Excerpt
Fight Right
Why We Fight“What’s going on?” he says, as she says at the same time, “So what do you want to talk about?”
They both laugh.
The couple is sitting close together, propped up against crisp white pillows on their comfy bed, facing the camera. They’re angled toward each other; so far, they both seem warm and relaxed—maybe a little nervous about being filmed. We’ve asked them to turn on the camera of their laptop, begin recording, and then simply talk about how their day is going. That’s all.
Meanwhile, our AI system is watching them. This system was built to assist our Gottman-trained therapists and couples wanting to assess their relationship at home—to gather illuminating data about how partners respond to each other in casual interactions and in conflict. The AI can read their heart rate from the video feed, without any other devices. Using machine learning, it does emotion coding, pinpointing each partner second by second on a wide range of possible emotional categories. And it rates each person’s trust level in their partner on a scale of 0 to 100 percent.
The AI—designed by our brilliant colleagues, Rafael Lisitsa and Dr. Vladimir Brayman—gathers all this data as this couple chats briefly about their workweek, how they’re looking forward to the weekend and having the chance to relax. So far, the AI has coded their interaction as progressing from “neutral” to “interested.” Both are relaxed—heart rate is around 80. Trust metric is fairly high.
Then, she says, “Oh, by the way. I told my parents they could stay in our room this weekend when they come to visit. We’ll sleep on the couch.”
There’s a pause.
“You already told them?” he says.
“Well, yeah,” she says, a bit dismissively. “They’re my parents. I—”
“You know I don’t sleep well on the couch.”
“Oh, come on.” (Eye roll) “It’s just for the weekend. What’s the big deal?”
“Well, I want to be at my best with your parents. I don’t want to be grumpy because I didn’t—”
“Like you’re ever at your best with my parents anyway, so . . .”
“Wow.” His voice is laced with hurt and sarcasm. “Okay.”
“Why are you making that face? You know it’s true!”
“Hey, I’m trying to make an effort for your parents, and—”
“Oh yeah? Well, why has it taken three years for you to do that? Why is this the weekend?”
“Three years? You don’t think that I’ve made an effort for three years?”
From here the temperature of the interaction spikes rapidly. They interrupt, talk over each other. She accuses him of making her dad cry during a phone conversation recently; he tries to defend himself.
“You just had to slip in a snide little comment, didn’t you,” she says, “while I was just trying to say, ‘Happy Birthday’ to him.”
“I was just trying to be funny!” he shouts.
The AI has clocked both partners’ heart rates rising—his more significantly, to 107 beats per minute. The trust metric has plummeted; his goes critically low, to below 30 percent. Emotion rating nosedives for both. The interaction rapidly skews negative; she attacks, he defends, both speak to each other with contempt. Less than thirty seconds later, the couple turns away from each other, exhausted and angry, giving up on the conversation completely. As the video captured by the AI cuts off, they each stare off in opposite directions.
“Coding” Conflict
The couple above is a real couple who agreed to participate in a new platform we set out to build, designed to help other couples just like them: normal couples who had hit a tough patch (or a tough year . . . or a tough decade) and needed some support and guidance.
In recent years—and accelerated by the pandemic—the demand for skilled therapists has overwhelmed the existing network. Plus, for busy couples who are working full time and perhaps caring for young children or other family, getting in to see someone can be difficult. A lot of couples who really could use some professional guidance are going without for various reasons. We wanted to figure out a way for couples who are struggling to get immediate relief. And we saw that a lot of couples were struggling. So we went to work on a platform that couples could access through a phone, laptop, or tablet—a place they could go to find guidance and tools. And for this to work, we needed an AI that could observe couples interacting and, like the most skilled and experienced trained professional, identify the signs and signals of a conversation going into toxic territory. Therapists are trained to look for these signs: subtle cues in body language and physiology, vocal tone, language choices, and more. Could a computer be programmed to be this sensitive?
In short: yes. And when it comes to the coding of conflict, not only has the AI matched its human counterparts, it’s actually outperformed them.
Much of the data and observations about couples in conflict in this book comes from our decades of work in the Love Lab and from other important and groundbreaking observational studies by ourselves and other researchers. But now we are getting even more sophisticated and granular information from the AI we trained with John’s emotional coding system, called SPAFF, short for Specific Affect Coding System.1
When John was first beginning his research into couples, the field of psychology was struggling to nail down consistent patterns when it came to the personality and behavior of one individual, much less two. The general belief in the field was that studies of couples would be too unreliable to be scientifically useful. Studying a single individual was already so unreliable, the thinking went, that studying two would simply square that unreliability, making it exponentially worse. John, ever the mathematician, set out to prove that false.
He began searching for patterns of behavior in individuals and couples—specifically, sequences of interactions that could be indicative of the couple’s overall happiness and the success (or not) of their relationship.2 Across a series of observational research studies, he and his colleagues worked out a coding system that measured every possible nuance of an interaction between two people: facial expressions, tone of voice, language and rhetoric, physical cues, and more. He and his research partner, Robert Levenson, developed ways for couples participating in the studies to actually rate their own experiences during conflict conversations, offering even more essential data about how people experienced conflict and whether or not their intentions matched their impact. Following couples over time, they were able to track how these coded sequences of interactions between couples aligned with the outcome of their relationship: Did they break up? Did they stay together? If together, were they happy or miserable?
John studied all three groups, gathering data from couples who divorced, couples who stayed happily together, and couples who stayed unhappily together. That data was anything but unreliable.
John found that interactions between couples were incredibly stable over time and highly predictive of the future of their relationship. Using SPAFF to code a couple’s interactions, John was able to predict, with over 90 percent accuracy, the future of that couple’s relationship.3 And a huge part of that prediction was how these couples behaved in conflict.
One of the key pieces to John’s studies on relationships was the “conflict task,” where couples were asked to choose a topic of ongoing conflict and discuss it. Their fight would be taped, and a team of researchers would then pore over the footage, working to code every expression and interaction, down to the hundredth of a second. It’s demanding work, and our researchers needed to be highly trained in order to do this accurately. Before SPAFF, the other coding systems for looking at behavior and interactions were cue based—they looked at actions and expressions, elements of human behavior that you could note visually. The problem is, that left out a ton of essential context. What about vocal tone? A major key will suggest positive emotion, while a minor key indicates the opposite. What about emphasis on certain words versus others? We call that a paralinguistic cue: the same sentence, read with an emphasis on different words, could communicate either frustration or flexibility—you have to take that into account. And what about cultural differences in the use of language and physicality? Our coding system accepts that emotion is conveyed in an interactive way across all communication channels.
