Independent method confirms M87 black hole ring from the same EHT data

A CSIRO team rebuilt the M87* image from eight days of public Event Horizon Telescope data with a calibration-proof AI method and found a ring 42.2 ± 1.5 microarcseconds across, matching 2019, though it reanalyses the same data rather than new observations.

Peer reviewed · Reanalysis of existing public data · Machine-learning imaging method

The famous orange ring around the black hole in galaxy M87 has passed an outside check. Samuel Lai and three colleagues at CSIRO, Australia’s national science agency, report in Nature that a different imaging method, built to ignore telescope calibration errors and to start with no idea of what a black hole looks like, recovers the same ring the Event Horizon Telescope (EHT) published in 2019. That matters because a few earlier reanalyses argued the ring might be an artefact of how the image was made. The limit: this is the same telescope data, processed a new way.

Evidence at a glance

Is this a big deal?

A solid confirmation, not a new discovery. The ring was already the mainstream reading. What this paper adds is a check by people outside the EHT Collaboration (none of the four authors is on the 2019 EHT paper’s author list, by our check), using a method that removes two common worries: calibration errors and the hand-set starting guesses that standard imaging codes need. It narrows, but does not end, the debate.

What scientists found

Across all eight datasets, the team’s images converge on a compact ring about 40 μas wide. On the best day of each year (11 April 2017, 21 April 2018 and 18 April 2021) the ring kept the same size and orientation whatever image size the algorithm was told to search, and its brightest side sat to the south or southwest.

The authors measured an average diameter of 42.2 ± 1.5 μas from their images. A separate test that fitted a simple mathematical ring model directly to the data gave 42.3 ± 1.1 μas. In the paper’s words, the method unequivocally reproduce[s] a ring-like morphology.

From that size, an assumed distance of 16.8 megaparsecs (about 55 million light-years) and a standard scaling, they estimate a mass of 6.5 ± 0.5 billion Suns. The ring’s appearance best matched simulated black holes viewed about 20° from the jet that points toward us.

How small is 42 μas? By our calculation, it is the width of an object about 8 cm across, roughly a doughnut, seen on the Moon from Earth. Lai used the same doughnut comparison in the CSIRO release. At M87’s distance, that angle spans about 710 times the Earth–Sun distance, or about 4 light-days.

How they found it

The EHT links radio dishes on several continents so they act like one Earth-sized telescope observing at 1.3 mm wavelength. The catch is that the dishes sample only scattered patches of what a real Earth-sized dish would see, and each station’s signal is distorted by the atmosphere above it. Turning that into a picture needs assumptions, and the choice of assumptions can change the result.

Nithyanandan Thyagarajan had developed the maths of closure invariants: combinations of measurements from triangles of telescopes in which each station’s own errors cancel out. Lai built GenDIReCT around them. Its diffusion model, the same family of AI used to generate pictures from text, learned what images generally look like mainly from CIFAR-10, a library of small photos of things like cars and animals. It was not trained on any astronomical images.

The team ran it on public EHT data from 5, 6, 10 and 11 April 2017, 21 and 25 April 2018, and 13 and 18 April 2021, at three different image sizes up to 450 μas square. Before M87, they had tested it on simulated data and on real EHT observations of the quasar 3C 279 and the galaxy Centaurus A, which have jet shapes rather than rings; it reproduced the published EHT images of those at 98% and 91% similarity.

Why it matters to you

Nothing about daily life changes. The payoff is trust: a widely shared science image now has an outside check, and Einstein’s prediction of a dark shadow ringed by light passes again with a different toolkit. The authors also say calibration-free imaging could help other sparse-data problems, but that is a suggestion, not something tested here.

What the evidence does not show

It does not provide new observations. Both the EHT and the CSIRO team start from the same calibrated public data, so a flaw in the raw recordings would not be caught.

The AI is not prior-free. It learns an implicit sense of what images look like from its training photos, and the authors note that its spread of solutions is not a formal Bayesian uncertainty. On some lower-quality days, a few reconstructions showed a less clear ring.

The study also leaves the fine detail open. Ring width, the brightness of the dark centre and the lopsidedness of the ring varied between methods and days, so the authors say those features probably do not reflect the source itself. Black-hole spin is not constrained.

How this fits what we already knew

The EHT published the first image in April 2019, a ring 42 ± 3 μas across implying 6.5 ± 0.7 billion Suns. In 2022, Makoto Miyoshi and colleagues reanalysed the public 2017 data and argued in The Astrophysical Journal that the ring was an artefact of a narrow field of view and gaps in the telescope coverage. The new paper directly tests that claim. It finds the field-of-view sensitivity disappears when the higher-quality days are used, and suggests the earlier result may reflect the sparse 5 April 2017 data.

