In a laboratory in Denmark, researchers have taught a machine to read a human life the way a scholar reads a text — and to predict, with unsettling precision, when that life might end. The algorithm, life2vec, achieved 78% accuracy forecasting four-year mortality among millions of citizens, not by magic but by pattern, drawing on the quiet data trails we leave behind in registries of income, illness, and occupation. What the researchers built is less a crystal ball than a mirror, reflecting back the social determinants of survival that epidemiologists have long suspected but rarely quantified
Danish AI Model Predicts Mortality With 78% Accuracy, Raising Ethical Questions
You can't run insurance against risk if everybody knows exactly what the risks are.
Mark
Why did the researchers choose people between 35 and 65 for the mortality test?
Mimi
Because that's the age range where death is hardest to predict. Younger people rarely die; older people die more predictably. The middle years are where the algorithm had to actually work.
Mark
The algorithm learned from sentences about people's lives. How does that actually work?
Mimi
It's treating your life like a story. "You worked as a guard in 2012 and earned this much." "You attended this school." The algorithm finds patterns in thousands of these stories and learns what sequences of events tend to lead to early death.
Mark
78% accuracy sounds impressive, but what does it actually mean?
Mimi
Out of 100 people the algorithm predicted would die, 78 actually did. But it also means 22 people it said would die didn't. And there are people it predicted would survive who didn't. The margin matters when you're talking about someone's life.
Mark
Why would insurance companies care about this?
Mimi
Because right now they price insurance based on broad categories—age, gender, smoking status. If they could know with 78% accuracy who will die soon, they could either charge those people much more or refuse to insure them entirely. That's the nightmare scenario.
Mark
The researchers said the algorithm can't predict car accidents. So what's it actually predicting?
Mimi
Mostly the slow things. Disease progression, the cumulative effect of depression, the wear of a difficult job. The things that show up in medical records and income statements. Not the sudden, random things.
Mark
Does this algorithm work in America?
Mimi
Nobody knows. Denmark has universal healthcare, stable employment records, and a homogeneous dataset. America has gaps everywhere—uninsured people, undocumented residents, people who avoid hospitals. The algorithm might not even see half the population.
Mark
What's the real danger here?
Mimi
Not that the algorithm is wrong. That it's right. If it works, and companies use it, then your entire future becomes legible to machines. Your choices, your health, your lifespan—all predictable. And once something is predictable, it can be controlled.
El Pulso
- A Danish algorithm has crossed a threshold that once seemed distant: predicting individual human death within a four-year window with 78% accuracy, outpacing all existing models by 11 percentage points.
- The findings expose uncomfortable hierarchies — mental illness, male gender, and skilled labor correlate with earlier death, while managerial rank and high income tilt the odds toward survival.
- The model's architects are sounding their own alarm, flagging blind spots in the data that erase the undocumented, the homeless, and anyone outside the formal healthcare system.
- Medical ethicists warn that insurance companies are watching closely, and that commercially deployed versions of such tools could price vulnerable people out of coverage within five years.
- Beyond discrimination, a subtler fear is taking shape: that lives rendered fully predictable by algorithmic scrutiny may lose something essential — the open horizon that gives human choice its meaning.
In a laboratory in Denmark, researchers have taught a machine to read a human life the way a scholar reads a text — and to predict, with unsettling precision, when that life might end. The algorithm, life2vec, achieved 78% accuracy forecasting four-year mortality among millions of citizens, not by magic but by pattern, drawing on the quiet data trails we leave behind in registries of income, illness, and occupation. What the researchers built is less a crystal ball than a mirror, reflecting back the social determinants of survival that epidemiologists have long suspected but rarely quantified so sharply. The deeper question it raises is not whether the machine is right, but who will be allowed to ask it questions — and what they will do with the answers.
A team at the Technical University of Denmark has built a machine-learning model called life2vec that predicts whether a person will die within four years with 78% accuracy — a result that outperforms existing models by at least 11 percentage points. Led by researcher Sune Lehmann, the team trained the algorithm on national registry data covering 6 million Danes, spanning education, health, income, and occupation across eight years. Their method was unusual: they borrowed language-processing techniques from artificial intelligence, converting life events into readable sequences — a salary received, a class attended, a diagnosis recorded — so the algorithm could learn from a person's history the way a reader learns from a story.
To test mortality prediction specifically, the researchers focused on 2.3 million people between 35 and 65, an age range where death is harder to forecast than in older populations. Life2vec was asked to predict survival among 100,000 individuals over a four-year window, and it was correct nearly four out of five times. The patterns it identified were familiar to epidemiologists but newly precise: men, skilled laborers, and those with mental health diagnoses were more likely to fall on the wrong side of the prediction; high earners and managers were more likely to survive.
