It seems like an age ago that Large Language Models such as ChatGPT or Gemini entered our lives with the promise of changing them. In reality, it has been just over three years. However, artificial intelligence has been used in research for decades whenever large amounts of data need to be analyzed. One example is tracking asteroids: a crucial step in defending ourselves from catastrophes coming from space.
It was February 2013 when a small asteroid, about the size of a four-story building, exploded in the sky over the Russian city of Chelyabinsk. There were no direct impacts and no deaths from the incident, but the shockwave from the explosion shattered countless windows, causing more than 1,500 injuries.
This is the most recent asteroid-related incident, and it reminded us of something fundamental: the danger posed by asteroids is not just material for Hollywood blockbusters—it is real, concrete, and it can hurt.
The Chelyabinsk event gave a strong push to space agencies around the world to work on planetary defense, the set of techniques designed to protect us from asteroids. It is no coincidence that in September 2022 NASA struck the asteroid Dimorphos with the DART spacecraft. It was what is known as a kinetic impactor: a probe that, like a projectile, smashed into the asteroid to test whether we would be able to deflect it from its trajectory if we were to discover one on a collision course with Earth. The test result was positive, and now we have a possible weapon to defend ourselves from asteroids.
A key element of planetary defense, however, is monitoring: if we do not know which asteroids exist, where they are, and what their future trajectories are, how can we know if and when we need to defend ourselves? The problem is the sheer volume of data. Today we know of almost two million asteroids, and we discover new ones every day. Only a small portion of these are classified as Near-Earth Asteroids (NEAs), and only a fraction of those—fewer than 3,000 objects—are considered potentially hazardous (PHAs, Potentially Hazardous Asteroids). Keeping track of all these objects and their characteristics is a titanic task.
Just as DART was about to impact Dimorphos, the American company OpenAI was preparing to launch its Large Language Model, ChatGPT, which would come out a couple of months later. For many people, AI was born at that moment, but in reality only its more popular use was born then.
Alan Turing laid its foundations in 1950 with an article titled Computing Machinery and Intelligence, in which he proposed the famous Turing Test to determine whether a machine can be considered intelligent. A few years later, in 1959, the concept of machine learning was introduced—of which LLMs are only a very specific case.
In broad terms, Machine Learning algorithms are trained on samples of data in order to develop strategies that allow them to act even when they are fed different kinds of information. In other words, put simply, they learn. In the case of LLMs, these algorithms are trained on vast amounts of text so that they can then process the prompts they receive and generate coherent responses.
In asteroid research, Machine Learning has long been used for various purposes. For example, to classify asteroids by comparing their orbits to understand which ones are related to each other. Or to find new asteroids within photographs. This latter field became a necessity when, at the beginning of the 2000s, asteroid hunting became a job for robots.
In the past, “discovering” an asteroid mostly relied on the human eye, comparing photographs of the same region of the sky taken at different times. But when known asteroids grew into the hundreds of thousands and then millions, carrying out this kind of observational work became impossible. Algorithms were therefore developed to automate the process, flagging possible moving points of light within databases containing millions of photographs. One example among many is LINEAR (Lincoln Near-Earth Asteroid Research), which at the beginning of the millennium discovered more than 140,000 new objects.
These algorithms have gradually become more sophisticated and specialized (today they fall under the name Deep Learning) and are a cornerstone of today’s planetary defense and asteroid research.
Only with similar algorithms will it be possible, for example, to analyze the immense quantities of data produced by the Vera Rubin Observatory, active since the end of 2025, which for the next ten years will generate up to 20 terabytes of images and sky data every single day, mapping large portions of the sky countless times. It is estimated that thanks to its data six million new asteroids will be discovered throughout the Solar System. Managing and analyzing them will, inevitably, be a job for robots.
Luca Nardi