Patterns / Recommendation & Prediction
When an AI product or system has made a classification that plausibly has more than one confident output, presenting multiple top outputs (for example above 80% confidence) is better than forcing the AI to present one output that it is not fully confident about.
When an AI product or system is undertaking a classifying task it should have an option to return ‘I don’t know’ or the equivalent rather than forcing a best guess.
When an AI product or system is undertaking a prediction task like showing the valuation of a property or the weather in the future show a range not an exact value which is false precision.
AI products and systems should surface recommendations that fall outside the established patterns to improve discovery and prevent the narrowing of preference over time.
When designing, recommenders in AI products and systems give a way for the user to bypass the system and discover something random with a clear call to action (‘Surprise me’).
AI Products and systems using recommender systems should flag when a recommendation is driven by popularity over personal relevance. Recommender systems systemically surface popular items.
When designing AI products and systems, present recommenders as a personalisation feature that improves over time. Do not present a recommender system as a degraded version of the full experience, full of random items.