Lehmann is candid about the model's limits. The data only captures people visible to Denmark's healthcare and tax systems, leaving out undocumented residents and those who avoid institutions. The findings may also not translate beyond Denmark's wealthy, universally insured population. But the limitations of the prototype are almost beside the point — what matters is the trajectory.
Arthur Caplan, a medical ethicist at NYU, frames the commercial risk plainly: insurers have always sought to quantify individual risk, and a tool this precise would be irresistible to them. Premiums could be set — or coverage denied — based on algorithmic verdicts. More advanced versions, Caplan predicts, could arrive within five years. His deeper concern, though, is philosophical. When enough of a life becomes predictable — its choices, its outcomes, its likely end — something of its value quietly disappears. The question regulators now face is whether they can move faster than the technology that is already outpacing them.
A team of Danish researchers has built an algorithm that can predict whether a person will die within the next four years with 78% accuracy. The model, called life2vec, represents a significant leap in what machine learning can do with personal data—and it has already started conversations about what should happen if such tools ever leave the laboratory.
The researchers, led by Sune Lehmann at the Technical University of Denmark, trained their algorithm on information from Denmark's national registry covering 6 million people. The dataset spanned 2008 to 2016 and included details about education, health, income, and occupation. To make sense of this information, the team adapted language-processing techniques typically used to understand text, converting life events into sentences the algorithm could learn from. One example: "In September 2012, Francisco received twenty thousand Danish kroner as a guard at a castle in Elsinore." Another: "During her third year at secondary boarding school, Hermione followed five elective classes." By treating a person's life as a sequence of events, life2vec learned patterns that humans might miss.
To test the algorithm's ability to predict mortality, the researchers focused on a specific group: over 2.3 million people between 35 and 65 years old, tracked from 2008 through 2015. This age range was chosen because deaths in this bracket are harder to forecast than in older populations. The team then asked life2vec to predict who among 100,000 individuals would survive the four years following 2016. The algorithm had no knowledge of who actually lived or died; the researchers did. When the results came back, life2vec was correct 78% of the time—and it outperformed existing prediction models by at least 11 percentage points.
The patterns the algorithm identified align with what epidemiologists already know, but the precision is striking. Men were more likely to be predicted as not surviving. Skilled workers like engineers faced higher predicted mortality than those in managerial roles. A diagnosis of depression or anxiety pushed someone toward the "death" column. High income and managerial positions pushed toward survival. These correlations are not new, but life2vec's ability to weigh them together and make individual predictions is.
Lehmann is careful to note that life2vec remains a research prototype. The study had significant limitations. The experiments were not randomized, and researchers knew the outcomes they were testing against. The algorithm only saw eight years of data. Perhaps most importantly, it only captured information about people who engaged with Denmark's healthcare system and earned reported income. Anyone outside that net—undocumented residents, the homeless, those who avoid hospitals—simply does not exist in the data. The researchers also acknowledge that their findings may not travel well. Denmark is wealthy, with robust infrastructure and universal healthcare. Whether life2vec's predictions would hold in the United States or other countries remains unknown.
Yet the real conversation is not about the algorithm's current limitations. It is about what happens next. Arthur Caplan, head of medical ethics at New York University's Grossman School of Medicine, sees the commercial future clearly. Insurance companies have long wanted to predict risk with precision. If life2vec or something like it becomes commercially available, insurers will use it. "You can't run insurance against risk if everybody knows exactly what the risks are," Caplan said. The math is simple: if an algorithm can tell an insurer that you are likely to die soon, they will price your premiums accordingly—or decline to insure you at all. Caplan expects more advanced models within five years, built on larger databases, offering not just predictions but recommendations for prolonging life.
What troubles Caplan most is not the accuracy of the prediction but what prediction removes from human existence. An algorithm cannot tell you whether you will die in a car accident or from an undiagnosed illness. But if enough aspects of your life become predictable—your behavior, your choices, your likely outcomes—then the mystery that makes living interesting begins to evaporate. "We're worried about the robots taking over the world," Caplan said. "What we need to worry about is the robots manipulating information and being able to predict a lot about our behavior that we wind up having lives that are so predictable that they take some of the value out of living." The question now is whether regulators will move faster than the technology.
Citas Notables
We can make any kind of prediction with highly specific data about a person.— Sune Lehmann, lead researcher, Technical University of Denmark
If someone does not have a salary or chooses not to engage with the healthcare system, we do not have access to their data.— The research team, on limitations of the Danish dataset
Quieres la nota completa?
Lee el original en Cable News Network ↗
Technology