An AI combed through 35 years of Hubble images and discovered more than 800 strange objects that had never been documented before.

Two researchers from the European Space Agency ran an AI tool on nearly 100 million cropped images from the Hubble Space Telescope archive and returned a short list of unusual objects, more than 800 of which had not been described in the scientific literature. The tool, called AnomalyMatch, was developed by David O’Ryan and Pablo Gomez, who reported the work in the December 2025 issue of the journal Astronomy and Astrophysics.

It’s worth noting the numbers carefully, as the title version flattens them. The archive in question, the Hubble Legacy Archive, contains data dating back to the telescope’s launch in 1990, so the search covers about 35 years of observations. According to the ESA/Hubble announcement, this is the first time the archive has systematically searched for anomalies of this type. The tool took about two and a half days to work through the image cutouts, each a few dozen pixels and roughly seven to eight arcseconds per page.

What the tool actually did

AnomalyMatch itself doesn’t find anything. It ranked the images by how unusual they looked against the trained ones, then handed a shortlist to two astronomers, who examined the highest-scoring candidates by eye. Of the candidates it came up with, the researchers confirmed that more than 1,300 were visually anomalous, and the catalog they published lists 1,255 unique objects in 18 classifications. More than 800 of them have not appeared in the published literature before.

The difference is important. The task of converting the ranked list into a collection of real objects was still done by the image viewer.

AI changed the scale. It would be impractical for a human team to sift through the tens of thousands of Hubble datasets in a comprehensive manual, so much of this material has never been examined with inconsistencies in mind. This tool made hay searchable. This ultimately did not change the verdict.

what happened

Most of the flagged objects are merging or interacting galaxies, pulled into irregular shapes, or trailing long streams of stars and gas. The catalog lists more than 400, along with 86 new gravitational lens candidates, where the gravity of a foreground galaxy bends light from something behind it into curves or rings. Closer to home were ring-ring galaxies, jellyfish galaxies with gas filaments, galaxies embedded in massive star-forming clusters, and planet-forming disks in our own galaxy.

A small group, several dozen objects, do not fit existing classification schemes. They are the ones that can reward follow-ups, and are the easiest to sell.

What “previously undocumented” does and does not mean

Undocumented is not unprecedented. If more than 800 objects do not appear in the literature, no one has written about them before, and the search turns up 800 new types of objects. Most of the above categories, including galaxies, lenses, and ring galaxies, are already well understood classes. What is new is that these specific events.

It is also worth being clear about the installation of the sheet. These are objects that are flagged by appearance and confirmed as visually contradictory, and the published list treats unspecified as candidates rather than resolved cases. An individual identified from the shape of a gravitational lens still requires spectroscopic follow-up before anyone can be confident about what lenses it is and at what distance. The same caveat applies to the unclassified handle. In our reading, the result is a well-prepared list of findings, not a collection of closed cases.

Why it matters and what to look for

The interesting part of this is that 800 objects are less than method and time. Hubble’s archive is large but limited. Now there are no surveys coming online. ESA’s Euclid mission and Vera c. Rubin’s observatories produce blocks of images that no team can examine by hand, and the only way to find rare objects in them is to have the software first rank the candidates and rank them at the top of the list. The Hubble run is read as an illustration of that workflow in a dataset whose contents are at least partly known.

So the thing to look at is not the number. Whether lens candidates and unclassified objects survive follow-up and whether the same approach will hold up if it points to archives many times larger than Hubble’s and far less studied. Shortlisting is the beginning of the work, not the end.