Would you like to know when you’re going to die?
Opinion A new AI predicts when we’ll die. It says even more about how we live.
January 15, 2024 at 7:30 a.m. EST
Communal jump scares are irresistible, and research that cast AI as the oracle of death was the perfect way to end 2023. The academics knew it, too. “We thought it sounds kind of “Minority Report,” so maybe it’ll get a little bit of attention for the paper,” says Sune Lehmann, one of the authors and a professor of networks and complexity science at the Technical University of Denmark. “Clearly that was a miscalculation, because it got this insane amount of attention! But really it’s not what the paper is about.”
It’s true. The marketing was a success, but it did the scholarship a disservice. Because “Using Sequences of Life-Events to Predict Human Lives” isn’t just a gimmick. It’s a fascinating merger of AI and social science that raises way more uncomfortable questions about how we live than about when we’ll die.
The project started from a basic observation about artificial intelligence. In large language models, words are transformed into numeric tokens. Then the tokens are examined in impossibly large combinations to find their relationships. Humans use grammar and logic to sort words, but the sheer size of a neural network allows computers to find patterns and correlations among the tokens that we could never spot.
The researchers wondered what would happen if, instead of text, they tokenized reams of data about everyday life. To exclude recent events such as the pandemic and ensure they could determine, ahem, final outcomes, they relied on data recorded from 2008 to 2015. “If you think about the social sciences, a lot of what they try to process are high-dimensional event sequences,” Lehmann, 47, told me over Zoom from his home in Copenhagen. (Yes, there was a very fast-looking bike in the room, and although he had a mild case of covid-19, he could not have been cheerier.) “So what if we use AI and think of each life as words in a sentence? You’re born, you see the pediatrician, you go to school, you get a job, and so on. What are the correlations that we can learn?”
To contemplate this kind of project, you need data from people willing to reveal much more about themselves than you’d get from an ordinary census. You also need that data to be as obsessively organized as a chef’s pantry. You need the Danes. “In Denmark, we have this wild thing — we trust the government! We like them,” Lehmann says.
That trust is the foundation of Statistics Denmark, a well-funded state agency established in 1966 to keep track of everything that might happen to a Dane. The data is available to any researcher affiliated with a Danish university, but it must be anonymized; a single lapse can cause the whole institution to lose access. There are highly specific categories for medical conditions, schooling, income, personality nuances and … duck strikes. “It’s a category for, like, if you’re moving along the landscape, and then you’re struck by a duck,” Lehmann says. “That’s how granular it is.”
When all the data was first converted to tokens, the model saw only an arbitrary string of figures. But over time, it discerned patterns. With nudges from the professors, it trained itself to recognize that income is different from concepts such as health and education. Then it made distinctions within each concept, sorting broken bones and various types of cancer by lethality. It taught itself how to put individual lives into the 100 salary quantiles used by Statistics Denmark. The end result is an orderly map of all the seemingly messy and random stuff that happens in a modern Danish person’s life.
The entire project was conducted in compliance with the European Union’s digital privacy laws, which are far clearer and stricter than the United States’ patchwork shrug of a digital privacy regime. That’s just one of the reasons it’s hard to imagine anything like this project happening here. (Even though one of the paper’s eight co-authors is American.) There’s simply not enough trust in government to gather data, keep it anonymous and execute on the findings — and that’s going to put the United States at a huge disadvantage in the near future. Denmark and other countries that combine AI with rich data sets will be able to play Moneyball at a national level. By finding unique correlations between life events, they’ll manage budgets and direct social services for their citizens with incredible efficiency. Nerdtopia is coming.
There’s a knee-jerk American argument that having an ineffective government is a small price to pay for keeping it away from your secrets. I’d give that more credence if Americans weren’t so promiscuous in sharing those secrets with Google, Meta, Amazon, the Container Store (don’t judge) and Peloton, or whatever. If we don’t start entrusting government institutions with the basic means to be effective, there’s a real possibility we won’t have any government institutions left. There are some people who’d like nothing better. But give me fiscal hygge or give me death.
Now, about death. It’s true that by excluding people under age 35 (“They rarely die,” Lehmann says) and over 65 (“They are always dying”) from the model, the researchers were able to predict mortality within the data with spooky accuracy. Except it’s not that spooky. If you want the one-word explanation: cancer. It’s not much of a guessing game if the data tells you someone’s body is under attack. There are, of course, deaths by misadventure, accident and chronic disease, but an actuary can produce similar (albeit less precise) results as the model without the benefit of AI.
Still, the research packs an existential punch. When Lehmann and his colleagues saw how AI mapped everything that people do, they realized: People don’t do very much. “Even if you live like a rock star, we’re pretty predictable,” Lehmann says. “Think about our mobility patterns. Most people wake up at home, have breakfast, go to work and go shopping, pick up the kids and come back home. We are all kind of doing the same thing. Our lives are much more boring than they appear.”
Denmark is rich and homogenous. Maybe there’s another country where every day is an adventure. But the conclusion that we’re not terribly distinct from one another — delivered by a technology that’s proving we’re not that distinct from machines — is kind of a bummer for Team Human. For decades, one of the chief arguments against AI ever working was that humans are too special, too magical to be reduced to patterns and probabilities. But the evidence keeps mounting. What we do is replicable because we are always repeating ourselves.
That doesn’t mean there’s no magic left to be made, just that AI might have called our bluff. If it can increasingly tell us the future outcomes of our choices, will we make the same ones? Or have the courage to change directions and seek new territory? Hamlet isn’t the relevant source here. It’s Ophelia: “We know what we are, but not what we may be.”