Artificial intelligence fortune telling predicts income and deaths

This may be the closest thing we have to a fortune-telling marble: an AI algorithm trained on the resumes of millions of people predicts parameters like income and risk of early death with chilling accuracy, a technology that promises benefits in preventive medicine but also raises new ethical questions.

The model is within 78% of its predictions, its Danish creators reported in the review Computational natural sciences. The system identified several factors that increase the risk of premature death, such as low income and psychiatric diagnosis. The cases he fell into often involved accidents or heart attacks, events that are difficult to predict.

The algorithm was trained on data from 2008 to 2016 and was asked to make predictions for 2020.

The study relied on “large language model” technology such as the famous ChatGPT, which is trained with large amounts of text and learns to predict which word is most likely to follow another word.

Sonny Lehmann and his colleagues at the Technical University of Denmark wanted to study whether these models could also detect patterns in sequences of things other than words. For example, the likelihood of surviving cancer may improve if health insurance is obtained earlier.

“As with language, the order in which life events occur is of great importance,” Lehman commented on the magazine’s website. Sciences.

The lives of others

The model, called life2vec, was trained using data from Denmark’s rich national registries, which contain medical and employment information for the entire population of 6 million. The researchers encoded the data so that the model could place events in the correct chronological order.

The algorithm was trained on data from 2008-2016 and was asked to make predictions for 2020, to guess for example whether someone would die in the meantime.

Researchers believe that such models will be useful in preventing disease and premature death.

However, the algorithm is unlikely to give reliable predictions for populations outside Denmark, and it is possible to reproduce hidden biases and stereotypes in the data.

This could have major implications if such algorithms are used to make decisions in insurance funds, warns Youyou Wu, a psychologist at University College London. For example, overdiagnosis of schizophrenia among blacks may lead to incorrect estimates of the risk of premature death.

The researchers also acknowledge that practical applications of these models face privacy issues that have not yet been resolved.

Who would be able to make such predictions? And what if someone does not want to know mantras for his future?