When AI Combs the Sky: What Does It Actually Discover?

The astronomer keeping watch at the eyepiece is no longer the whole picture. Today an eighth planet in a known star system passed across scientists' screens without their noticing, and was then seen by a programme combing the numbers with tireless patience.

A flood of images

On 23 June 2025 the Vera C. Rubin Observatory released its first public images. Its camera has 3,200 megapixels and is expected to produce about 20 terabytes of data every night, in a ten-year survey known as the Legacy Survey of Space and Time (LSST). The accompanying briefing material estimates that it will observe about 17 billion stars, 20 billion galaxies and millions of transient events such as supernovae. No human team could examine that volume in a lifetime, which is why machine learning has become a need, not a luxury.

How are planets seen without an eye?

Most planets beyond the solar system are not seen directly; they are inferred from a slight dimming of their star's light as they pass in front of it. The Kepler spacecraft built its work on measuring this dimming across an enormous number of stars. But not every dimming is a planet; it may be noise in the instrument or some other astrophysical effect.

In 2018 Christopher Shallue and Andrew Vanderburg published the results of a convolutional neural network trained to tell planetary signals from false alarms. It ranked likely signals above false ones in 98.8% of test cases, a ranking measure rather than classification accuracy in the usual sense. Applied to new signals in systems already known to hold several planets, it led to two being confirmed: one in the Kepler-80 system, and an eighth planet in Kepler-90, which now matches the Sun in its number of known planets. No new observation of the sky was involved; the signal lay in existing data that earlier analysis had not picked up.

The human in the loop, and the tool's limits

The algorithm does not announce a discovery on its own; it ranks candidates, and statistical validation and scientific review follow. A fine example of this complementarity is the Gravity Spy project at the LIGO gravitational-wave observatory. Brief bursts of noise in the detectors (glitches) can mimic a real signal. Volunteers on the Zooniverse platform sort images of them into known classes, the models are trained on their labels, and the volunteers are asked to alert researchers to images that may represent a new class. The aim is to learn the causes of the glitches and remove them from the data or the instruments.

Notably, it is the human who notices the new, while the machine learns from what people have sorted.

Some of these limits follow from the nature of the tool rather than from any particular study:

• A model trained on known patterns is good at finding their like, and may treat the strange as noise.
• Many deep models are hard to interpret, and science seeks understanding, not prediction alone.
• A polished result with precise numbers invites acceptance, so follow-up observation and human review are essential.

The most useful division of labour therefore remains: the machine scans and ranks, and the person asks, interprets and guards the anomalous.

The Saudi context

In late December 2018 a royal decree establishing the Saudi Space Agency was announced, according to The New Arab, which reported that it aims at satellite communications and at investing in the space industry to diversify the economy. Analysing astronomical data is a field that joins astronomy and data science, and it is open to anyone who masters both.

Beyond wonder

The question of the first human before the stars remains the same: what lies in the distant dark? Only the instrument has changed: the naked eye, then the lens, then the sensor, then the algorithm. In all of this there is a lesson for those who reflect on the greatness of creation, and on God's grace in letting human beings learn what they did not know, provided they verify before they accept.