Lee’s first app, called Crushh, promised exactly that. The “texting relationship analyzer” offered a romantic interest score on a scale of zero to five, as well as insights on the power dynamics in a conversation (i.e., who likes who more). It also prompted users to say a little about each repartee: How old were the people in the conversation, what were their genders? Was the contact a colleague? A spouse? A crush?
Lee says the app processed “hundreds of thousands” of these conversations, many of them self-labeled with those context clues. That provided a hefty data set of what real text conversations looked like, across various demographics and in different types of relationships. Some of the patterns were obvious—a person who says “I miss you” early in a conversation likely has the feels—but others were more Delphian. “Based on the data, people who have romantic intent use the words ‘night’ and ‘dream’ a lot more,” says Lee.
Other apps have used similar models to juice up sales pitches, advise employees on messaging the boss, or generate context-specific email replies. Boomerang, a plug-in for Gmail and Outlook, makes an AI tool that proofreads emails and suggests ways to improve them before you hit “Send.” An app called Keigo combines “advanced psychology” and “cutting-edge AI” to determine the personality of a person based on their emails or tweets, and then provides helpful suggestions on how to approach them.
Like any good assistant, Keigo can slide deftly into many situations: to prepare for the job interview, to win the second date, to better understand a partner after a big fight. But Teemu Huttunen, Keigo’s managing director, says people are mostly using it for love. “To be honest, we were hoping that people would use this in other forms than just dating, but the dating one is the most obvious,” he says. “When you have a Tinder match and you agree to go on a date, the next step is that you would have to say something interesting.”
The app borrows a model from IBM’s Watson, which performed a series of studies to map basic personality characteristics onto peoples’ output on social media. IBM’s version translates tweets into its own “big five” traits: openness, conscientiousness, extraversion, agreeableness, and neuroticism. Keigo uses a different framework, based on Meyers-Briggs’ personality assessments. Feed it a snippet of text and it’ll deliver recommendations on how to talk to someone.
By way of demonstration, Huttunen showed me a graph that had mapped my tweets against Oprah Winfrey’s. The insights suggested that Oprah and I are 77 percent “compatible,” and that in a conversation with her, I’d want to emphasize teamwork, my “future journey,” and intuitive reciprocity. (Later, Huttunen would send me an email that referenced our “inspiring” phone call, and I would wonder if Keigo had planted that choice of words.)
All of these apps require a real suspension of privacy—they are, after all, parsing intimate conversations. Lee says Mei anonymizes all of its conversational data, and allows users to scrub their uploads from the company’s servers. By way of caution, the app also displays this pop-up before you upload anything: “In order for Mei to give you analysis on your conversation, the conversation history needs to be uploaded to our servers. If you are not comfortable with this, PLEASE GO NO FURTHER.”
For the intrusion, Lee seems to think the payoff is enough. Right now, Mei is a novelty crush analyzer. But he likes to think about what might happen in the future, with a much bigger data set. “I could go, ‘OK, this is a crush, but what type? Are you just flirting? Are you married? You might be able to start building models for those things,” he says. “When you have enough data, it’s almost like an encyclopedia of people.”