AI feature UX checklist

PAIR's first recommendation is to find where user needs and AI strengths meet. Many problems don't need AI, and a rule-based solution is easier to explain and test. PAIR: User Needs + Defining Success

PAIR suggests automating tasks people find tedious or repetitive, and augmenting tasks people enjoy, that carry social value, or where people's preferences vary widely. PAIR: User Needs + Defining Success

This decides whether you tune for precision (fewer wrong results) or recall (fewer missed ones). PAIR's example: a fire alarm that fails to go off is far worse than an occasional false alarm. PAIR: User Needs + Defining Success

Apple's guidelines ask for a great experience even without the generative feature, and a non-AI fallback where possible. Apple HIG: Generative AI

People build a mental model from the first thing they see. Say what the feature does and, just as plainly, what it doesn't do. Example prompts or a short tour help with open-ended inputs. Microsoft HAX: Guidelines for Human-AI Interaction

The paper's description is to help the user understand how often the AI system may make mistakes. Hedged wording (“we think you'll like”) is one way; telling people where it's weakest is another. Microsoft HAX: Guidelines for Human-AI Interaction

Decide when the system should act or interrupt based on what the person is doing and where they are. A suggestion that arrives mid-task can cost more than it saves. Microsoft HAX: Guidelines for Human-AI Interaction

Output should relate to the person's current task and environment, not to everything the model could say. Microsoft HAX: Guidelines for Human-AI Interaction

Tone, formality and behaviour should be what people expect given their social and cultural context. Check this per market, not only in your own language. Microsoft HAX: Guidelines for Human-AI Interaction

Models learn from data that over-represents common cases, so they can repeat stereotypes. Test with a diverse set of people and inputs, and ask for information rather than inferring personal characteristics. Microsoft HAX: Guidelines for Human-AI Interaction

Make it easy to call up the AI when people want it, for example from a visible button or a keyboard shortcut, instead of only when the system decides to appear. Microsoft HAX: Guidelines for Human-AI Interaction

An unwanted suggestion should take one action to ignore or close, and should not block the content underneath. Microsoft HAX: Guidelines for Human-AI Interaction

Make it easy to edit, refine or recover when the output is wrong. Put Edit, Undo and Retry next to the generated content. Microsoft HAX: Guidelines for Human-AI Interaction

When the system is unsure what someone wants, it should ask, offer a few options, or do less, rather than act confidently on a guess. Microsoft HAX: Guidelines for Human-AI Interaction

Let people see an explanation of why the AI behaved as it did: the inputs it used, the source it drew on, or the rule it applied (“fastest route”). Microsoft HAX: Guidelines for Human-AI Interaction

Keep short-term memory so people can refer back (“make it shorter”, “who is he married to?”) without repeating themselves. Microsoft HAX: Guidelines for Human-AI Interaction

Personalise from what people do over time, where that helps them, and within what they have agreed to. Microsoft HAX: Guidelines for Human-AI Interaction

Limit disruptive changes. A list that reshuffles while someone is reading it breaks their place and their trust. Microsoft HAX: Guidelines for Human-AI Interaction

Let people say what they want during normal use, for example marking an item as important or not relevant, not only through a survey. Microsoft HAX: Guidelines for Human-AI Interaction

Show how an action will change what the AI does next, such as “hiding this ad will change which ads you see”. Microsoft HAX: Guidelines for Human-AI Interaction

Let people customise what the AI monitors and how it behaves, including switching personalisation or the feature off. Microsoft HAX: Guidelines for Human-AI Interaction

Tell people when the AI adds or updates its capabilities. A model upgrade that changes outputs is a change people should hear about. Microsoft HAX: Guidelines for Human-AI Interaction

PAIR's goal is not maximum trust. Over-trust leads people to accept wrong answers; under-trust means a useful feature goes unused. PAIR: Explainability + Trust

PAIR recommends explaining the scope and reach of the data used, and how to remove it. PAIR: Explainability + Trust

A music recommendation needs little explanation; a health, money or legal suggestion needs enough for the person to judge it. PAIR: Explainability + Trust

PAIR notes that numeric confidence needs people to understand probability; categories or a set of alternatives are often clearer. PAIR: Explainability + Trust

PAIR advises against a long list of features up front, and suggests stating benefits, current limitations and how the product will change over time. PAIR: Mental Models

PAIR: “make it extremely clear that the product is not a human”. Human-like framing raises expectations the system can't meet. PAIR: Mental Models