Machine Learning
Synonyms: ML, statistical learning, pattern recognition, trained models, “the algorithm”
Definition
Use cases
- The cold start flop: A new user opens a music app and gets a stranger’s playlists. The model has no data on them yet, so the first session feels broken.
- The confident wrong tag: A photo app labels your mom as your aunt, and there’s no fix button. Users stop trusting every tag after that.
- The one-click spiral: One sourdough video turns someone’s feed into a bakery for a month. The model read a casual click as a strong preference.
How it's used in practice
- Check if ML is even needed: If a simple rule works (sort by date, filter by price), skip the model.
- Design for zero data: Ask 2 or 3 preference questions at signup, or show popular picks until the model learns.
- Show confidence where it matters: Use soft labels such as “Suggested for you” or “Is this Sarah?” on low-confidence results.
- Give users a correction path: Add “Not this person” and “Not interested” controls, then send those corrections to the data team.
- Prototype with real outputs: Test 50 real model results with users, bad ones included. Static mockups always show the perfect answer.
Challenges & limitations
- Black box outputs: Many models can’t explain their choices, so honest “Why am I seeing this?” copy is hard to write.
- Bad data, bad results: Thin or biased training data produces skewed outputs, and no amount of UI polish fixes that.
- Models change over time: Retraining shifts behavior, so a flow that tested well in March can feel off by June.
Free resources
- People + AI Guidebook (Google PAIR) — Google’s methods, patterns, and examples for designing with AI, written for UXers and PMs.
- Apple HIG: Machine learning — Apple’s guidance on ML roles, confidence, mistakes, and limitations in app UI.
- Guidelines for Human-AI Interaction (Microsoft) — the 18 evidence-based guidelines behind the HAX Toolkit.
