An artificial intelligence analyzed 35 years of observations from the Hubble telescope and located more than 800 strange objects that had not previously been recorded in the scientific literature, revealing that even one of the most studied astronomical collections still contains visible surprises.
Two researchers from the European Space Agency applied an artificial intelligence tool to almost 100 million image clippings extracted from the Hubble Space Telescope archive and generated a selection of unusual objects, more than 800 of which had never been described in scientific publications.
Two researchers from the European Space Agency applied an artificial intelligence tool to almost 100 million image clippings extracted from the Hubble Space Telescope archive and obtained a selection of unusual objects, more than 800 of which had never been described in scientific literature. The tool, called AnomalyMatch, was created by David O’Ryan and Pablo Gómez, who published the study in the magazine Astronomy & Astrophysics in December 2025.
The numbers call for caution because initial disclosure may distort perception. The collection in question, known as the Hubble Legacy Archive, brings together data since the launch of the telescope in 1990, which means that the scan covered around 35 years of observations. According to the ESA/Hubble announcement, it was the first time that this file underwent a systematic search for anomalies of this type. The tool needed around two and a half days to process the clippings, each with a few dozen pixels and around seven to eight arc seconds in length.
What the tool actually did
AnomalyMatch did not make discoveries autonomously. He organized the images according to their degree of strangeness in relation to what he had learned during training and handed a prioritized list to the two astronomers, who visually examined the most promising candidates. Of the items identified by the program, the researchers validated more than 1,300 as visually anomalous, and the published catalog contains 1,255 unique objects distributed in 18 categories. More than 800 of them were not included in previous scientific literature.
The distinction is relevant. The step that transformed the machine-generated list into concrete objects still depended on human analysis of the images.
What artificial intelligence changed was the scale of the search. A complete manual scan of tens of thousands of Hubble data sets would not be practical for any team of people, which is why much of this material had never been examined with specific attention to anomalies. The tool made it viable to explore this volume of information. It did not eliminate the need for final human judgment.
What appeared
Most of the detected objects consist of galaxies in the process of merging or interacting, with irregular shapes or long streams of stars and gas. The catalog records more than 400 such cases, as well as 86 new candidates for gravitational lensing, in which the gravity of a foreground galaxy distorts the light from a background object into arcs or rings. Ring galaxies resulting from collisions, galaxies with the appearance of jellyfish and gas filaments, galaxies rich in large star-forming clusters and, closer, planetary disks observed in profile within the Milky Way also appeared.
A smaller set, with a few dozen objects, did not fit into any existing classification category. These cases are those that most justify additional investigations and also those that present the greatest risk of overestimation.
What “previously undocumented” does and doesn’t mean
Undocumented does not equate to unpublished. More than 800 objects missing from the literature indicate that no one had described them before, not that research revealed 800 new types of objects. Most of the categories mentioned merging galaxies, gravitational lensing, ring galaxies are already well known. What is new are the specific examples identified.
It is also worth clarifying the scope of the study. Objects were identified by visual appearance and confirmed to be anomalous in that respect, and the catalog treats unreferenced items as candidates rather than definitive conclusions. A gravitational lensing candidate still requires spectroscopic observation to confirm what is being lensed and how far away it is. The same caution applies to the few objects that have escaped known classifications. In the authors’ opinion, the result is a well-constructed catalog of candidates, and not a set of closed cases.
Why it matters and what to watch
The most relevant point is not so much the 800 objects themselves, but the method and context in which they were gathered. The Hubble archive is extensive but finite. Large ongoing surveys do not have this limit. ESA’s Euclid mission and the Vera C. Rubin Observatory will generate quantities of images that no team will be able to review manually, and the only practical way to locate rare objects will be to allow algorithms to prioritize candidates so that astronomers can check the most promising ones. The Hubble analysis serves as a proof of concept for this workflow on a dataset that is already partially known.
Therefore, the key is not the volume of finds, but whether lens candidates and unclassified objects will withstand follow-up observations and whether the same strategy will prove effective when applied to archives much larger and less explored than Hubble’s. The list of finalists only represents the beginning of the work, not the end.
