An AI trawled through 35 years of Hubble images and found more than 800 strange objects that had never been documented before, showing that one of astronomy’s most famous archives is still hiding discoveries in plain sight

Two researchers at the European Space Agency have run an AI tool across nearly 100 million cropped images from the Hubble Space Telescope archive and come back with a shortlist of unusual objects, more than 800 of which had never been described in the scientific literature. The tool, called AnomalyMatch, was built by David O’Ryan and Pablo Gómez, who reported the work in the journal Astronomy & Astrophysics in December 2025.

The numbers are worth stating carefully, because the headline version flattens them. The archive in question, the Hubble Legacy Archive, holds data going back to the telescope’s launch in 1990, so the search covered about 35 years of observations. According to the ESA/Hubble announcement, this was the first time the archive had been systematically searched for anomalies of this kind. The tool took about two and a half days to work through the image cutouts, each only a few dozen pixels and roughly seven to eight arcseconds on a side.

What the tool actually did

AnomalyMatch did not discover anything on its own. It ranked images by how unusual they looked against what it had been trained on, then handed a shortlist back to the two astronomers, who inspected the highest-rated candidates by eye. Of the candidates it surfaced, the researchers confirmed more than 1,300 as visually anomalous, and the catalogue they released lists 1,255 unique objects across 18 classifications. More than 800 of those had not appeared in the published literature before.

The distinction matters. The work that turned a ranked list into a set of real objects was still done by people looking at pictures.

What the AI changed was the scale. A comprehensive manual pass through tens of thousands of Hubble datasets is not practical for a human team, and so most of this material had simply never been examined with anomalies in mind. The tool made the haystack searchable. It did not replace the judgement at the end.

What turned up

Most of the flagged objects were galaxies in the middle of merging or interacting, pulled into irregular shapes or trailing long streams of stars and gas. The catalogue lists more than 400 of these, alongside 86 new gravitational lens candidates, where the gravity of a foreground galaxy bends light from something behind it into arcs or rings. There were collisional ring galaxies, jellyfish galaxies with trailing filaments of gas, galaxies studded with large star-forming clumps, and, closer to home, edge-on planet-forming disks within our own galaxy.

A smaller group, several dozen objects, did not fit existing classification schemes at all. Those are the ones most likely to reward follow-up, and also the ones it is easiest to oversell.

What “previously undocumented” does and does not mean

Undocumented is not the same as unprecedented. More than 800 objects not appearing in the literature means no one had previously written them up, not that the search turned up 800 new kinds of thing. Most of the categories above, merging galaxies, lenses, ring galaxies, are already well understood as classes. What is new is these particular instances of them.

It is also worth being clear about what the paper establishes. These are objects flagged by appearance and confirmed as visually anomalous, and the released catalogue treats the unreferenced ones as candidates rather than settled cases. A gravitational lens candidate identified from its shape still needs spectroscopic follow-up before anyone can be confident about what is being lensed and at what distance. The same caution applies to the unclassified handful. In our reading, the result is a well-made finding list, not a set of closed cases.

Why it matters, and what to watch

The interesting part of this is less the 800 objects than the method, and the timing. Hubble’s archive is large but bounded. The surveys coming online now are not. ESA’s Euclid mission and the Vera C. Rubin Observatory will produce image volumes that no team can inspect by hand, and the only way to find rare objects in them will be to let software rank candidates first and have people check the top of the list. The Hubble run reads as a demonstration of that workflow on a dataset whose contents are at least partly known.

So the thing to watch is not the count. It is whether the lens candidates and the unclassified objects survive follow-up, and whether the same approach holds up when it is pointed at archives many times larger and far less explored than Hubble’s. The shortlist is the start of the work, not the end of it.