In September, Pavel Rabtsevich asked an AI model to discover a planet. He pointed it at a batch of telescope data, handed it a high-level plan, and about two weeks later it returned two possible worlds circling a star 116 light years from Earth. Professional astronomers are now checking the result. The story is a useful test case for what happens when expert-grade analysis becomes something a non-expert can summon in plain language.
What the bot actually did
Rabtsevich works in Spain as an AI specialist and has no formal astronomy education. For his search, he used Claude Code, a coding tool made by Anthropic, and asked it to sift through data from more than 120,000 stars observed by NASA’s Transiting Exoplanet Survey Satellite (TESS) in late 2025 and early 2026.
The method itself is classic. A star that dims regularly, again and again, may have a planet crossing in front of it and blocking a sliver of light on each pass. Rabtsevich guided Claude to use a popular planet-detection algorithm to hunt for exactly those dips.
Claude zeroed in on a star called TIC 4206066, which showed two separate dimming cycles: one roughly every 3 days and another every 11 days. That pattern suggests two planets, with the inner one looping the star faster than the outer one. Rabtsevich then had Claude check archival TESS data, and the same dimming appeared in records from 2018 and 2020. To validate the signals, Claude found a planet-detection algorithm written and publicly released by researcher Michelle Kunimoto, and ran it.
Notice the division of labor. The human supplied direction and judgment about what to ask. The AI handled the volume: scanning an enormous number of stars, locating existing tools on the internet, and chaining them together.
Why professionals are uneasy
Amateurs have used public tools from NASA and other researchers to find planets for decades. What has changed is the barrier to entry. An AI that can scour the internet for those tools and run them on massive datasets means a person no longer needs to understand how the software works in order to operate it.
That is where Kunimoto, an exoplanet researcher at the University of British Columbia in Vancouver, raises a pointed concern. Some planet hunters, she says, don’t fully understand the algorithms the AI is using, which means they can’t validate what it produces. She describes a huge, overwhelming surge of AI-driven papers, which tend to carry planet candidates that are not the best quality. Her conclusion is blunt: the systems used to disseminate science are being overwhelmed.
Here is the part most coverage misses. The bottleneck in science was never only finding signals. It was deciding which signals deserve belief. When discovery gets cheap, verification becomes the scarce resource, and it lands on a small pool of reviewers and editors who now face a flood of submissions.
A case that went right, and what made it work
Rabtsevich’s result is worth reading carefully, because it shows the safeguards that separate a useful contribution from noise. He posted his findings to the platform Zenodo as a preprint on 25 September. That work has not been peer reviewed. But TESS astronomers completed an independent analysis of the signals and added them to their official list of planet candidates.
Three things stand out. The detection was cross-checked against older archival data from 2018 and 2020. The validation step used a respected, publicly available algorithm. And, critically, professionals independently reanalyzed the signals before they entered an official list. The word “candidate” matters here too: a candidate is a promising lead, not a confirmed planet. Rabtsevich’s own framing, two possible planets, reflects that restraint.
The lesson is not that AI discoveries are suspect. It is that an AI finding earns trust through independent confirmation, the same way any other finding does.
What this means beyond astronomy
The pattern extends well past exoplanets. AI is turning many specialist workflows into something a motivated outsider can attempt, and that is genuinely valuable. People with curiosity and time, but without a degree in the field, can now contribute to real research. Professionals, in turn, gain a larger pool of leads to examine.
But the benefit depends on a shift in what counts as expertise. Operating the tool is becoming easy. Understanding what the tool did, why it might be wrong, and how to test it is the skill that matters now. An amateur who can explain the dimming method and check a result against archival data adds something. An amateur who forwards raw output adds to the pile.
For institutions, the challenge is capacity. Journals, preprint servers, and review processes were built for a world where producing a plausible analysis took real effort. As that effort drops, the filtering layer needs support, and the people doing that filtering become more important, not less. Human judgment is the part of the pipeline that cannot be skipped.
In short
A non-specialist used an AI coding tool to scan more than 120,000 stars and surfaced two candidate planets around TIC 4206066, signals that TESS astronomers independently analyzed and added to their official candidate list. The same tools that make this possible are also producing a surge of lower-quality candidates, which Kunimoto warns is overwhelming the systems that spread science. The difference between the two outcomes is verification: cross-checking against archival data, using trusted algorithms, and having professionals confirm the result. As AI makes finding signals easy, the real value shifts to the people who can judge which ones are real.
Based on reporting from Nature.