The black hole photo wasn't taken, it was computed by four teams that couldn't talk to each other
The 2019 image of M87's black hole: four imaging teams worked in isolation, partly deceiving one another, using Frosty the Snowman as a test image — all to prove they hadn't computed what they wanted to see into the data.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
To understand how knowledge travels, first look at how it is blocked
Galison says he studies secrecy not for its own sake but as proof by contradiction: to understand a car you need to know that racing stripes and tire color don't affect function, but pull the spark plug and it stops. He gives the example of Freud — Freud's theory of unconscious repression and censorship came directly from the physical censorship technologies of World War I Vienna: Russian censors blotted out passages with a thick black ink they called caviar, while the Germans reset the type so you couldn't tell where something had been cut. These concrete blocking techniques echoed into Freud's theory of how the mind represses traumatic experience.
— Peter GalisonThe moment of seeing the image was a double panic
The team saw the image nearly a year before the official release, and alongside the excitement came two kinds of panic. One was the fear of a leak, of immediately losing control. The other was more fundamental — the history of science is full of retracted claims. He mentions a local episode that seemed to confirm inflation and later had to be retracted, and the dozens of "gravitational wave detections" in the 60s and 70s that were all eventually retracted, until LIGO actually measured them in 2015. Galison's doctoral thesis was titled How Experiments End, on exactly how people distinguish artifacts from real effects, so the whole following year was spent asking: what if it's this? What if it's that?
— Peter GalisonFour teams were isolated — and partly deceived one another
Inside the imaging group there were four teams, forbidden to talk to each other while working, each covering their windows, even slightly deceiving one another about what they were doing. The procedure: each team was first given a batch of practice images, without being told what they were, and asked to faithfully reproduce each one using their best method. Some of these images were random grabs from the internet, some were astronomical images, some were black hole simulations, and one was Frosty the Snowman — because nobody could possibly have anticipated it. Only once all four teams could accurately reproduce a test image whose content they didn't know were they given the real data; the four teams then produced very similar results, and that was the first time they genuinely hoped the image was reliable.
— Peter GalisonCollaboration without a center is what constitutes the experiment
The camera itself is a composite: eight telescopes at six sites, which only combined can synthesize a telescope the size of the Earth and achieve enough resolution. Galison stresses this is not "many people doing the same thing" — not a crowd counting fish in a river and adding up the totals — but different people in different places doing very different things, where the coordinating activity itself constitutes the experiment. He contrasts CERN: CERN has an office you can walk into to meet the director-general, whereas EHT has a director, but the funding is dispersed, the telescopes are dispersed, the work is dispersed, and nobody would agree that this experiment has a single center.
— Peter GalisonTwo imaging methods, each with its own narcissism risk
The four teams used two classes of technique. Two teams used clean, the old method beloved by astronomers: treat the image as a collection of point sources, restore blurred bright blobs into star-like points — the advantage being that you can see things that were otherwise invisible. The other two used more modern Bayesian or forward methods: presuppose that the object has edges, is convex, sits roughly somewhere in the frame, and construct a cost function from deviations in those parameters — the advantage being that you can extract signal from noise that would otherwise be lost, which is exactly how phone cameras work. The danger: assume too much and you get the thing you are looking for, like Narcissus gazing at his own reflection in the spring. So what they feared most was precisely this: knowing what you want to see, and therefore actually seeing it.
— Peter GalisonLet the computer find the parameters itself, then try them on real data
To rule out the four teams sharing one set of expected biases, they added another layer: let the computer search the test images on its own for the parameter set that best reproduces them, then apply the settings it found to the real data. Galison says this is closer to machine learning than to AI in today's sense — not many layers, traceability not entirely out of control, quite simple by today's standards. He describes that year as "living dangerously," constantly inventing new ways to challenge the image.
