The 43 Unclassifiable Objects: What AI Found Hiding in 35 Years of Hubble Data That Astronomers Cannot Explain

For 35 years, the Hubble Space Telescope has been staring at the universe and sending back images. Astronomers have analyzed those images by hand, searching for objects that fit into known categories: galaxies, nebulae, gravitational lenses, supernovae. They built catalogs, published papers, and moved on to the next observation. In January 2026, an AI built by two ESA researchers scanned the entire Hubble archive in 2.5 days and found what humans had been walking past for decades.

The numbers are striking. Out of 99.6 million image cutouts, the algorithm flagged 1,339 anomalous objects. Of those, 811 had never been documented in any scientific publication. Most could be retroactively classified once identified: galaxy mergers caught mid-collision, gravitational lenses bending light around unseen mass, jellyfish galaxies trailing plasma as they move through dense cluster gas.

But 43 of them could not be classified at all. They do not match any known category in astrophysics. Their morphologies have no precedent, their physical processes map to no established mechanism. The researchers described them as objects of "uncertain astrophysical nature" that "warrant community discussion." That is scientific language for: we do not know what these are.

This article is not about 43 weird blobs in old telescope data. It is about what happens when you combine AI pattern recognition with the growing number of cracks in our fundamental physics models, from dark energy that appears to be weakening, to early galaxies that should not exist, to particle decays that violate the Standard Model. Something is wrong with our understanding of the universe, and the evidence is accumulating faster than our frameworks can absorb it.

The AnomalyMatch Discovery

The tool is called AnomalyMatch, developed by David O'Ryan and Pablo Gomez at the European Space Agency's European Space Astronomy Centre in Madrid. Their paper, published in Astronomy and Astrophysics in December 2025, describes a deceptively simple approach to a problem that had never been attempted at this scale.

The Hubble Legacy Archive contains 35 years of observations, roughly 1.7 million individual pointings of the telescope. From these, the team extracted 99.6 million small image cutouts, each covering 7 to 8 arcseconds of sky. The question they asked was not "what are these objects" but rather "which of these objects do not look like the others."

AnomalyMatch uses semi-supervised learning built on an EfficientNet neural network backbone. The critical detail is how little labeled data it needed to begin. The algorithm started with exactly three known anomalies, 128 normal objects, and 99.6 million unlabeled images. Through three rounds of active learning, where researchers verified the AI's highest-confidence candidates and corrected its mistakes, the system expanded its training set to 1,400 labeled samples and achieved 76 to 94 percent precision in identifying genuine anomalies.

The entire scan took 2.5 days on a single GPU. Training took less than four hours. For context, a human astronomer examining one image per second, working eight hours a day, would need approximately 34 years to look at every cutout once. The Hubble archive has existed for 35 years. An AI accomplished in a long weekend what a human career could not.

What the AI Found

The 1,339 anomalous objects break down into categories that tell their own story about what was hiding in plain sight.

The largest group, 629 galaxy mergers and interactions, represents collisions between galaxies caught at various stages. These are not mysterious in principle. Astronomers understand galaxy mergers well. What is remarkable is that 65 percent of these had never been cataloged. They were sitting in the archive, visible to anyone who looked, and no one had looked.

The 140 gravitational lens candidates are scientifically valuable. Each one represents a massive object, often a galaxy cluster, bending spacetime enough to distort the image of something behind it. Finding new gravitational lenses is like finding new natural telescopes. The 95 high-redshift galaxies are among the most distant objects ever observed by Hubble, each one a window into the early universe.

Then there are the 43.

The Unclassifiable 43

The paper describes these objects carefully. Some display curved morphologies with no matching type in existing astrophysical literature. Others show signatures that resemble ram-pressure stripping, the process where gas is torn from a galaxy as it moves through a dense medium, but they are not located in galaxy cluster environments where ram-pressure stripping occurs. The researchers explicitly noted that some of these objects "may not be galaxies but rather other objects" entirely.

This is a significant statement. Astrophysics has developed a comprehensive taxonomy over the past century. Stars, galaxies, nebulae, quasars, pulsars, magnetars, black holes, and dozens of subcategories within each class. When professional astronomers say an object does not fit any known category, they are saying something that carries weight. These are not ambiguous readings on a noisy detector. They are clearly imaged objects in 35 years of the most productive space telescope ever built, and they match nothing in the catalog.

The conventional response is reasonable: these are probably exotic configurations of known phenomena. A multi-body galaxy interaction seen from an unusual angle. An object caught in an extremely brief transitional phase between recognized states. The human verification step in the AnomalyMatch pipeline was specifically designed to filter out imaging artifacts and optical contamination, so simple instrumental explanations are unlikely.

But here is the question that nags: if 811 previously undocumented anomalies were hiding in the archive, including 43 that cannot be classified, what exists in datasets that have never been systematically scanned? Hubble is one telescope. Its archive, while vast, covers a fraction of the sky. The Vera C. Rubin Observatory will generate 20 terabytes of imaging data every night for ten years. ESA's Euclid mission is surveying one-third of the sky. If AI anomaly detection scales to these datasets, the number of unclassifiable objects will not be 43. It will be orders of magnitude larger.

