Independent AI reanalysis confirms 2019 black hole image was accurate

An independent analysis of the data behind the world's first picture of a black hole has reproduced the same glowing ring reported in 2019, offering what its authors call the first truly independent confirmation of the landmark image. The study, led by astronomers at Australia's national science agency, CSIRO, was published Wednesday in the journal Nature.

The original image, released in April 2019 by the Event Horizon Telescope (EHT) collaboration, showed a bright, uneven ring of light surrounding a dark central shadow at the heart of the galaxy Messier 87, roughly 55 million light-years away. It was hailed at the time as the first direct visual evidence of a black hole's event horizon, the boundary beyond which nothing, not even light, can escape. The EHT project linked eight radio observatories on four continents, effectively turning the planet itself into a single telescope large enough to resolve an object that small at that distance, and the resulting image was assembled by a collaboration of several hundred scientists. Because the picture was reconstructed from sparse, noisy radio data rather than captured the way an ordinary photograph is, some astronomers have since questioned how much of its fine structure was shaped by the computer algorithms used to fill in the gaps, rather than by the black hole itself.

A method designed to avoid bias

The new CSIRO-led team, headed by Dr. Samuel Lai and Dr. Nithyanandan Thyagarajan, built a reconstruction pipeline from scratch rather than adapting the EHT's own imaging software. Their method, called GenDIReCT, combines two safeguards against bias. First, it works with "closure invariants," mathematical combinations of signals from three telescopes at a time rather than the usual pairs, a step the researchers say cancels out atmospheric distortion and instrumental calibration errors before any image is built. Second, it uses a machine-learning model that was trained only on ordinary photographs, of cats, dogs and cars, among other everyday objects, so that it carried no built-in assumption about what a black hole ring should look like.

To guard further against unconscious bias creeping in during analysis, the team says it locked its processing pipeline in December 2025, before applying it to the 2017 M87 observations that produced the original image. A second, more conventional statistical method was used as a cross-check on the results.

Run on the same raw interferometry data the EHT collaboration used, the new pipeline produced a ring of consistent size and orientation across the observing nights, matching a black hole with a mass of roughly 6.5 billion Suns viewed nearly face-on, in line with the EHT's published figures. Some fine details, including the exact width of the ring and the brightness distribution around its circumference, differed modestly between the two methods, which the authors describe as expected given the different assumptions built into each approach.

Why independent verification matters

The 2019 image relied on imaging algorithms that fill in gaps in sparse data using statistical assumptions, a standard but inherently interpretive step in radio interferometry. A 2022 reanalysis by a separate group of researchers, working with different imaging software, produced a markedly different picture showing a core and jet rather than a clean ring, fueling a public debate over how much the EHT's chosen algorithms had shaped its famous result. The EHT collaboration itself later published additional images using newly developed techniques that it said recovered the same ring structure, but Wednesday's study is the first to build an entirely separate pipeline, with no shared code or training assumptions, from the ground up.

"Our independent results enhance confidence in one of the most significant modern scientific results," the researchers wrote.

Laurent Loinard, director of the EHT Collaboration, who was not an author on the paper, said the agreement between the two independent methods strengthens the case that the ring is a genuine astrophysical feature rather than a processing artifact. According to the CSIRO researchers, Loinard described it as "gratifying" that an entirely independent method recovers the ring structure of M87* across several years of data in agreement with the EHT's own published results.

Lai, describing the technical challenge of resolving a black hole roughly 55 million light-years away, compared it to "standing in your backyard and trying to take a photo of a doughnut on the Moon." He called the result "the first truly independent validation of the EHT image, an important step in the scientific process."

Beyond astronomy

The CSIRO team previously tested GenDIReCT on EHT observations of two other galaxies, Centaurus A and 3C 279, in work posted earlier this year, before applying the frozen pipeline to M87. Thyagarajan said the closure-invariant approach, which is built to strip out instrument-specific noise from data gathered by physically separate sensors, is not limited to radio astronomy. He suggested it could eventually be adapted for medical imaging techniques that combine readings from multiple detectors, or for triangulating the source of earthquakes from seismometer networks.

The findings were reported Wednesday by multiple outlets, including a detailed account in The Register, which noted that the four-person Australian team worked separately from the hundreds of scientists involved in the original EHT collaboration. The 2019 image itself remains freely available through the European Southern Observatory, one of the EHT's partner institutions.

A related, not yet peer-reviewed preprint from the same research team describes extending GenDIReCT toward Sagittarius A*, the much smaller black hole at the center of the Milky Way, whose 2022 EHT image proved harder to reconstruct because the surrounding gas changes brightness on timescales of minutes. In tests on simulated data standing in for that target, the method recovered the expected structure in nearly every configuration tried, suggesting the same bias-resistant approach could eventually be applied to the Milky Way's own black hole.