A team of astronomers led by Joshua Roth, a graduate researcher at Princeton University, has used artificial intelligence to sift through more than 80 million brightness recordings from NASA's TESS space telescope, turning up over 10,000 new candidate planets circling other stars. It is one of the largest single batches of new planet candidates ever reported at once, and it points to how central AI has quietly become in modern astronomy. The findings, described in a paper titled "The T16 Planet Hunt: 10,000 New Planet Candidates from TESS Cycle 1," were reported this week and have been accepted for publication in The Astrophysical Journal Supplement Series. The work involved researchers from Princeton, MIT, UCLA, and the Las Campanas Observatory in Chile.

How the AI did the heavy lifting

TESS (the Transiting Exoplanet Survey Satellite) hunts for planets by watching for the tiny, repeating dip in a star's brightness that happens when a planet passes in front of it, an event astronomers call a transit. Over its first observing cycle alone, the telescope generated more than 83 million of these brightness recordings, known as light curves — far too many for any team of humans to check one at a time.

Roth's team built a two-stage AI pipeline to handle the flood of data. First, an algorithm called CETRA (the Cambridge Exoplanet Transit Recovery Algorithm), designed to run quickly on graphics processors, scanned every light curve for the faint, periodic dimming pattern that signals a transit. Notably, CETRA was sensitive enough to catch signals from stars fainter than those covered by TESS's official detection pipeline. Next, a machine-learning classifier called Random Forest ranked and filtered the results, narrowing the pool from roughly 80 million light curves down to about 2.5 million that were worth a closer look by the team.

What they found

From that shortlist, the researchers identified 11,554 total transit candidates, of which 10,091 are newly discovered (the rest had already been catalogued by earlier surveys). The haul leans toward objects that are relatively easy to detect: researchers say the bulk are so-called "hot Jupiters," gas giants that orbit very close to their host stars, alongside more than 100 Neptune-sized worlds and 11 candidate super-Earths — smaller, rocky planets that are of particular interest to scientists searching for potentially habitable worlds elsewhere in the galaxy.

As a check that the whole pipeline actually works, the team followed up on one candidate with ground-based observations: using the PFS spectrograph on the Magellan telescope in Chile, they measured the subtle gravitational wobble of a star and confirmed a hot Jupiter orbiting TIC 183374187, an unusual metal-poor star located in the Milky Way's thick disk.

Why AI made the difference

Roth put a concrete number on how much time the AI approach saved. He said that manually vetting 50,000 of the candidates by eye, checking each light curve one by one, took his team about six weeks of focused work. Without the machine-learning filtering step, he estimated the team would have needed to manually check roughly 32 times as many light curves overall — a task that, at the same pace, would have taken about 192 weeks, or well over three and a half years. Instead, the AI-assisted pipeline compressed that workload dramatically, letting a small team accomplish in a reasonable stretch of time what would otherwise have demanded a small army of human reviewers working around the clock.

Caveats and what comes next

It is worth stressing that "candidate" is not the same as "confirmed planet." Transit-like dimming can sometimes be mimicked by other phenomena, such as pairs of stars eclipsing each other, so a substantial share of the new candidates will need follow-up observations, additional transits, radial-velocity measurements, or direct imaging, before they can be confirmed as genuine worlds. Astronomers describe this AI-driven sorting as a first pass that flags the most promising targets for that harder, slower confirmation work.

Even with those caveats, researchers say the approach is quickly becoming essential rather than optional. NASA's Roman Space Telescope, expected to launch in August 2026, is projected to detect roughly 100,000 transiting planet candidates on its own — a scale of data that would be unmanageable through manual review alone. Roth's team says it plans to add further AI tools, including convolutional neural networks, to sharpen the search in future TESS observing cycles, suggesting that AI-assisted triage is set to become a standard part of how astronomers comb the sky for new worlds.