Scientists built a "death calculator" AI trained on 6M people

What if your entire existence could be structured like a sentence, with health diagnoses, jobs, and paychecks acting as the words? If a language model can guess the next word in a paragraph, can it predict the final chapter of your life?

Researchers from the Technical University of Denmark and Northeastern University built an AI system that does just that. Dubbed life2vec, the model was detailed in a paper published in Nature Computational Science, demonstrating the ability to anticipate individual life events, personality traits, and even mortality with striking accuracy.

Turning life histories into tokens

Instead of scraping billions of internet forum posts, life2vec was trained on an extraordinary dataset: the entire population of Denmark, spanning roughly 6 million people. Granted rare access by Statistics Denmark, the researchers tapped into exhaustive civil registry data detailing medical visits, education milestones, job titles, and earnings.

Sune Lehmann, a professor at DTU Compute and lead author of the study, explained that human life can effectively be modeled as a single, sprawling sequence of events. The team adapted the same transformer architecture that powers systems like ChatGPT to digest these chronologically ordered milestones.

By plotting millions of real-life sequences into a high-dimensional vector space, the system maps intricate ties between seemingly unrelated life markers, such as income shifts and hospital admissions. When visualizing mortality predictions, the model produced what Lehmann described as a long cylinder stretching from low to high risk. Those placed on the high-risk end had a significantly elevated rate of death within the tracked window, while individuals who died unexpectedly on the low-risk end mostly succumbed to unpredictable incidents like car crashes.

Beyond lifespan, the model proved capable of predicting how individuals would score on standard psychological questionnaires, specifically regarding extroversion.

An observatory, not a crystal ball

Despite the uncanny predictive performance beating state-of-the-art baselines, the creators are adamant that this tool must never be pointed at living individuals to forecast their personal fates.

Tina Eliassi-Rad, a computer science professor and AI ethics expert at Northeastern University, stressed that the system reflects correlations inside a very specific environment. Denmark boasts a distinct universal healthcare system, unique labor market rules, and its own cultural fabric. Running the same architecture on data from a country like the United States would yield vastly different conclusions.

Eliassi-Rad views the algorithm not as an oracle, but as an observatory of society—a diagnostic mirror showing how policies, regulations, and institutional biases shape life trajectories.

Lehmann also noted that major tech companies have likely been building similar predictive systems behind closed doors for years. By publishing their methodology in the open, the researchers hope to pull the curtain back on what these models can actually infer, forcing a public debate before proprietary versions dictate real-world outcomes.

Why it matters

We think the real breakthrough here is conceptual: showing that transformer models are not limited to text and can decode the complex grammar of human lives. Treating life events as sequential tokens opens up a powerful new lens for public health researchers and sociologists. But you will not be plugging your medical history into a web app anytime soon, nor should you; the data is strictly locked down, and the researchers are intentionally keeping it out of commercial hands.

The darker implication lies in who else might replicate this. Insurance underwriters, credit bureaus, and large employers would love nothing more than an algorithm that calculates a customer's biological clock or long-term earning potential. If private tech giants feed their massive surveillance troves into similar architectures, algorithmic discrimination could shift from targeting what you buy to deciding how you live.

Lehmann and his team made a smart move by framing this as a warning shot rather than a consumer product. As predictive AI moves from generating text to forecasting human outcomes, transparent academic work like life2vec draws a line in the sand, giving policymakers a rare glimpse at the machinery before commercial actors weaponize it.