cannot correct what you cannot see. Confirm intent in plain language. Propose a draft. Expose key assumptions compactly. Offer deeper rationale only when requested. Sometimes you want the view from 30,000 feet. Sometimes you want street level. Let the user choose.

Put “done” back into the interface. Attention-harvesting systems remove stopping cues. Infinite scroll is the canonical example. Copilots import the same pattern into work: a sidebar that always has one more suggestion, a chat that never ends, a stream of follow-ups that turns a simple task into a conversation you cannot leave without feeling rude. Humane copilots create stopping cues. They mark completion. They summarize what changed. Then they go quiet.

A small, optional wrap-up is more respectful than five enthusiastic follow-up questions.

The last mile Link to heading

If you want a single test for humane copilot design:

After a month of use, is the user more capable, or merely more dependent?

AI copilots sit right next to thought, next to the line you are about to write, the decision you are about to make. That proximity is not a reason to panic. It is a reason to be precise.

A good backcountry guide does not shout directions every ten meters. They point when the trail forks. They make sure you still know where you are. Then they let you walk.

But here is the part the backcountry metaphor misses: the guide’s employer is not paid by how long you stay on the trail. If it were, the guide would find reasons to keep walking. The trail would never end.

Calm design principles tell you how the guide should behave. Incentive alignment decides whether the guide is allowed to behave that way.

You need both. The pattern and the structure. The interface and the scoreboard.

A calm surface means nothing if the engine underneath is still revving.