An AI trained on cat photos redrew the first black hole image, and the ring held

The most famous picture of a black hole has passed a blind test. A small team at CSIRO, Australia’s national science agency, built an imaging program that learned what pictures look like from photos of cats, dogs and cars, pointed it at the raw Event Horizon Telescope data behind the first image of a black hole, and got back the same glowing ring.

That matters because the ring had a challenger. A rival reanalysis argued that the doughnut seen around the giant black hole at the heart of a nearby galaxy might be an artefact of how the data were processed, and that the real object could be an elongated bright spot. The new work, published in Nature, finds the ring in every high-quality dataset it examined, and its size points to a black hole billions of times heavier than the Sun, in line with independent estimates.

“Importantly, our model had no idea what a black hole looked like when we applied it to the M87 data,” said lead author Samuel Lai, an astronomer at CSIRO. “This is the first truly independent validation of the EHT image, an important step in the scientific process.”

Why a picture of a black hole can be argued over

M87* sits about 55 million light-years away, and its shadow spans a sliver of sky so small that Lai compares the task to “standing in your backyard and trying to take a photo of a doughnut on the Moon.” No single dish can do it. The Event Horizon Telescope links radio observatories in Chile, Mexico, Hawaii, Arizona, Spain and elsewhere, so that the whole Earth acts as one telescope.

That virtual telescope is mostly empty space. Each pair of dishes samples only a thin track of the information needed for an image, and every station adds errors from its electronics and from the turbulent air above it. Software has to fill the gaps, and the choices made in that software can nudge the final picture.

The original collaboration knew this and ran four imaging teams in parallel, blind to one another. All four found a ring, but they disagreed on fine details such as where it is brightest. Then a Japanese group led by Makoto Miyoshi reprocessed the public data with a classic technique and reported that, if the analysis allowed for a wider patch of sky, a single elongated bright spot fitted better than a ring. The dispute turned on assumptions, and assumptions are exactly what an independent test has to remove.

How they did it

The CSIRO team attacked both weak points at once. First, the data. The standard approach combines measurements from pairs of telescopes, which carry each station’s calibration errors along with the signal. The team instead used “closure invariants,” quantities built from three telescopes at a time, in which those errors cancel out. “Our technique combines triplets rather than the standard pairs of telescope data to eliminate those errors,” said project leader Nithyanandan Thyagarajan.

Second, the software. The team built a generative AI system called GenDIReCT, similar in spirit to the tools that turn text into pictures, but steered by the closure invariants instead of a written prompt. It learned what real images look like from a standard library of everyday photos plus simple shapes such as disks and rings. To make sure those rings were not biasing the answer, the researchers retrained it without any geometric shapes and found no notable difference.

Before it saw any M87* data, the program had to pass tests on synthetic data and on real observations of other galaxies, reproducing published images of the quasar 3C 279 with 98% similarity and of the radio galaxy Centaurus A with 91%. Then the team froze the model, so nobody could tweak it after seeing the answer. It worked through eight nights of public data from three observing seasons, each with three sky windows, producing around 300 candidate images per dataset. As a cross-check that owed nothing to machine learning, the team also fitted the same triplet data with a conventional statistical model of a ring.

What the ring tells us

Every dataset with good telescope coverage produced a compact ring about 42 microarcseconds across, whatever the size of the sky window. That stability is the direct answer to the rival analysis, which found its bright spot only when the window grew large. The CSIRO team argues that the earlier effect most likely stems from a night with sparse coverage and lower-quality data. The machine learning images gave 42.2 microarcseconds and the conventional fit 42.3, matching within the error bars. Resolving detail this fine, the paper notes, is like picking out the width of a human hair from more than 1,000 kilometres away.

From that size, the team estimates a mass of 6.5 billion Suns, give or take 0.5 billion. That sits almost exactly on an independent estimate from the motions of stars in the galaxy, about 6.6 billion Suns, and well away from an estimate based on gas motions, about 3.5 billion, which would have produced a ring roughly half as wide. Sagittarius A*, the black hole at the centre of the Milky Way, weighs about 4 million Suns, so M87* is more than a thousand times heavier.

The images also suggest we see the black hole almost face-on, about 20 degrees from its jet, close to the 17 degrees measured from the jet itself. The brightest part of the ring shifts between seasons, which the authors attribute to turbulent gas falling into the black hole rather than to the method.

What it doesn’t settle

The new analysis is independent in method, not in data. It reuses the same public Event Horizon Telescope measurements, so it cannot rule out a problem baked into the raw recordings; it shows that the ring is not a product of the original teams’ processing choices.

Some features remain soft. The ring’s thickness, the brightness of its centre and how lopsided it looks all changed between methods and between nights, and the authors say these probably do not reflect the true source. One night, when a key telescope in Mexico was missing, gave the most ambiguous results.

The mass is less firm than the headline figure suggests. Converting a ring size into a mass relies on a calibration factor taken from computer simulations, and the full range of that factor moves the estimate between 6.0 and 6.7 billion Suns. The images cannot pin down how fast the black hole spins, although they favour some rotation. And the team warns that a good fit to the data is necessary but not sufficient proof that an image is correct, since there is no ground-truth picture of M87* to check against.

Common questions about the M87* black hole image

Is the first black hole image real?

Yes, according to this independent check. A team outside the original analysis, using a different way of combining the data and an AI that had never seen a black hole, recovered a ring of the same size and orientation. Fine details such as the ring’s thickness still vary between methods.

How do you photograph a black hole?

You photograph the hot gas around it. Light from that gas bends around the black hole and forms a bright ring around a dark shadow. To see a shadow that small, the Event Horizon Telescope combines radio dishes across the planet so they work like one telescope as wide as the Earth.

How big is the M87* black hole?

The new analysis puts its mass at about 6.5 billion times that of the Sun, more than a thousand times heavier than the black hole at the centre of our galaxy. Its ring spans about 42 microarcseconds on the sky, far too small for any single telescope to resolve.

What comes next

The data span four nights in April 2017, two in April 2018, when the Greenland Telescope joined the array, and two in April 2021. The original M87* picture was unveiled on April 10, 2019, and the challenge from Miyoshi’s team appeared in 2022 and 2024. The CSIRO researchers say their frozen model first ran on the M87* data on December 4, 2025, and the paper appeared online on November 11, 2026, with its reconstructions and code made public.

The authors expect planned observations with the James Webb Space Telescope’s NIRSpec instrument to map gas and star motions in the core of M87 and tighten the mass. Thyagarajan says the triplet technique could also help fields that stitch many sensors into one map, from medical imaging to tracking earthquakes. A bigger prize sits further out: telescopes in space could resolve the thin photon ring inside the shadow and, with it, finally measure how fast M87* spins.

Reference: Lai et al., “Independent interferometric observable confirms black hole shadow in M87*,” Nature, 2026. DOI: 10.1038/s41586-026-11241-9

Tags: artificial intelligence, astronomy, black holes, CSIRO, Event Horizon Telescope, M87