Typing, routing, and systems that change what they measure

Three small models you use hourly

Your keyboard. The word suggestions above the letters come from a small language model running on the phone itself, not in a data centre, which is why they work in flight mode. Swipe typing is the same model reading a smeared path across the keys and predicting the intended word. Autocorrect is that prediction overruling what your fingers did, which is why it is confident and occasionally maddening.

Most keyboards learn your own vocabulary locally. Some contribute to a shared model through federated learning, in which the phone computes an update from your typing and sends only that update, never the text. It is a genuine privacy improvement and it is not the same as sending nothing.

Search autocomplete. The suggestions that appear as you type a query are predicted mainly from what other people have typed. That makes them a mirror held up to a population rather than an answer, which is why they periodically cause scandals and why providers maintain long block lists.

Maps. Two separate predictions are running. Travel time is estimated from historical speeds on each road segment, live traffic, road type, and time of day; Google reported in 2020 that a graph neural network cut its inaccurate arrival times substantially in several cities. Then routing picks the path that minimises predicted travel time.

The idea worth taking from this lesson

Maps illustrate something no other example in this block shows as cleanly: a prediction that people act on changes the thing it predicted.

The app finds a quiet residential street that is faster. It routes traffic down it. Enough traffic arrives that the street is no longer quiet. The model observes the new congestion and stops recommending it. The prediction was correct at the moment it was made and became false because it was made.

Communities have responded by petitioning for road closures and speed bumps, and several towns in the United States and Europe have altered street layouts specifically because navigation apps redirected motorway traffic through them. This is a case of a model reshaping the physical world in order to keep being right about it.

You will meet this loop repeatedly. A hiring model trained on past hires shapes who gets hired next, and the next model trains on that. A feed that predicts what you will watch determines what you are shown to watch. Once a prediction drives an action, the record it learns from is no longer independent of it.

The arrival time is a range pretending to be a moment

The app tells you 34 minutes. What the model produced was closer to a spread of plausible durations, and 34 is a chosen point within it.

Which point gets chosen is a product decision. Providers generally lean towards the pessimistic side, because arriving five minutes early feels like the app worked and arriving five minutes late feels like it lied. The asymmetry is in how people react, not in the traffic.

This explains something people notice and misread. If your arrival estimate is usually a little conservative, that is not the model being poor. It is the model being tuned for the complaint it wants to avoid. A route on an unpredictable road (one flood, one level crossing, one market day) has a wide spread, and the single number hides exactly how wide. When a journey must not be late, the estimate is the wrong tool: leave earlier than any of it suggests.

What routing optimises, and what it does not

The route you get minimises predicted travel time. It does not consider whether you find the road frightening, whether it is safe to cycle, whether it passes a school at closing time, or whether sending you through a narrow lane is fair to the people living on it. Those quantities are not in the objective, so the system is indifferent to them, not hostile.

Some apps now offer a fuel-efficient or low-emission route, which is a different objective made available as an option. That is the honest way to handle it: expose the choice rather than pretend a single best route exists.

Practical notes

• Offline maps work on your phone with no connection, which is worth setting up before travelling; routing quality drops because live traffic is missing, but the map and the route remain.
• OpenStreetMap, with free apps such as OsmAnd or Organic Maps, is the free path here, and in many towns it is better than the commercial maps because local people edit it directly. You can fix your own street.
• Keyboard predictions can be cleared. Every major keyboard has a “reset learned words” control, useful if it has picked up a typo or something private.
• If autocorrect keeps mangling a name or a word in your language, adding it to the personal dictionary once fixes it permanently, because that dictionary overrides the model.