The Observation Bias Problem

The AnomalyMatch discovery exposes something uncomfortable about how science has operated for centuries. Astronomers look for what they expect to find. Telescope time is allocated through competitive proposals that describe specific scientific objectives. Researchers study objects they already have frameworks to understand. Careers are built by publishing results that fit within existing paradigms. There is no incentive structure for spending years staring at blobs that defy classification.

This is not a conspiracy. It is a structural property of how scientific institutions function. Funding agencies want measurable outcomes. Peer reviewers evaluate papers based on existing theoretical frameworks. Journal editors prioritize results that advance known fields. An astronomer who submits a paper saying "I found 43 objects and I have no idea what they are" faces a different reception than one who submits "I found 140 new gravitational lenses that constrain dark matter halo profiles."

The result is a systematic blind spot. Objects that fit known categories get cataloged, studied, and published. Objects that do not fit get ignored, deferred, or explained away as artifacts. The Hubble archive accumulated these orphan anomalies for 35 years because no human had the bandwidth or the institutional incentive to look for things that broke the taxonomy.

AI does not have career incentives. It does not need tenure. It does not submit grant proposals. AnomalyMatch treated every image with the same computational attention regardless of whether the object inside it matched a known category. The 43 unclassifiable objects were not hidden. They were visible in the data the entire time. They were simply never the object of anyone's search.

The Convergence of Cracks

The Hubble anomalies arrive at a moment when the foundations of modern physics and cosmology are under simultaneous pressure from multiple directions. Taken individually, each anomaly has conventional explanations or at least conventional caveats. Taken together, they form a pattern that is harder to dismiss.

Dark Energy Is Weakening

The Dark Energy Spectroscopic Instrument released its second data set in March 2025, analyzing over 14 million galaxies and quasars. The results suggest dark energy, the force accelerating the expansion of the universe and constituting 68 percent of its total energy content, may not be the cosmological constant Einstein proposed. Instead, it appears to be weakening over time.

The statistical significance ranges from 2.8 to 4.2 sigma depending on which supernova dataset is combined with the DESI measurements. This is below the 5-sigma discovery threshold but well above noise. Cornell physicist Henry Tye published an analysis arguing the cosmological constant may actually be negative, which would mean the universe will expand for roughly 11 billion more years before contracting into a "Big Crunch."

If dark energy is evolving, the Lambda-CDM model that has served as the standard framework for cosmology since the late 1990s requires fundamental revision. This is not a minor adjustment. It is the equivalent of discovering that the gravitational constant changes over time.

The Universe Expands at Two Speeds

The Hubble tension has escalated from a measurement discrepancy to what researchers at the January 2025 American Astronomical Society meeting formally called a "crisis." Local measurements of the universe's expansion rate yield approximately 73 kilometers per second per megaparsec. Early-universe measurements from the cosmic microwave background yield approximately 67. The difference, confirmed at 5-sigma confidence, means either our measurements of the nearby universe are systematically wrong, our understanding of the early universe is incomplete, or something fundamental about cosmology needs to change.

In 2025, new measurements did not resolve the tension. They deepened it. One JWST-calibrated result yielded 70.4, tantalizingly close to splitting the difference. But the TDCOSMO collaboration's gravitational lensing measurement yielded 72.1, siding firmly with the local value. The universe appears to be expanding at two different speeds depending on when you measure it.

Galaxies That Should Not Exist

The James Webb Space Telescope continues to find galaxies in the early universe that are too massive, too bright, and too chemically enriched for their age. JADES-GS-z14-0, the most distant galaxy ever confirmed at redshift 14.32, contains an estimated half a billion solar masses despite existing only 300 million years after the Big Bang. Standard models of galaxy formation cannot produce objects this massive in the time available.

Some "impossible" galaxies have been partially explained by improved calibration and a newly identified class of objects called "little red dots" that are likely accreting black holes rather than conventional galaxies. But even after corrections, roughly twice as many massive early galaxies exist as the standard model predicts. A January 2026 study proposed "dark stars," hypothetical objects powered by dark matter annihilation rather than nuclear fusion, as a solution. The fact that such exotic mechanisms are being seriously considered indicates the depth of the problem.

The Standard Model Springs Leaks

Particle physics faces its own accumulating anomalies. The NA62 experiment at CERN achieved a 5-sigma observation of an ultra-rare kaon decay, the rarest particle process ever measured. The observed rate is approximately 50 percent higher than the Standard Model predicts. While the excess itself sits at roughly 2 sigma due to experimental uncertainties, it points in the same direction as other hints of physics beyond the Standard Model.

Fermilab's muon g-2 experiment delivered its final result in June 2025, measuring the muon's magnetic moment to 127 parts per billion, surpassing its design precision. In a twist, the theoretical prediction shifted rather than the experimental measurement. Lattice QCD calculations produced a new Standard Model prediction that agrees with experiment, but this prediction is incompatible at 3 sigma with the previous data-driven theoretical calculation. The theory disagrees with itself.

