Spoken language unfolds at a rapid pace, approximately 4-5 words per second. Have you ever wondered how listeners are generally able to keep up? A popular hypothesis is that people do so by anticipating what comes next, including which words come next. This anticipation, also called ‘prediction’, is thought to occur continuously, routinely, and implicitly. This means that it happens all the time, it does not require any conscious effort, and it happens even without someone trying to guess the next word. But why is prediction beneficial? This blog post explores the presumed benefits of prediction, but also its costs.

The key benefit of prediction is efficiency, which means “doing more with less“. Let’s demonstrate this with an example.

A friend tells you a story about a birthday and utters “and then the birthday boy blew out the …”. Research shows that, in such situations, you probably predict the next word to be ‘candle’, based on your knowledge of the described events (a birthday, blowing something out). This means that you activated the associated meaning of ‘candle’ stored in your long-term memory (what a candle looks like, what it is made of and is used for, etcetera), and possibly information about the word ‘candle’ itself (what sounds it consists of, that it is a noun, for example).

When the speaker continues, you would rapidly know whether your prediction is incorrect. For example, if the next word starts with ‘f’ your prediction was wrong, whereas ‘c’ would be consistent with your prediction. You only need to process any deviation from your prediction (called prediction error), and perhaps change your prediction. For example, upon hearing ‘f’ you might predict that the word is ‘flame’.

Without a prediction of ‘candle’, the word-initial ‘c’ tells you the word could be ‘candle’ but may equally likely be ‘captain’ or ‘candidate’. You would need to wait until you have more information to recognize ‘candle’ beyond any doubt.

Prediction also lets you quickly recognize incomplete or distorted words, or recognize different pronunciations of the same word. Let’s say your friend utters something like “condle“. Thanks to your prediction of “candle”, you can chalk up the deviation to an error rather than trying to match the input to words that start with ‘co’.

In sum, prediction is efficient because it limits the amount of new information that is processed at a given time, it allows for quick use of small chunks of input to be used as soon as they become available.

That all sounds great, but surely there must be a downside? A prediction cost? Despite a great deal of research, psycholinguists have yet to converge on a clear answer to this question.

Recent studies suggest that wrong predictions (e.g., “flame” when you predicted “candle”) incur a processing cost compared to when there was no prediction. This cost becomes visible in slower reading times, for example. However, this topic is still rather controversial, because some studies show that predictions are even beneficial when they’re wrong, especially when the encountered word is related in meaning to the predicted word (like candle-flame).

An even more controversial issue is whether the generation of the prediction itself is costly. Some researchers argue that prediction is a separate brain process that incurs a metabolic cost. This could explain why prediction does not occur all the time and in everyone; sometimes the mental resources for prediction are not available. However, an actual metabolic cost has not yet been demonstrated. Others argue that predictions naturally emerge from the same processes with which we comprehend input, and in that sense, they are for free. Predictions might be reduced if comprehension does not proceed entirely smoothly.

The potential cost of prediction is currently a trending topic in psycholinguistics. Researchers try to figure out whether such costs are real and what role they play. This will pave the way towards a better understanding of what makes language comprehension efficient.