Another machine-learning reanalysis, PRIMO, reported in 2023 a sharper ring of 41.5 ± 0.6 μas, but PRIMO was trained on simulations of black holes. The CSIRO model’s training on unrelated photos is what makes its independence argument stronger.

The new mass also takes a side in an older dispute. Star motions gave about 6.6 billion Suns; gas motions gave about 3.5 billion. The ring size fits the stellar value. A 3.5-billion-Sun black hole would cast a ring only about 22 μas across, by our calculation, close to the EHT’s resolution limit of about 20 μas.

What the researchers and the EHT said

“This is the first truly independent validation of the EHT image, an important step in the scientific process,” Lai said in a CSIRO release republished by Phys.org. Laurent Loinard, director of the EHT Collaboration, welcomed the work in the same release; he leads the team whose result was being checked. The paper was reviewed for Nature by Andre Offringa and an anonymous referee. SciQuest has not yet obtained comment from a researcher outside both groups, including the authors of the earlier critiques.

How other headlines framed it

• Phys.org, carrying CSIRO’s release, framed it as an independent technique validating a landmark image, with endorsement from the EHT director.
• The Conversation posed the accuracy question in its headline; the piece is written by Lai and Thyagarajan themselves.
• The Register framed it as scientists checking their work, although the four authors were not part of the 2019 EHT paper.

None of the three stresses that the check reuses the same telescope data, which is the main reason to read “independent” as independent method.

What happens next

The authors point to two upgrades. Observations of M87’s core with the James Webb Space Telescope’s NIRSpec instrument should tighten the stellar mass estimate. Further out, radio dishes in space could add longer baselines and resolve the thin photon ring, which would allow a direct measure of the black hole’s spin.

In one sentence

A calibration-proof AI method trained on everyday photos, run by researchers outside the EHT, finds the same 42-microarcsecond ring around M87’s black hole in data from three years, though it reuses the original telescope measurements.

Read next

• JWST measures the gas near Sagittarius A*: the other black hole the EHT has imaged, close to home.
• What is the biggest black hole? How black-hole masses like M87*’s are weighed, and why most are estimates.
• How is a black hole formed? The background on what the EHT is looking at.

Sources and disclosure

1. Lai S, Thyagarajan N, Wong OI, Diakogiannis F. “Independent interferometric observable confirms black hole shadow in M87*.” Nature 659, 317–323 (2026). doi:10.1038/s41586-026-11241-9. Open access, CC BY 4.0. Authors declare no competing interests; no specific funding.
2. Event Horizon Telescope Collaboration. “First M87 Event Horizon Telescope Results. I. The Shadow of the Supermassive Black Hole.” The Astrophysical Journal Letters 875, L1 (2019). doi:10.3847/2041-8213/ab0ec7.
3. Miyoshi M, Kato Y, Makino J. “The Jet and Resolved Features of the Central Supermassive Black Hole of M87 Observed with the Event Horizon Telescope (EHT).” The Astrophysical Journal 933, 36 (2022). doi:10.3847/1538-4357/ac6ddb.
4. Medeiros L, et al. “The Image of the M87 Black Hole Reconstructed with PRIMO.” The Astrophysical Journal Letters 947, L7 (2023). doi:10.3847/2041-8213/acc32d.
5. CSIRO release “Independent technique validates landmark M87* image,” republished by Phys.org, 11 November 2026 (source of the Lai quote and the Loinard comment).

Disclosure: This article was drafted with AI assistance and checked against the sources above, including the full Nature paper. Calculations by SciQuest: 1 μas = 4.848 × 10⁻¹² radians, so 42.2 μas = 2.05 × 10⁻¹⁰ rad; at the Moon’s 384,400 km that spans 7.9 cm; at 16.8 Mpc (5.18 × 10²⁰ km, about 55 million light-years) it spans 1.06 × 10¹¹ km ≈ 709 AU ≈ 4.1 light-days; the ring scales with mass, so 3.5 billion Suns gives about 42 × 3.5/6.5 ≈ 23 μas by simple proportion, and about 21 μas using the textbook shadow formula 2√27 GM/(c²D), hence “about 22 μas”; 1.3 mm divided by Earth’s 12,742 km diameter ≈ 21 μas resolution. The graphic is a SciQuest original built from the values in the cited papers. Spotted an error? See our corrections policy.