Is the Life2vec algorithm really capable of predicting the date of your death?

An artificial intelligence that would be capable of predicting death with a 78.8% success rate? This is the news relayed by several media and Internet users on social networks after the publication of a Danish study on December 18, 2023. In the scientific journal Nature Computational Science, six researchers from the Technical University of Denmark revealed that they had looked into life expectancy using health data from six million Danes. They derived an algorithm, called Life2vec.

Thanks to the Danish Statistical Institute, these scientists had access to a lot of data: the history of medical appointments with diagnoses, income, education, employment, working time… Up-to-date information every day from 2008 to 2020. To study life expectancy, they focused on data from 2008 to 2015 per 100,000 people aged 35 to 65. “Life expectancy in this age category is difficult to predict”admits Sune Lehmann, one of the authors of the study, contacted by franceinfo.

A prediction in specific cases

In this sample of people, the researchers used 50,000 people to “train” the algorithm and make it assimilate the mechanisms of death. For example: a person with a poor lifestyle, who has lived in polluted places all their life, has a greater chance of dying in the next four years than a other. In this context, the algorithm had 78.8% success.

“We fed people’s data into the algorithm whether they were going to die or survive in the next four years, then we did it again and the algorithm had to guess.”

Sune Lehmann, researcher and author of the study

at franceinfo

But the algorithm developed by Danish researchers cannot under any circumstances succeed in predicting the exact age of death of a person. First, as the study was based on Danes aged 35 to 65, testing it on another population would create bias. “A 28-year-old Argentinian does not have the same life expectancy as a 52-year-old Dane”, states the researcher. Life expectancy differs depending on the year of birth, a person’s country of origin, etc. Finally, the researchers argue that the objective of their study is purely scientific. “Our goal is not to alarm people by giving them their presumed age of deathexplains Sune Lehmann. We just want to advance science.”

A method that breaks with a classic study

The objective has been achieved, judges Florian Bonnet, researcher at the National Institute of Demographic Studies (INED): “The size and precision of the database used represent a major advance in the study of demography.” For comparison, in France, researchers have access to data from 4% of the population: a considerable gap with the Danish database. The method used for the study is also innovative. “We calculate life expectancy based on a particular variable: gender, socio-professional category, geographic origin”, lists Florian Bonnet. But in the Danish study, all these variables were integrated simultaneously, while taking into account temporality. “The algorithm created thus represents a break with the classic statistical analysis of demography”explains the French researcher.

Researchers are aware that by pushing the limits of the predictive skills of artificial intelligence, their work can interest companies which might be tempted to see an economic interest in it. “Insurance companies already use this type of statistics, but this new model is much more reliable”notes Laura Tocmacov Venchiarutti, director of the ImpactIA foundation, in The Parisian. Aware of these ethical issues, the Danish researchers specify that they have taken steps to secure their data. They also see their work as a way to raise awareness among governments so that they better regulate the use of AI. “We warn that this type of prediction already exists behind the doors of Silicon Valley without any law to regulate them”notes Sune Lehmann.

Fraudulent sites that take advantage of it

Although the data was protected, this did not prevent some people from creating sites to deceive Internet users and make a financial profit from this study. By searching for “Life2vec” on Google, you can come across a site with a “chatbot” which offers to answer a few questions such as “Do you go to the doctor regularly? Do you drink alcohol? Do you smoke?” After a few seconds, an age is displayed, the presumed age of your death.

But this online tool turns out to be misleading. If the site takes the name Life2vec, “this artificial intelligence is not linked to our study, it was created to defraud Internet users”, assures Sune Lehmann. The site, initially free, quickly offers numerous paid options to refine the presumed age of your death. “Unfortunately, humans are curious and willing to pay for such information, even if it is false”deplores the researcher.