A famous American psychologist, John Gottman, watched married couples talk in his lab. He said he could predict divorce with 93 percent accuracy. That sounds amazing.

But other researchers found a problem. When the method was tested on new couples, it did not work as well. A big study in 2020 also showed that whether a relationship gets better or worse is very hard to predict.

John Gottman is well known for observing couples in his "love lab." In one study, about 130 newlywed couples discussed a topic they disagreed about, and the researchers followed them for years. Gottman later reported that a model based on couples' emotional exchanges predicted divorce with 93 percent accuracy over 14 years. He also named four harmful patterns, including contempt, or looking down on a partner.

However, in 2001 other researchers pointed out that the prediction model had been built and tested on the same data. When they tested a similar method on new data, the accuracy dropped sharply. This does not mean Gottman's observations were useless. Knowing which kinds of talk can hurt a relationship is different from predicting the future of a particular couple.

Self-help books often claim that a few minutes of conversation can show whether a couple will divorce, and many of these claims trace back to the psychologist John Gottman. In his laboratory, newlywed couples talked about topics they disagreed on while their facial expressions, words and physical responses were recorded. Gottman singled out four especially damaging patterns, namely criticism, contempt, defensiveness and stonewalling.

A later 14-year analysis reported that a model combining emotional behaviors predicted divorce with 93 percent accuracy. Critics argued in 2001 that the figure was inflated. The model had been built and tested on the same couples, and without cross-validation a model can end up fitting the quirks of one data set instead of real patterns. When a similar model was tried on new data, its accuracy dropped steeply.

The critics also pointed out that when only a minority of couples divorce, a prediction of divorce is especially likely to be wrong. A 2020 project used machine learning to pool data from 43 long-term studies of couples. How partners saw the relationship, such as feeling appreciated and believing the other person was committed, explained how satisfied they were at the time. Changes in satisfaction, though, were almost impossible to forecast.