University of Arizona astrophysicist Chi kwan Chan, a member of the Event Horizon Telescope collaboration, describes the narrow, iterative role OpenAI's Codex plays in his plasma modeling, and the line it does not cross.
The honest version of this story starts with what the source actually is: a vendor profile. OpenAI published a post on June 11, 2026, titled "How an astrophysicist uses Codex to help simulate black holes," and it reads like a customer case study. The researcher, Chi-kwan Chan of the University of Arizona's Steward Observatory, is real, and so is his membership in the Event Horizon Telescope collaboration. But the framing is OpenAI's, and the reader should know that before reading another sentence.
Chan works on the computational side of black hole physics. According to the OpenAI Applied AI blog post, he uses Codex to refine and test algorithms that simulate the movement of charged particles, electrons and ions, in the plasma swirling near a black hole's event horizon. That plasma is the medium EHT actually images. The 2019 M87* image released by EHT showed the silhouette of a supermassive black hole surrounded by glowing, lensed emission from that plasma, and it was a genuine milestone in observational astronomy, the product of years of work by a collaboration of hundreds of researchers and a global network of radio telescopes.
What Codex does, in the version of the workflow Chan describes to OpenAI, is sit inside the loop of writing simulation code, running it, checking the output, and rewriting it. The coding assistant is being used to compress that iteration cycle: trying variations of an algorithm, testing numerical behavior, and surfacing alternative implementations. It is not generating the physics, not interpreting the EHT observations, and not making the images. The boundary is honest: a productivity tool in a research loop, not a co-author.
This distinction matters because the EHT program is in the middle of a different kind of work right now. The collaboration is currently gathering observations aimed at producing the first video of a supermassive black hole, focused on M87's central object. That is a planning and execution challenge that lives in telescope time, data pipelines, imaging teams, and atmospheric calibration. None of which a coding assistant touches. Anyone reading the OpenAI post as a signal that AI helped produce the 2019 image, or is about to produce the M87* video, would be reading it wrong.
The wider point for a builder or a researcher using AI coding tools is closer to something humbler. The black hole simulation problem Chan works on is computationally hard because general relativity, the theory that describes gravity as mass and energy bending space-time rather than as a pulling force, has to be coupled to magnetized plasma physics. Getting that coupling right is a long exercise in numerical methods, and the bottleneck is often the time between proposing a new algorithm and producing the run that shows it works. That bottleneck is where Codex is being applied. It is iteration support, not discovery.
The OpenAI framing is the natural one to be skeptical of, because the post is built to make a customer profile feel like a research breakthrough. Chan is a real researcher with a real affiliation doing real work, and the workflow he describes is plausible and probably accurate in its narrow claims. But the marketing lens of the source means specific performance numbers, exact code changes, and any framing of Codex as transformative should be set aside until they appear in a peer-reviewed paper, a code release, or an EHT publication rather than in a vendor blog.
What to watch next is straightforward. EHT's M87* video program is the actual scientific deliverable on the horizon, and its progress will be reported by the collaboration itself, not by a coding tool's parent company. Chan's own work, like that of other computational astrophysicists in the collaboration, will show up in published simulation papers and in code made available to other groups. The honest story is the boring one: a working researcher uses a productivity tool to iterate faster on a hard numerical problem inside a large scientific collaboration. It is worth reporting because that loop is becoming common, and because the gap between "AI helped with code" and "AI did the science" is now the one to keep clear.