Can AI predict who splits up?

At first glance, it sounds appealing. Record a couple discussing a difficult issue for a few minutes, analyse the recording with the latest AI

At first glance, it sounds appealing.

Record a couple discussing a difficult issue for a few minutes, analyse the recording with the latest AI technology, and presto, a recent study suggests you can predict with 79 per cent accuracy who will divorce within the next two years.

BBC Future has just produced an article about this.

But think about the concept a little longer and the idea is both horrific and – fortunately – implausible.

If you can really predict an individual couple’s future so accurately, imagine telling that couple that they are doomed. that they should pack up now and save themselves any more heartache! That’s the horrific bit.

Of course the reality is that these tests are nowhere near as impressive as the headline percentages would have you believe. With so many variables influencing a couple over the subsequent years, predicting outcomes with such huge accuracy is simply implausible.

We’ve been here before.

Back in the 1980s and 1990s, a lot of research was done on predicting who will stay and who will split using videos and trained observers. Don’t forget video technology only became cheaply available in the late 1970s. The well known marriage guru Professor John Gottman made his name using this method, making predictions and helping couples at his Love Lab in Seattle.

He and others filmed couples, coded the behaviours – whether positive and negative, how each partner reacted – and from this were able to predict with more than 80 per cent accuracy who would stay together. The longest version of this kind of analysis I’ve seen was a study that followed newlyweds for 13 years.

In fact the voice AI study quoted above used the videos from one of these studies, because it saved them the huge expense of having to do a new study and the time needed to find out what happened to the couples later on.

But don’t get too excited just yet.

Although predicting with 79 per cent accuracy sounds impressive, it’s really not.

In this particular study, the couples had been married for an average of 10 years and were assessed as ‘chronically distressed’. In a similar study conducted in the US a few years earlier, 9 per cent of couples had split within 2 years and 38 per cent within 4 years.

So the baseline prediction – the minimum you can predict without knowing anything else about these couples – should be 91 per cent accuracy. You can do that just be saying all the couples will stay together. You’d be wrong in 9 per cent of cases but right for 91 per cent. Not bad.

Even if you said all will stay together over a four year period, you’d be wrong in 38 per cent of cases and right for 62 per cent.

Suddenly 79 per cent accuracy isn’t looking so hot.

The really good bit about this kind of objectively-coded observational research is that it has allowed researchers to focus on the relational processes that helps couples best. Being aware, for example, that demand-withdraw behaviour is destructive means that you can teach couples how to identify this kind of negative behaviour and nip it in the bud.

The best known and best researched relationship education programme of this kind is PREP, which is widely used in the US. I used a variant of this in the UK myself with seven hundred couples.

But we should treat with a large pinch of salt claims that you can predict with great accuracy how couples will turn out.

Thankfully it’s nonsense.

Harry Benson, Research Director