What AI Already Does in SETI
Machine learning helps astronomers sort telescope observations, recognise interference and rank unusual signals for investigation. It works on measurements such as signal strength, frequency and time. These are detection tasks. They do not establish who transmitted a signal, whether it contains a message or what that message means.
In radio searches, a spectrogram maps how power at different frequencies changes over time. Researchers can train a model with simulated signals inserted into real observations, or use clustering to group similar detections. An outlier earns closer inspection, not an extraterrestrial identity.
- 1. ObserveRecord the target and comparison directions.
- 2. FilterUse algorithms to rank signals and flag interference.
- 3. VerifyInvestigate the source with further observations and independent teams.
- 4. DecodeIf information is present, test ways to recover its structure.
- 5. InterpretTest what that structure could mean using independent evidence.
The last two stages are conditional. A verified transmitter would not automatically provide a readable message.
What the Studies Actually Found
These projects used different instruments, datasets and methods. Their results cannot be combined into a single AI accuracy score.
| Study | Data and Method | Result and Limit |
|---|---|---|
| Ma et al., 2023 | Deep learning searched 150 TB of Green Bank Telescope data covering 820 nearby stars. | Eight signals of interest. Follow-up observations did not detect them again, so they did not qualify as confirmed technosignatures. |
| Pardo et al., 2025 | Anomaly detection ranked roughly 100 billion spectrograms from Green Bank and Parkes observations. | About 20,000 examples were inspected, including validation samples. None survived scrutiny as a viable extraterrestrial candidate. |
| Zhao et al., 2026 | DBSCAN clustering filtered residual interference from five hours of FAST observations recorded in July 2019. | One candidate group remained after additional checks. Its roughly 2.8-second duration was insufficient to resolve its origin. |
The 2023 result is sometimes retold as AI discovering eight alien signals. Breakthrough Listen's own project page says the signals failed its candidate criteria because they were not detected again. The achievement was a different way to select unusual observations from a large archive.
Read the Breakthrough Listen follow-up statement and access its code and data.
The FAST paper, published in January 2026, does not announce contact. After checking frequency bands and signal durations, the authors retained one short event with no observed frequency drift. Radio interference remained unresolved; they proposed longer observations. Keeping an unexplained event on a follow-up list is a normal research outcome.
Why AI Can Miss or Misclassify a Signal
Radio-frequency interference, or RFI, comes from human technology. A narrow signal or a changing frequency can attract attention, but satellites and other terrestrial sources can also produce structured detections. Models must work within that contaminated environment.
AI also inherits choices made by researchers. Selecting narrow signals, simulating certain shapes, excluding noisy frequencies or setting a detection threshold determines what a search can find. An algorithm does not become neutral because it processes more data.
- False positives. Unfamiliar interference can look unusual enough to reach the shortlist.
- False negatives. A real signal outside the model's tested range could be discarded or never detected.
- Limited coverage. A search samples particular directions, frequencies and times. No detection in that sample does not establish that the Galaxy is empty.
Pardo and colleagues explicitly describe the trade-off. Prioritising quieter parts of their data reduced the human workload, but could reject a genuine signal in a noisy region. They also acknowledge that simulations reflect the kinds of signals researchers choose to create.
Could AI Translate an Alien Message?
There is no demonstrated alien-language translator in these studies. Detecting a possible transmission, recovering an encoding and identifying its meaning are separate problems. A repeating pattern might help test an encoding hypothesis; it would not, by itself, tell us whether the contents describe chemistry, navigation or something else.
AI could help researchers compare possible structures or generate hypotheses. A convincing interpretation would still need evidence beyond the model's fluent explanation. A system trained on human text has no verified alien dictionary against which to check its answer. Mathematical regularities could offer clues, but they do not supply an automatic translation.
A real exercise illustrates the gap. In May 2023, Daniela de Paulis's A Sign in Space project transmitted a human-designed, simulated extraterrestrial message from ESA's ExoMars Trace Gas Orbiter. Ken and Keli Chaffin decoded it in June 2024, recovering an image containing five amino acids. ESA's October 2024 account still separated that decoding achievement from interpreting the message's meaning.
Read ESA's account of the decoding exercise.
This was an art and science project with a message created on Earth. It was not an alien detection, and the reported solution was the work of a father-daughter team, not proof that AI can translate an unknown civilisation.
What Would Count as Confirmation?
The International Academy of Astronautics updated its SETI principles in June 2026. They call for efforts to authenticate a candidate, ideally involving multiple facilities, organisations and methods, and for a peer-reviewed verification report supported by the underlying data. These are scientific guidelines, not a guarantee that any candidate will be resolved.
Read the June 2026 detection and verification principles.
When a new AI discovery headline appears, check three things before treating it as evidence of contact.
- Does the original paper report a detection, a provisional candidate or only a method tested on simulations?
- Were interference and instrument effects investigated, and did independent observations support an extraterrestrial origin?
- Is there actual information to decode, or only a signal that the search software selected?
AI can make the search more efficient and expose observations worth revisiting. It cannot turn an anomaly into a verified message by interpretation alone. The wider question of why searches remain inconclusive is part of the Fermi paradox.
