Between 2018 and 2019, NASA’s Transiting Exoplanet Survey Satellite photographed more than 56 million stars during its first sweep of the sky. The images were downloaded, archived, and made public. Then, for the most part, they sat there. Standard planet-hunting software skimmed the data for obvious signals and moved on. Nobody had the time or the tools to dig deeper.
Now a research team led by astronomer Megan Bedell at the Flatiron Institute’s Center for Computational Astrophysics, working with collaborators across several institutions, has turned machine learning loose on that same archive and pulled out roughly 10,000 planet candidates that every previous search missed. The results, described in a preprint posted in May 2026, represent one of the largest single batches of new exoplanet candidates ever reported. The signals were hiding in plain sight, buried in publicly available data that any astronomer on Earth could have downloaded at any time over the past seven years.
What the search actually turned up
The project, called the T16 Planet Hunt, ran a semi-automated transit search across 83,717,159 light curves extracted from TESS Cycle 1 full-frame images. That number is larger than the 56,401,549 stars in the underlying catalog because many stars were observed in more than one TESS sky sector, producing a separate light curve for each sector. Each light curve tracks the brightness of a single star over a given observing window; a brief, repeating dip can indicate a planet crossing the star’s face. The search flagged 11,554 planet candidates in total. Of those, 10,091 are entirely new, 1,052 match candidates already in existing catalogs, and 411 are single-transit events where the telescope caught only one dimming episode, hinting at planets on longer, slower orbits.
The raw data behind the search is massive and independently verifiable. A companion paper describes the construction of the T16 light-curve dataset, which covers 56,401,549 stars observed during TESS Cycle 1. The processing pipeline used image subtraction and detrending techniques to squeeze usable brightness measurements out of full-frame images that TESS captured at 30-minute intervals. Those full-frame images are standard TESS data products, documented in NASA’s own instrument handbook, and they record wide swaths of sky rather than individual pre-selected targets. That breadth is exactly what made the new search possible: millions of stars were photographed as a matter of routine, but most were never individually scrutinized for planets.
The team also confirmed at least one planet outright: a hot Jupiter orbiting the star cataloged as TIC 183374187. Its short orbital period and deep, repeatable transits made it an ideal proof of concept, demonstrating that the pipeline can move from flagging candidates to producing validated discoveries when follow-up data supports them.
Why these signals went unnoticed
Standard planet-detection pipelines are built for efficiency. They scan enormous datasets quickly, but they are tuned to catch clean, textbook-shaped transit signals: symmetric dips that repeat at regular intervals with consistent depth. Many of the T16 candidates produced signals that broke those rules. Some showed only a handful of shallow dips buried in the natural flickering of their host stars. Others had asymmetric or V-shaped profiles that resemble grazing eclipses or blended light from nearby stellar systems. Traditional software flagged these as noise or ambiguous and moved on.
Machine learning handles ambiguity differently. Trained on large sets of known transits and known false positives, the algorithm learned to recognize subtler patterns that rigid threshold-based searches discard. That does not mean every flagged signal is a real planet. It means the net was cast wider, and what it hauled in now needs sorting.
This is not the first time AI has pulled confirmed worlds out of old telescope archives. A separate NASA effort used a deep-learning system called ExoMiner to re-analyze data from the earlier Kepler mission. That algorithm distinguished planetary transits from common impostors and ultimately added 301 confirmed planets to the Kepler catalog. The precedent matters: it showed that machine learning can accelerate established vetting processes, not replace them. ExoMiner did not bypass traditional checks. It made them faster.
What ‘candidate’ actually means
In exoplanet science, the gap between “candidate” and “confirmed” is wide and consequential. NASA draws a clear line between confirmed planets and unconfirmed candidates. A candidate is a signal that looks like a planet transit but has not yet survived independent observations or statistical tests rigorous enough to rule out false positives: eclipsing binary stars, instrumental glitches, background contamination. All 10,091 new T16 signals sit on the candidate side of that line. Only the single hot Jupiter around TIC 183374187 has been promoted to confirmed status so far.
The headline framing of these candidates as “impossible” planets is editorial shorthand, not a term used in the preprint itself. It reflects the unusual nature of many signals in the catalog: transits so faint, so brief, or so oddly shaped that standard automated searches passed over them. Whether they are genuinely planets, instrumental artifacts, or astrophysical false positives will only become clear through follow-up spectroscopy and additional transit observations. The T16 preprint does not claim all 10,091 signals are real planets. It presents them as candidates worthy of further scrutiny.
Comparing the T16 results directly to ExoMiner’s Kepler work is tempting but imprecise. The two projects used different instruments (TESS versus Kepler), different detection methods, and different definitions of what counts as a starting candidate. ExoMiner confirmed 301 planets from an already-vetted pool of Kepler candidates; the T16 search generated its candidate list from scratch using raw light curves. A similar confirmation rate is not guaranteed, and the true yield of real planets from the T16 catalog could be higher or lower. Both outcomes are plausible, and the answer will only emerge through independent follow-up work.
The 411 single-transit events pose a particular challenge. With only one observed dimming, astronomers cannot determine orbital periods, which means they cannot predict when the next transit will happen or schedule targeted follow-up efficiently. These signals could represent some of the most scientifically valuable finds in the catalog, potentially planets on wide orbits in or near habitable zones, but they are also the hardest to confirm. In many cases, verification would require long-term radial-velocity monitoring or waiting for a future space telescope to catch another transit.
What happens next
No public timeline exists for systematic follow-up on this catalog. The NASA Exoplanet Archive, operated by IPAC at Caltech, serves as the official ledger for confirmed exoplanets and tracks candidates separately. A new catalog of this size would need to be cross-matched against existing records before any bulk validation could begin, and that process depends on telescope time, funding, and community interest, all of which are finite.
The preprint has not yet passed formal peer review as of June 2026. That means the methodology, false-positive rates, and candidate classifications have not been independently scrutinized by journal referees. Preprints on arXiv are standard practice in astrophysics and often precede publication by only a few months, but readers should treat the specific numbers as provisional until a peer-reviewed version appears. Review could uncover biases in the training data, problems with how the pipeline handles crowded stellar fields, or miscalibrated confidence thresholds.
For researchers and citizen scientists, the practical next step is straightforward. The T16 data release is publicly available through NASA’s Mikulski Archive for Space Telescopes, complete with a DOI, provenance records, and bulk-download instructions. Independent teams can re-analyze the light curves with alternative pipelines, apply their own transit-search algorithms, or zero in on subsets of stars where small, rocky planets are easier to detect. Others might focus on vetting specific categories: ultra-short-period candidates, signals around bright stars suited for spectroscopic follow-up, or the tantalizing single-transit events that could point to worlds in temperate orbits.
10,000 places to look next
The T16 Planet Hunt did not discover 10,000 new worlds. What it produced is something almost as valuable: a map of 10,000 places where worlds might be hiding, drawn from data that was already in hand but never fully explored. Turning those candidates into confirmed planets will take years of additional observations, careful modeling of stellar behavior, and the slow, unglamorous work of ruling out every alternative explanation. But the starting point has shifted. Seven years of overlooked TESS photographs now have 10,091 new reasons to be looked at again.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
