How AI is widening the search for alien technology

A telescope observation can have more than one scientific life. An archive searched for one kind of signal can later be examined for another; a survey of changing stars can supply leads for researchers hunting signs of extraterrestrial technology. AI is helping those questions travel between datasets.[1]

Breakthrough Listen astrophysicist Steve Croft’s review, posted on 3 October, maps established radio analyses, optical anomaly detection and early language-model experiments. It reports no confirmed technosignature—a sign of technology beyond Earth—and distinguishes deployed methods from prototypes.[1]

A signal with somewhere to be

One revealing example is a study published in Nature Astronomy in 2023. Peter Ma and colleagues searched more than 480 hours of Green Bank Telescope observations covering 820 stars. After machine filtering and human inspection, they identified eight signals of interest from five stars that earlier analyses had missed. Reobservations did not recover similar signals, leaving them unconfirmed.[2]

The learning task was quite concrete. Training examples placed simulated radio signals inside real telescope backgrounds. A promising pattern appeared when the dish pointed at a target and disappeared when it looked elsewhere. The system compressed these examples into a learned representation, then used a separate classifier to distinguish that pattern from interference.[2]

This was still a substantial sorting job: the researchers visually assessed 20,515 flagged events before choosing the eight. The AI supplied a different route through existing measurements; it did not replace the observations needed to decide what a signal was.[2]

Learning what ordinary explosions look like

Optical searches offer another route. In February’s Anomaly Hunter for Alerts preprint, Leyla Iskandarli and colleagues used three neural networks to examine Zwicky Transient Facility alerts: one read measured object features, one inspected image cutouts and one followed brightness over time. Each learnt to reconstruct predominantly ordinary supernova examples; a poor reconstruction made an object worth examining.[3]

Across alerts from 1–25 December 2025, the team selected 87 high-priority, unusual supernova candidates for spectroscopic follow-up. The different representations selected substantially different objects. Croft discusses how such complementary anomaly searches could support future technosignature work.[3][1]

The distinction is useful. A detector taught a particular radio pattern asks whether that pattern is present. A reconstruction model asks where an observation departs from its learnt examples. Neither question covers everything, but together they give a search more opportunities to notice an exception.

The observation also has to survive

Archives can be searched again; discarded raw data cannot. Croft describes fast models examining incoming streams, with more expensive analysis reserved for selected events. Transferring optical methods to Rubin Observatory at scale remains work to be done.[1]

The computing behind that idea is becoming tangible. Stelline’s current project documentation describes telescope packets travelling directly from a network interface into GPU memory, with processed astronomy files written to fast storage. Its supplied examples include a live spectrum display, beamforming and a correlator, targeting the Allen Telescope Array. Keeping data moving through the processors makes room for analysis while observations are arriving.[4]

The resulting opportunity is wider than an automatic “alien detector”. More kinds of observations can become starting points for a search, and more searches can share the same observation. An unusual result may lead to an instrumental explanation, an unfamiliar natural object or a harder question. The scientific gain begins when something worth investigating becomes visible.

Report an error or suggest a correction