A star dimmed by five hundredths of a percent. An amateur and AI found a planet-sized question

The interesting thing about TIC 4206066 is not that somebody has found a planet there. Nobody has established that. The interesting thing is that a recurring weakness in a star's light—about five hundredths of one percent—survived enough checking to justify another look. The researcher who surfaced it was not running an observatory. Pavel Rabtsevich, a Spain-based independent astronomy enthusiast, used AI-assisted code and public NASA TESS records to search an immense collection of stellar light curves.[1][2]

Reported in Scientific American in October, the search involved more than 126,000 target stars. One showed a small, transit-like dimming approximately every 3.18 days. A planet crossing the star could block just such a fraction of its light. So can less romantic things: blended light from another star, stellar variability, detector systematics or an eclipsing companion. The difficult part of exoplanet science begins precisely when a pattern looks persuasive.[1][2]

The significance of 0.05%

Imagine the star's recorded brightness as a steady one-million-unit signal. A dip of five hundred parts per million corresponds to roughly 500 of those units. That is small enough for calibration and noise to matter, yet if the same shape recurs on a reliable schedule, it can be investigated statistically. The period and dip depth are observations or fitted estimates; the proposed size of a potential planet depends on assumptions about the host star and what fraction of its light comes from the target.[1][2]

Public TESS data are a remarkably democratic raw material. A satellite may have already collected the light but there are many stars, many sectors of observations and multiple ways of cleaning imperfect signals. An AI-assisted coding workflow can make searching and comparing possibilities easier for a person who is not in a large professional team. It cannot turn a false positive into a planet by enthusiasm alone.[1][2]

A claim with a way to lose

The most appealing feature of this case is the emphasis on audit and prediction. A published independent examination of the candidate labels it explicitly unconfirmed and not peer-reviewed. It lays out the known periods, preprocessing choices, competing stellar explanations and ways in which future observations might invalidate the signal. TESS is expected to revisit the target later in 2026, offering the prospect of data outside the original search window.[1][2]

That is a more demanding test than checking whether another algorithm can reproduce a pattern in the same already-inspected frames. Once a researcher has selected an unusual star from more than a hundred thousand targets, post-selection bias becomes a central issue: chance patterns are bound to appear somewhere. A fresh observation at the predicted transit time—and an independent assessment of whether the dimming comes from the target rather than a neighbouring object—provides much stronger evidence.[1][2]

A weaker secondary signal has also been discussed. That is not grounds for declaring a multi-planet system. It is a tentative feature within a candidate analysis, with additional possible instrumental or astrophysical explanations.[1][2]

The new division of labour

It is tempting to say that AI discovered an exoplanet. The defensible statement is more interesting: AI-supported analysis helped an individual find a scientifically testable lead in a vast public archive. The human researcher chose the question, examined the candidate and put it before others who can test it. The planet, if it exists, does not care who wrote the search code.[1][2]

This is what a serious version of AI-accelerated amateur science can look like: accessible data, reproducible methods, public disagreement and a prediction that has not happened yet. In November, the crucial development will not be a louder headline. It will be what the telescope records.[1][2]

Report an error or suggest a correction