Here’s a use of AI that appears to do more good than harm. A pair of astronomers at the European Space Agency (ESA) developed a neural network that searches through space images for anomalies. The results were far beyond what human experts could have done. In two and a half days, it sifted through nearly 100 million image cutouts, discovering 1,400 anomalous objects.

The creators of the AI model, David O’Ryan and Pablo Gómez, call it AnomalyMatch. The pair trained it on (and applied it to) the Hubble Legacy Archive, which houses tens of thousands of datasets from Hubble’s 35-year history. “While trained scientists excel at spotting cosmic anomalies, there’s simply too much Hubble data for experts to sort through at the necessary level of fine detail by hand,” the ESA wrote in its press release.

After less than three days of scanning, AnomalyMatch returned a list of likely anomalies. It still requires human eyes at the end: Gómez and O’Ryan reviewed the candidates to confirm which were truly abnormal. Among the 1,400 anomalous objects the pair confirmed, more than 800 were previously undocumented.

Most of the results showed galaxies merging or interacting, which can lead to odd shapes or long tails of stars and gas. Others were gravitational lenses. (That’s where the gravity of a foreground galaxy bends spacetime so that the light from a background galaxy is warped into a circle or arc.) Other discoveries included planet-forming disks viewed edge-on, galaxies with huge clumps of stars and jellyfish galaxies. Adding a bit of mystery, there were even “several dozen objects that defied classification altogether.”

“This is a fantastic use of AI to maximize the scientific output of the Hubble archive,” Gómez is quoted as saying in the ESA’s announcement. “Finding so many anomalous objects in Hubble data, where you might expect many to have already been found, is a great result. It also shows how useful this tool will be for other large datasets.”