The data had been sitting on NASA’s servers for years. Thousands of stars, observed by the Transiting Exoplanet Survey Satellite during its first sweep of the sky, had already been picked over by multiple research teams. The obvious planets had been found. The easy signals had been cataloged. And then a group of astronomers ran the entire dataset through a machine-learning pipeline and pulled out more than 10,000 planet candidates that nobody had noticed before.
The effort, described in a study titled “The T16 Planet Hunt” and posted to arXiv in 2025, processed 83,717,159 light curves from TESS Cycle 1 full-frame images. The search flagged 11,554 transit-like signals, objects whose host stars periodically dimmed in a pattern consistent with a planet crossing in front of them. Of those, 10,091 had never appeared in any previous catalog. To put that in perspective: as of early 2026, the total number of confirmed exoplanets across all missions and methods stood at roughly 7,000.
What the search actually found
The candidates span orbital periods between half a day and 27 days, meaning they orbit close to their host stars. That range covers scorching-hot gas giants, lava worlds, and potentially rocky planets baking in tight orbits. The team also flagged 411 single-transit events, cases where only one dip appeared during the observation window, hinting at planets on longer orbits that crossed their star just once while TESS was watching.
To validate that the pipeline could find real planets and not just noise, the researchers confirmed a hot Jupiter orbiting the star TIC 183374187. That confirmation used traditional follow-up techniques, including radial-velocity measurements that detected the gravitational tug of the planet on its star. One confirmed planet out of 11,000-plus candidates is a small fraction, but it demonstrates the pipeline works end to end, from raw photometry to a verified world.
The light curves themselves came from the publicly released T16 dataset, which covers 56,401,549 stars brighter than magnitude T=16. Building that dataset required image subtraction, a technique that strips away a reference image of each star field to isolate faint brightness changes. The method suppresses noise from crowded fields and instrumental quirks, making it easier to spot the tiny, periodic dimmings that betray a transiting planet. Both the raw and cleaned photometry were made public, giving any research group the ability to run independent searches on the same data.
Why AI made the difference
TESS watches nearly the entire sky and generates far more light curves than any team of humans could review by hand. Kepler, its predecessor, monitored about 150,000 stars in a single patch of sky. TESS Cycle 1 alone produced light curves for more than 56 million. At that scale, automated classification is not a convenience; it is the only way to keep up.
NASA has been building toward this moment for years. Its open-source tools ExoMiner and ExoMiner++ were designed specifically to triage transit signals at scale. Applied to older Kepler data, deep-learning models confirmed 370 exoplanets that had lingered as unverified candidates, proving that AI could reliably distinguish real planetary transits from lookalikes such as eclipsing binary stars and instrumental glitches. The T16 Planet Hunt team built their own machine-learning pipeline for the TESS search. How much it borrows from ExoMiner’s architecture versus developing its own approach is not fully detailed in the available materials, but the underlying principle is the same: train a model on known transits and known false positives, then turn it loose on millions of unexamined light curves.
The result is a kind of archaeological dig through data that already existed. No new telescope time was needed. No new spacecraft was launched. The discoveries were always there, encoded in brightness measurements that had been publicly available for years. What changed was the tool used to look.
The long road from candidate to confirmed planet
A transit candidate is not a confirmed planet. That distinction matters enormously. Eclipsing binary stars, background blends where two stars overlap along the line of sight, and systematic instrumental artifacts can all mimic the light-curve dip of a transiting planet. In previous large-scale transit surveys, raw candidate lists have historically seen false-positive rates of 30 to 50 percent, sometimes higher in crowded fields near the galactic plane.
The T16 study does not provide a publicly summarized false-positive rate or purity estimate in its abstract, so the community cannot yet gauge how clean this particular catalog is. That number will be critical. If the list runs at a 40 percent false-positive rate, roughly 6,900 candidates might be real. If it runs higher, the yield drops accordingly. Either way, even a conservative survival rate would add hundreds or potentially thousands of new planets to the confirmed tally over time.
Confirmation requires follow-up, and follow-up requires telescope time. Radial-velocity spectrographs on instruments like ESPRESSO at the Very Large Telescope or NEID at the WIYN Observatory can measure the stellar wobble induced by an orbiting planet, nailing down its mass. For atmospheric characterization, the James Webb Space Telescope is the current gold standard, but JWST time is fiercely competitive and limited. Astronomers will almost certainly prioritize candidates orbiting bright, nearby stars, where measurements are most precise and atmospheric studies are feasible. Fainter, more distant candidates may wait years for their turn.
The 411 single-transit events face an additional hurdle. Without repeated dips, their orbital periods are unknown, making it difficult to predict when the next transit will occur and schedule follow-up observations. Some of these could turn out to be planets on orbits of months or years, potentially placing them in more temperate regions around their stars. But confirming them will require patience and, in many cases, additional space-based photometry from future TESS observing cycles.
What this catalog could reshape
If a substantial fraction of these candidates hold up, the implications ripple across exoplanet science. Demographic studies, which ask questions like “how common are hot Jupiters?” or “what fraction of Sun-like stars host small, rocky planets?”, depend on large, well-characterized samples. Adding thousands of new planets from a single TESS cycle would sharpen those statistics considerably, especially for short-period planets that are easiest to detect via transits.
The study also carries a methodological message. TESS has completed multiple observing cycles since Cycle 1, and each cycle generates a comparable volume of data. If the same pipeline is applied to Cycles 2 through 6, the candidate count could multiply further. Other archival datasets, from Kepler’s extended K2 mission, from ground-based surveys like WASP and HATNet, may also contain overlooked signals waiting for a similar computational pass. The era of discovery-by-reanalysis is not a one-off event; it is becoming a sustained mode of doing science.
The T16 Planet Hunt paper is currently hosted on arXiv and may not yet have completed formal peer review. That is standard practice in astrophysics, where preprints routinely circulate for months before journal acceptance, but it means the methods, training data, and candidate classifications have not yet been independently vetted through the journal process. Peer reviewers will likely probe the robustness of the detrending algorithms, the composition of the training set, and the statistical thresholds used to declare a candidate. Until that review is complete, the numbers should be treated as preliminary.
Old data, new eyes
What makes this story striking is not just the number of candidates. It is the reminder that discovery does not always require building a bigger telescope or launching a new mission. Sometimes the planets are already in the data, hiding in plain sight, waiting for an algorithm sharp enough to find them. The TESS archive is vast, public, and growing. The tools to mine it are improving faster than the data itself accumulates. And somewhere in those 83 million light curves, confirmed worlds are almost certainly still waiting to be recognized for what they are.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