The Pattern Recognition Paradox

There is a deeper question beneath the specific anomalies. The AnomalyMatch discovery demonstrates that AI pattern recognition can find things in existing data that human experts have missed for decades. But it also raises a paradox: if the AI identifies objects that do not fit any known category, how do we know they are real anomalies rather than artifacts of the algorithm itself?

The answer lies in the verification pipeline. Every high-confidence anomaly candidate was reviewed by human experts. Imaging artifacts, optical contamination, and detector noise were filtered at each stage. The 43 unclassifiable objects survived this process. They are real objects in real Hubble images that real astronomers examined and could not categorize.

But the paradox extends further. The AI was trained to distinguish "anomalous" from "normal" starting from just three labeled anomalies. Its definition of anomalous was shaped by what its human trainers showed it. If the three initial anomalies had been different, the algorithm might have found different objects in the same archive. We are not seeing what the universe contains. We are seeing what a particular AI, trained by particular humans, with particular initial assumptions, flags as unusual. The 43 unclassifiable objects are not a complete census of the unknown. They are one algorithm's interpretation of strangeness.

This does not diminish the discovery. It contextualizes it. The Hubble archive may contain thousands of anomalous objects that AnomalyMatch did not flag because they did not match its particular definition of anomalous. Different algorithms, trained with different seed data, will find different anomalies. The universe's true inventory of unexplained phenomena is almost certainly larger than any single AI scan can reveal.

The Mystery Object at 44 Minutes

The Hubble anomalies are not the only unclassifiable objects challenging astrophysics. In May 2025, NASA announced that ASKAP J1832-0911, a radio source 14,700 light-years away in the Milky Way, emits bursts lasting approximately two minutes every 44.2 minutes. It is the first long-period radio transient ever detected in X-rays.

This object is not a pulsar. Pulsars rotate in milliseconds to seconds, not 44 minutes. It is not a magnetar, though it may involve extreme magnetic fields. Researchers have proposed it could be a highly magnetic neutron star, a white dwarf, or a precessing black hole with a disk-jet system. None of these explanations account for all observed behavior. NASA's description was unusually candid: "unlike anything we have seen before."

The MeerKAT telescope has since discovered 26 additional galactic radio transients, several of which pulse at similarly anomalous periods. A class of objects that did not exist in the astronomical taxonomy two years ago is rapidly growing. Like the Hubble 43, these objects force the question: are we discovering new phenomena, or recognizing phenomena that were always there but invisible to our methods?

What Would It Mean

Consider the convergence. An AI finds 43 objects in Hubble data that match nothing in astrophysics. Dark energy appears to be changing over time. The universe expands at different speeds depending on when you look. Galaxies exist in the early universe that should be impossible. A kaon decays 50 percent more often than predicted. A radio source pulses at a period that matches no known stellar mechanism.

None of these anomalies individually requires new physics. Each one has at least a partial conventional explanation or sits below the 5-sigma threshold for claiming discovery. But the pattern is harder to dismiss than any individual data point.

If the Standard Model of particle physics and the Lambda-CDM model of cosmology are both correct, then these anomalies are statistical fluctuations and measurement artifacts that will resolve with better data. This is the default expectation, and historically, most anomalies do eventually find conventional explanations.

But if these anomalies are genuine cracks, they point toward physics we do not yet have. The universe may contain objects, forces, and processes that our current theoretical frameworks are structurally incapable of describing. The 43 unclassifiable Hubble objects would then be not anomalies but the first visible symptoms of a deeper reality waiting for the right mathematics.

The AI did not answer this question. It simply showed us what was there all along, patient and uncategorized, waiting in the data for someone, or something, to finally look.

Open Questions

1. What are the 43 unclassifiable objects? Follow-up observations with JWST or ground-based telescopes could determine whether these are genuinely new classes of astrophysical phenomena or exotic configurations of known processes.

2. What would other AI architectures find? AnomalyMatch started with three labeled anomalies. Different seed data and different neural network architectures would produce different anomaly catalogs from the same archive. A systematic comparison of multiple AI approaches on the same dataset could reveal whether the 43 represent a robust signal or an algorithmic artifact.

3. How many anomalies exist in unscanned archives? Hubble is one telescope. Ground-based survey data, radio telescope archives, and X-ray observatory catalogs have never been systematically scanned for unclassifiable objects. The total population of unknown phenomena in existing astronomical data is entirely unconstrained.

4. Are the cosmological anomalies connected? The Hubble tension, DESI dark energy results, JWST impossible galaxies, and particle physics anomalies all point toward potential cracks in standard models. Whether these cracks are independent measurement issues or symptoms of the same underlying physics remains an open and testable question.

5. What is the institutional cost of observation bias? If 811 anomalous objects were hiding in the most studied space telescope archive in history, how many discoveries are delayed or prevented by the structural incentives of academic science to study known phenomena rather than search for unknown ones?

The data exists. The AI has shown us it exists. What happens next depends on whether the scientific community treats 43 unclassifiable objects as a curiosity or a crisis.