— Peter GalisonThree epistemological stages, walked backwards
In Objectivity, Daston and Galison distinguish three eras in the function of scientific images: idealization from the 18th century on (you want to see the ideal skeleton, not mine or yours), mechanical objectivity (erase the self as far as possible, let nature write itself onto the page), and expert judgment (not the apperception of genius but trainable judgment, such as distinguishing different kinds of epileptic seizure on an electromyogram). Black hole imaging walked these three steps backwards: first let the four teams each use expert judgment, then let the computer mechanically search parameters, and finally return to idealization with the averaged image.
— Peter GalisonThe averaged image was a compromise argued into existence
There was a big fight inside the collaboration over whether to release the averaged image. Supporters said that way no team could claim "this is our image," but Galison finds that reason too saccharine and epistemologically unconvincing; the stronger reason is that if the individual images agree (and they did all show a bright crescent on the south side), it will show up in the average, and the average can be compared against each individual image. He gives a counterexample: if King Kong smashed a hole in the Empire State Building that was exactly congruent to the building, averaging the two would give a flat field, resembling neither tower nor hole. Opponents said the averaged image doesn't correspond to any data that actually exists. In the end they decided to average — conservative, emphasizing commonality, downplaying individuality — and the image was released on April 10, 2019.
— Peter GalisonAn image is valuable because it can surprise you
Galison says that if you want to find a rare effect, you usually design an instrument to pick it out of the noise — but then you are designing to look for a particular thing. Images work the other way: through an image, nature can show us something we did not anticipate. That is the ultimate surprise of image-making — you can find something that isn't on the menu of your expected results. Which is also why they handed out Frosty the Snowman: to make sure that even something as unexpected as that fictional character could show up.
— Peter GalisonIn their own words · checked verbatim
Well, for me to understand something like how knowledge is transmitted, I want to know how it's blocked. And if I understand how it's blocked, then I understand better how knowledge works.
Peter Galison12:11
And then with with black holes, they're the ultimate invisible object. They're the objects that reflect no light. They emit no light. They seem impossible, right?
Peter Galison15:12
One of them was a Frosty the Snowman because we figured nobody could have anticipated that.
Peter Galison24:23
It's not just a lot of people doing the same thing. It's the coordinative activity itself that constituted the experiment.
Peter Galison26:25
The danger is if you assume too much you get what you are looking for and you end up like narcissists looking at the spring and seeing his own reflection.
Peter Galison29:26
Images are the other way around. They're nature can show us things that we had not anticipated. And I think that is the ultimate surprise that image making holds for us.
Peter Galison39:37
So sometimes things that seem not even real become real, fundamental and practical.
Peter Galison43:38
Figures
| Size of the EHT imaging team (when the 2019 image was completed) | about 200 people | 19:16 |
| Number of independent teams inside the imaging group | 4 | 23:23 |
| Telescope sites and count | 6 sites, 8 telescopes | 25:24 |
| Telescope altitude | 13,000 feet | 25:24 |
| Release date of the M87 black hole image | April 10, 2019 | 33:30 |
| Year LIGO actually detected gravitational waves | 2015 | 20:17 |
| Year Einstein proposed general relativity | 1915 | 42:37 |
Glossary
- clean / CLEAN algorithm
- A classic astronomical deconvolution algorithm that restores blurred bright blobs into point-like stellar sources.
- Bayesian / forward methods
- Imaging that presupposes object features (edges, convexity, etc.) and builds a cost function from deviations in those parameters.
- mechanical objectivity
- The epistemological stage of scientific imaging proposed by Daston and Galison, which holds that the self should be erased and nature allowed to write itself.
- tacit knowledge
- Polanyi's concept: knowledge that is hard to put into words and is acquired through personal transmission and feel.
- photon ring
- The thin ring formed by light orbiting around a black hole, the target the next generation of telescopes wants to resolve.
How to listen
Engineers and researchers interested in how scientific knowledge is "made" and "verified" — especially those working on large-scale distributed collaboration, data pipelines, or machine-learning interpretability.
The opening 0:00–11:00, the academic-career and HPS-discipline-history reminiscence, can be skipped.