Week 3: Query completion and the autocomplete confession

What a search box's suggestions reveal about the people typing into it, and why that turned out to be a design problem

8 March 2027

A T9 disambiguator predicts from one person’s keypresses and a fixed dictionary. A search box’s autocomplete predicts from everyone’s past queries, which means it isn’t really finishing your sentence — it’s telling you what other people, in aggregate, tend to ask. That’s a genuinely different kind of prediction, and it produces a genuinely different kind of failure: not “wrong word,” but “accurate reflection of something the platform would rather not display in a dropdown.”

Search engines have spent a long time patching this after the fact — filtering suggestions that complete a person’s name into a harassment query, or a medical symptom into a worst-case diagnosis, or a demographic group into a slur. Every patch is a value judgement about which aggregate truths are fit to surface, made by whoever owns the suggestion box. The lecture treats query completion as a case where the “it’s just showing what people search for” defence runs out fastest, because the completion is attributed to the platform, phrased as if the platform is finishing your thought, not reporting a statistic.

Outline#

• how query completion actually works: frequency, recency, and personalisation
• three real cases where a completion became a controversy, and what changed after each one
• the gap between “this is popular” and “this is what I meant to ask”
• what the lab’s audit is looking for