When AI should and shouldn't be in the UX

The Principle

AI is not a universal upgrade. It excels at pattern matching, generation, and scaling certain tasks, but it is also unpredictable, sometimes hallucinatory, and often opaque. The responsible approach is to treat AI as one tool among many, not the default solution.

A practical decision framework asks these questions:

• Does the task benefit from probabilistic creativity or strict accuracy?
• Can the user easily verify or correct the output?
• Does the AI reduce cognitive load or increase it through uncertainty?
• Is the benefit worth the transparency, latency, and failure costs?
• Would the feature work well (or better) without AI?

If the honest answer leans no on several points, the feature should remain deterministic or human-driven. Defaulting to AI often stems from hype, competitive pressure, or technical excitement rather than user need.

In my own building, this framework was a corrective. Early on I added AI suggestions to almost every surface because the APIs were available and the demos looked impressive. Some features improved; many became noisy distractions or sources of user frustration. Stepping back and applying these questions helped me remove or redesign several features. The products became simpler, more reliable, and ultimately more useful.

Why It Matters for Design & Building

Treating AI as a default creates bloated, unpredictable interfaces that erode trust and increase cognitive load. A good decision framework forces intentionality: we add AI only where it genuinely serves the user, not where it serves our desire to appear cutting-edge.

As a Design Engineer, this framework now acts as a gatekeeper. In one project, we evaluated an AI-powered search enhancement. The model produced creative results but introduced frequent hallucinations in a domain where accuracy mattered. We kept a traditional deterministic search as primary and made AI an optional “explore” mode with clear uncertainty signals. Users got the best of both worlds without the risk of misleading defaults.

For calm technology, this restraint is essential. AI can be magical when used appropriately, but when added indiscriminately it introduces uncertainty, latency, and opacity that work against calm. The honest practice is to design human-first experiences and use AI as a supporting actor only when it clearly improves the user’s ability to achieve their goals.

Real-World Examples

Notion’s AI features show thoughtful restraint. AI is available as an opt-in assistant for writing and brainstorming rather than inserted everywhere by default. Users can invoke it when they want creative help, maintaining control and reducing noise in everyday workflows.

Many consumer apps during the 2023–2025 AI boom illustrate the opposite: AI chat buttons, auto-summaries, and generative features added to every screen with little user control or clear value. The result was often overwhelming interfaces that felt slower and less predictable than their non-AI predecessors.

A task management tool I worked on offered a mixed case. We initially added AI task suggestions by default. After user testing revealed frequent irrelevant recommendations, we changed to an explicit “suggest similar tasks” button. The feature became genuinely useful for some users without cluttering the core experience for everyone else.