Somewhere in the southern sky, a giant planet roughly the size of Jupiter whips around a faint star every few days, so close that its atmosphere is likely scorched beyond recognition. It had been hiding in data NASA already owned. Nobody noticed until a team of astronomers turned machine-learning algorithms loose on 80 million stars that the space agency’s standard software had essentially ignored.
That planet, now confirmed through ground-based telescope observations, is just the first catch from a sweeping new search called the T16 project. In a preprint posted in early 2025, the team reported 11,554 planet candidates extracted from archival data collected by NASA’s Transiting Exoplanet Survey Satellite (TESS), with 10,052 of those never flagged by any previous search. To put that number in perspective, NASA’s total count of confirmed exoplanets across all missions and methods currently sits just above 5,800. If even a small fraction of these new candidates survive follow-up scrutiny, the haul could meaningfully reshape the known catalog of worlds beyond our solar system.
The findings have not yet undergone formal peer review, and “candidate” is a long way from “confirmed planet.” But the work demonstrates something that planetary scientists have suspected for years: existing space telescopes have already recorded far more discoveries than anyone has had time to find.
Why 80 million stars went unexamined
TESS photographs enormous swaths of the sky using Full Frame Images, capturing light from millions of stars in a single exposure. The mission’s main data pipelines then generate detailed brightness measurements, called light curves, for individual stars. Astronomers comb those light curves for the tiny, periodic dips that occur when a planet crosses in front of its host star.
But those pipelines have a practical ceiling. According to the T16 preprint, the MIT Quick-Look Pipeline, one of the primary tools for processing TESS Full Frame Images, produces light curves only for stars brighter than roughly TESS magnitude 13.5. Stars fainter than that cutoff, stretching down to magnitude 16, number in the tens of millions. They were photographed by TESS but never systematically searched for planets.
The T16 team closed that gap using image subtraction, a technique that isolates changes in brightness by digitally removing a reference image from each new frame. Applied to TESS Cycle 1 data, the method generated light curves for stars as faint as magnitude 16. The team validated the approach by recovering transit signals from already-known planets, confirming the pipeline could reliably pick up the subtle brightness dips that betray an orbiting world. The resulting light curve files are publicly available through NASA’s Mikulski Archive for Space Telescopes (MAST).
What the AI search turned up
With millions of new light curves in hand, the team ran a semi-automated, machine-learning-based transit detection system to sift through the data. The search produced 11,554 planet candidates with orbital periods ranging from about half a day to 27 days. Of those, 10,052 had never appeared on any earlier candidate list.
The single confirmation so far is a hot Jupiter orbiting the star TIC 183374187. According to the team’s preprint, radial velocity measurements taken with the Planet Finder Spectrograph on the Magellan 6.5-meter telescope in Chile detected the gravitational wobble the planet induces in its host star, and the measured orbital period matched the TESS transit signal. That agreement, obtained through a standard and well-understood verification method, demonstrates that the T16 pipeline can produce genuine planetary detections, not just noise or stellar impostors.
Machine learning has pulled hidden worlds out of space telescope archives before. In 2017, a deep-learning model trained by Google identified two planets, Kepler-90i and Kepler-80g, that earlier searches of NASA’s Kepler mission data had missed. That effort used a neural network applied to Kepler light curves, a different approach from the T16 project’s combination of image subtraction and machine-learning classification on TESS Full Frame Images. Still, both cases share a core insight: algorithms can recover weak transit signals that traditional searches overlooked. The T16 results apply that insight to a far larger and fainter stellar sample, effectively opening a new slice of the sky for discovery without launching a new mission.
The long road from candidate to confirmed planet
In exoplanet science, the gap between “candidate” and “confirmed” is wide and unforgiving. NASA defines a candidate as a star showing brightness dips consistent with a transiting planet. Confirmation requires independent verification, typically through radial velocity measurements, transit timing variations, or high-resolution imaging that rules out other explanations like eclipsing binary stars or background blends.
Of the 10,052 new candidates, only one has crossed that threshold. The rate at which the rest will survive scrutiny is unknown. False positives from binary stars, instrumental artifacts, and intrinsic stellar variability routinely thin candidate lists during follow-up campaigns. A meaningful reliability estimate will require a statistically representative follow-up sample, testing both strong and marginal candidates to measure how often the machine-learning system is fooled.
NASA has not publicly stated whether the T16 candidates will be folded into the mission’s official catalog or prioritized for follow-up through the ExoFOP portal, the coordination site where astronomers organize observations of promising targets. Absorbing thousands of additional faint-star candidates would require significant triage. Observers must weigh the scientific payoff of each potential discovery against the risk that a given signal will turn out to be a false alarm, all while competing for scarce telescope time.
Scaling up and what comes next
The T16 preprint describes plans to extend the analysis beyond TESS Cycle 1, but no data products for later cycles have been released as of June 2025. Whether the team can scale its approach to the growing volume of TESS observations, which now span multiple years of sky coverage, is an open technical question. Running image subtraction and machine-learning classification across tens of millions of additional faint stars demands serious computational resources. Changes in detector performance, spacecraft pointing, and background noise between observing cycles could also complicate the task of training models that behave consistently over time.
Then there is the vetting bottleneck. Traditional pipelines rely on human experts to inspect light curves, check for telltale signs of false positives, and examine centroid shifts that might indicate a background binary masquerading as a planet. Applying that level of scrutiny to more than 11,000 new signals is not realistic without new automated tools. The T16 team’s classifiers provide a first-pass filter, but independent groups may want to retrain or recalibrate those models using their own criteria, potentially revising the reliability estimates.
Both preprints and the publicly archived dataset are accessible through MAST, which means any researcher can independently check, extend, or challenge the findings. If other groups reproduce the candidate list using different algorithms, confidence in the underlying detections will grow. For now, the T16 results represent a parallel, community-driven effort rather than an update to TESS’s core pipeline. Official mission catalogs carry institutional vetting that independent projects, however rigorous, do not automatically receive.
Why faint stars could hold the biggest surprises
Beyond the raw candidate count, the T16 project’s most lasting contribution may be methodological. By pushing TESS data analysis to fainter magnitudes, the search samples a different stellar population. Many of those dim stars are low-mass red dwarfs, whose small physical sizes make it easier to spot the transit signatures of smaller, potentially Earth-sized planets. The same dip in brightness that a tiny planet produces when crossing a compact star would be undetectable around a larger, brighter one.
The work also illustrates a broader principle: existing space missions can keep yielding new science long after their primary pipelines mature, simply by reprocessing the same photons with more sophisticated tools. Whether most of the 10,052 new candidates survive the gauntlet of follow-up observations or quietly fade from the record, the experiment offers a template. As TESS continues to scan the sky and future surveys like the European Space Agency’s PLATO mission come online, the ability to mine archival data with machine learning will only become more valuable. The planets, after all, were always there. The data recorded them. It just took a different kind of eye to see them.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
