Amateur astronomer uses Claude to find potential new planet

From spreadsheets to space

Pavel Rabtsevich, a 28-year-old product manager living in Spain, recently conducted a search for planets outside our solar system. Despite having no formal training in astronomy or owning a telescope, he used an internet connection and several AI subscriptions to hunt for new worlds. He focused his efforts on public data captured by the Transiting Exoplanet Survey Satellite (TESS), a space telescope launched by NASA in 2018.

TESS is designed to find exoplanets by measuring changes in stellar luminosity. When a planet passes between a star and our telescopes, it blocks some light, causing a measurable dip in brightness. Rabtsevich specifically looked at a narrow sliver of data captured between November of last year and early January. This dataset included measurements for over 126,000 different stars.

To process this information, Rabtsevich handed the data to three different AI systems: TypeSafe’s Jev, Anthropic’s Claude Code, and OpenAI’s Codex. He prompted these models to look for specific patterns of dimming that might indicate a planet is orbiting a star. He did not simply accept the first results the AI provided, but instead pushed the models through multiple rounds of iterative testing to improve their methods.

Identifying a candidate world

Through this process, Rabtsevich narrowed the list of candidates down to a specific star known as TIC 4206066. This star is located roughly 116 light-years from Earth and is slightly smaller than our own Sun. The data showed an unmistakable, slight dip in the star's brightness every 3.18 Earth days. To verify this, he pulled older data from 2018 and 2020 and found the same dips occurring at the same frequency.

Based on these findings, Rabtsevich calculated that the potential planet would be just over 1.4 times larger than Earth. Because its full solar year lasts only about three Earth days, he predicts the surface temperature would be around 1,000 degrees Fahrenheit. This would make it significantly hotter than Mercury, the planet closest to our Sun, and hot enough to melt lead.

The AI-assisted search also revealed a second dimming pattern occurring every 11.13 Earth days. While this could suggest a second previously unknown planet, Rabtsevich considers the evidence for that specific finding to be less convincing than the first. He eventually ran a total of seventy-four tests with Claude Code to try and disprove his primary hypothesis before moving forward.

Validating the AI's logic

To test the reliability of the AI, Rabtsevich conducted a specific experiment using two of the TESS datasets. He asked the agent to predict exactly when the dimming would occur in a third, separate dataset. The AI's predictions were correct in every single one of the three cases. This level of accuracy gave him the confidence to submit a formal report to the researchers at MIT.

The report requested that TESS be positioned to observe TIC-4206066 the next time the telescope was scheduled to focus on that specific sector of space. The request was approved, and the telescope will now measure the star's brightness every two minutes from October 31 to November 26. Rabtsevich has already publicly shared the specific times the star should dim during this window to see if the planet appears.

As gizmodo.com reported, Rabtsevich noted that without these tools, he would not have been able to process a dataset of this size. He explained that his previous work involved building endless spreadsheets and grinding through data, which consumed a massive amount of time. With AI, he was able to run dozens of tests on a single star in about two weeks, a task that would have been much slower manually.

The role of AI in science

The discovery of exoplanets is a massive undertaking, with astronomers already cataloging more than 6,400 worlds. However, observing subtle changes in light across vast distances is prone to errors, such as mistaking a binary star system for a planet. Kevin Hardegree-Ullman, a research scientist at the NASA Exoplanet Science Institute, notes that the biggest challenge is ruling out these false-positive scenarios while competing for limited telescope time.

AI has a superhuman ability to parse through galactic quantities of data, which makes it a powerful tool for detecting candidates. NASA actually deployed an AI model slightly less than a year ago to comb through TESS data, and it has already identified around 7,000 exoplanets. Because tools like Claude Code and Codex are publicly available, they could democratize the discovery process for amateur astronomers who lack professional resources.

Hardegree-Ullman suggests that while AI might not be the primary tool for confirming a planet's existence, it will likely continue to be used to string together publicly available code. These tools help researchers search for planets and run them through basic vetting processes. This allows individuals like Rabtsevich to contribute to the scientific community by identifying targets that professional teams can then investigate more deeply.

What comes next for TIC-4206066

The upcoming observation period in November will be the ultimate test for Rabtsevich's hypothesis. If the star dims exactly when predicted, the evidence for the planet becomes much more serious. If it does not, the data will likely suggest that the dimming was caused by something else, such as a binary star system or other celestial interference.

Rabtsevich is being careful to emphasize that correlation does not necessarily equal causation. He acknowledges that the pattern detected in his AI-assisted investigation might turn out to be rooted in something other than a planet. Regardless of the outcome, he has already seen his findings catalyze interest among other amateur astronomers who are now beginning to search TESS data themselves.

He hopes that this new wave of enthusiasm comes with careful checks to ensure that professional astronomers are not overwhelmed by poorly vetted signals. For now, the scientific community is watching the data from TIC-4206066 closely. This story serves as a notable example of how AI can move beyond