An AI trained on photos of cats, dogs and cars rebuilt the M87 black hole image and found the same ring in 2017, 2018 and 2021 telescope data
Four CSIRO researchers in Australia built an AI imaging method that learned from everyday photos, not black holes, and ran it on the Event Horizon Telescope's public M87 data. Every year it produced the same ring, 42.2 microarcseconds across, matching the famous 2019 image and pointing to a black hole of about 6.5 billion suns. Nature published the study on November 11, 2026.
Original AI-generated clay editorial illustration for Hooray Machine. Conceptual artwork, not a photograph or documentary evidence.
In April 2019 the Event Horizon Telescope team showed the world its first picture of a black hole: a glowing ring around the giant at the heart of galaxy M87, about 55 million light-years away. Radio dishes across the globe had acted as one Earth-sized telescope, leaving big gaps to fill in. Lead author Samuel Lai compares it to photographing a doughnut on the Moon from your backyard. Critics asked whether the ring came from the processing, and one 2022 reanalysis saw no clear ring at all.
The clever bit
A four-person team at CSIRO, Australia's national science agency, rebuilt the picture while dodging the fragile steps. The usual method pairs signals from two dishes, then has to estimate and strip out errors from each instrument and from the air above it. This team combined measurements from three dishes at once, a construction called closure invariants. That three-dish step cancels each station's own errors and keeps a cleaner read of the object's shape. An AI imaging model called GenDIReCT then turned those numbers into pictures. It learned what images look like from everyday photos of things like cats, dogs and cars. The team froze the model after testing it, and only on December 4, 2025 let it look at M87.
Hooray
All eight nights of public data, from 2017, 2018 and 2021, came back with a compact ring averaging 42.2 microarcseconds across, brightest toward the south or southwest. That size points to a black hole of about 6.5 billion suns, in line with earlier estimates from the motion of stars in the galaxy. Lai says the result is almost identical in size and shape to the 2019 image. A second, conventional statistics method using the same three-dish data found the same structure.
Studies like this strengthen our confidence in what we have learned about black holes, and enrich the toolkit available to the whole community.
— Laurent Loinard, director, Event Horizon Telescope Collaboration, via csiro.au
What comes next
The team thinks the three-dish trick could help anywhere several sensors are combined to map something hidden, such as ultrasound or MRI scans and earthquake imaging, though the paper doesn't test that. Sharper pictures from future telescopes in space could pin down how fast M87's black hole spins, which today's images can't settle. Earlier outside checks also found a ring, and this one leans less on human choices.
How solid is it?
Promising: Results measured in a study or bounded trial. Limitations and wider testing still matter.
Peer-reviewed, open-access paper in Nature by Samuel Lai, Nithyanandan Thyagarajan, O. Ivy Wong and Foivos Diakogiannis (CSIRO and ICRAR), received April 17, accepted October 2 and published November 11, 2026. It analyses public Event Horizon Telescope data from eight nights in 2017, 2018 and 2021 at three fields of view, with the code and reconstructions public. A separate statistical model of the same data gave a matching ring size (42.3 versus 42.2 microarcseconds), and the implied mass agrees with earlier star-motion estimates. The authors declare no competing interests.
The fine print Update, Nov 15, 2026: We revised the fine print and the clever-bit section after lead author Samuel Lai wrote in. This work reuses the Event Horizon Telescope's own public data rather than new observations. Combining measurements from three dishes at once (closure invariants) cancels each dish's own atmospheric and instrumental errors, so those station-specific problems do not simply carry over; other issues in the data can still remain. It isn't the first outside check either: in 2022 the collaboration listed four earlier independent reconstructions, one co-written by project leader Nithyanandan Thyagarajan, though those leaned more on analysts' starting guesses. Finer details, like the ring's width and brightness pattern, still differ between methods. The same day the paper came out, Nature published a news article by Davide Castelvecchi, "Iconic first image of black hole holds up to scrutiny," which reports that the new analysis largely corroborates the original findings and recounts the 2022 reanalysis, by Junichiro Makino and colleagues, that saw a glob of radio emissions rather than a ring.
An AI trained on photos of cats, dogs and cars rebuilt the M87 black hole image and found the same ring in 2017, 2018 and 2021 telescope data · Hooray Machine