Semantic Drift

What a model meant by a word last year is not necessarily what it means by the same word today.

This is one of the stranger facts about modern AI systems, and most of the people deploying them have not quite absorbed it. A language model trained on a corpus from a particular moment has a set of associations baked into it. The word inflation meant one bundle of things in the data from ten years ago, something slightly different three years ago, something different again today. The word friend had certain associations before social media and other associations after. When a model trained on old data is asked questions about new situations, it gives you answers that are internally consistent with how it once understood the world, not with how the world actually is now. This is not a bug. It is the nature of a model that was finalised at a point in time.

In educational contexts this matters more than in most others, because the vocabulary of a field shifts underneath its feet. What counted as good pedagogy a decade ago is not what counts now. What counted as safe adaptive design five years ago is not what counts now. A model built on older material can quietly give advice that sounds thoughtful and is several shifts out of date, and the people receiving the advice have no way to tell from the confident tone that the language has moved on. Worse still, some systems keep learning from user interactions, and their meanings drift in directions nobody planned. Detecting this drift, deciding when to retrain, deciding whose current usage counts as the right one, are all active problems without clean answers. The model sounds the same. The world it was trained for is gone. When was the last time you noticed that a word you were using had quietly changed its meaning?

Last week we looked at what consent really protects and what it does not. This week we look at how the very meaning of words drifts inside an AI system and at how rarely anyone notices.