Mental models for AI design
Co-pilot, Assistant, Learning Partner: three ways people understand AI, and how each one changes the interface you'll design.
Part of the guide Design for AI
When someone uses AI, they bring a mental model. They don’t pick it consciously. It just shows up. And the mental model the person carries into the interaction shapes the tone, the expectation of control, and what counts as failure.
Three models keep appearing. Not the only ones, but the ones that cover 90% of the experiences you’ll encounter.
1. Co-pilot
The AI sits next to the person. It suggests, it doesn’t decide. The person keeps the wheel.
Concrete examples:
- Code completion (GitHub Copilot, Cursor)
- Email autocomplete (Gmail)
- Suggestions inside Figma or Notion
- Auto-grouping in organisation tools
The typical interface is discreet: a grey suggestion, a sidebar, a shortcut to accept or dismiss. The user keeps doing the work. The AI saves seconds, not whole tasks.
The bigger risk is giving the co-pilot too much autonomy. The moment it starts deciding instead of suggesting, the person loses the feeling of control and disengages from the output. The name says it: the pilot is still human.
2. Assistant
The person asks, gets an answer. The AI is reactive. It doesn’t act unless called.
Concrete examples:
- ChatGPT, Claude, Perplexity in conversation mode
- AI search (Google AI Overviews, You.com)
- Support bots
- Siri, Alexa in basic cases
The typical interface is a text box, an answer, maybe follow-ups. The user takes the client role, the AI the expert. The relationship doesn’t evolve much across sessions. Each conversation starts almost from scratch.
The risk is mistaking an Assistant for something that learns. When the person notices the AI doesn’t remember what they asked last week, they feel uncontextualised and trust collapses. Be honest about what carries across sessions and what doesn’t.
3. Learning Partner
The AI learns with the person, the person learns with the AI, the relationship evolves over time. Not a transaction. A path.
Concrete examples:
- Adaptive tutoring tools
- AI-driven habit coaches
- Recommendation systems that change with use
- Some experiences inside Notion AI or Cursor after heavy use
The typical interface needs to show that the system has changed. States like “I noticed you usually start projects this way” or “I learned you prefer this tone” make learning visible. Without that visibility, the evolution goes unseen and the person can’t trust it.
The risk is promising learning the product doesn’t deliver. If the AI doesn’t actually change with behaviour, calling it a Learning Partner is lying to the user. Better to admit it’s an Assistant.
How to identify the right model
Three quick questions for the product you’re designing:
- Who decides? If it’s always the person, it’s Co-pilot. If it’s the AI but only when called, it’s Assistant. If there’s a negotiation that evolves, it’s Learning Partner.
- How visible is the AI? Co-pilot is discreet, Assistant is central, Learning Partner is a character.
- Does the relationship change over time? If not, it isn’t a Learning Partner. Don’t pretend.
The most common mistake is designing an Assistant and promising a Learning Partner. People arrive expecting the system to remember, adjust, anticipate. It doesn’t. The frustration is proportional to the expectation the design created.
The optional fourth model: Autonomous agent
Worth naming, even though it isn’t on the classic list. An autonomous agent runs entire tasks without constant supervision. It’s qualitatively different from the three above: not next to you, not waiting, not in dialogue. It’s doing.
The difference from Learning Partner is autonomy to act, not just to learn. When you design for this, observability becomes vital. Covered in Observability in agentic UX.
What to do now
Pick a product of yours with AI inside. Ask three people who use it: “do you think this AI is here to suggest, to answer, or to learn with you?” If the answers diverge, your mental model isn’t landing.
More on the background to these models in the Design for AI guide. On the human-AI interaction patterns that still hold up, read 5 human-AI interaction types.
