Week 9: Forecast UIs and the uncertainty they hide

What a weather app, a poll tracker, and a stock chart have in common with autocomplete — and what each one does with the uncertainty a good forecast can't remove

26 April 2027

A forecast is a prediction with a probability attached, and this week’s claim is that most forecast interfaces spend their design effort hiding the probability rather than showing it. “70% chance of rain” gets rendered as a single rain icon. An election forecast’s “62–38, with real uncertainty on either side” gets rendered as a number that reads like a final score. The uncertainty was calculated — someone’s model produced a distribution, not a point — and the interface flattened it into something that looks more confident than the forecast actually is, because a confident-looking interface is easier to use and, not coincidentally, easier to sell.

This connects back to the whole semester rather than starting a new argument: every predictive system so far has had a confidence it could, in principle, expose, and mostly hasn’t. Autocorrect could show you its second- best guess and how close the scores were; it usually just picks one. A forecast UI is where that omission becomes legible, because forecasting is a field that already has the vocabulary — calibration, confidence intervals, base rates — for saying exactly how sure a prediction is, and interfaces routinely choose not to use it.

Outline#

• three forecast domains, three different ways of hiding the same kind of uncertainty
• calibration: what it would mean for a forecast interface to earn trust rather than just project it
• why “the model was technically right” and “the interface was honest” are different claims
• Assignment 2’s brief, due at the end of next week