For more than six decades astronomers have searched through billions of narrowband radio channels, struggling to find any sign of intelligence. And adding to the challenge of listening at the right time, in the right place, there’s myriad confounding signals all over Earth, from cellphones to GPS, wi-fi to microwave ovens, which wash over a radio telescope and interfere with its strained hearing of the heavens. Even though in principle all that radio junk from human activity can be sifted through and sorted, the sheer amount that could obscure any true alien signal means the odds are really against us. Especially as we enter the era of ‘Big Data’ astronomy.

It’s becoming less of a job for a human, and more of a job for a machine.

A new machine-learning algorithm, written by an undergraduate student at the University of Toronto, Peter Ma, has cut through the terrestrial noise to uncover eight currently unexplained radio signals, each with some hallmark of bonafide extraterrestrial chatter.

Ma first fed his algorithm with simulated signals, training it to recognize what we think an alien radio signal might look like — one that exhibits Doppler drift incurred by the rotation of the transmitter and the receiver on rotating planets, and which also displays a clear on-off pattern when the telescope is moved away from the target star. If a signal really is just terrestrial interference, then it shouldn’t exhibit Doppler drift, and in principle should disappear when the telescope is moved away from the target.

Then, he unleashed it on 480-hours’ worth of data collected from 820 star-systems by West Virginia’s Green Bank 100-meter radio telescope. This amounted to millions of radio signals, mostly spurious, but once the algorithm had ruled out the terrestrial radio frequency interference (RFI) and astrophysical noise, eight candidate signals remained. These seem to be coming from five star -systems, all between 30 and 90 light years from Earth.

Despite being singled out by the machine-learning algorithm, the smart money is still on these eight signals being terrestrial RFI that slipped through the net. Much like BLC-1 — the anomalous signal detected by Breakthrough Listen, believed to originate from Proxima Centauri, but actually was from human-built electronics — they are almost certainly going to turn out to have a mundane origin.

What is important is that the algorithm was able to sift through the torrents of more obvious RFI to find the eight unusual signals quickly. More pertinently, when the same data was fed through the traditional search pipeline, the eight signals were all missed. In other words, Ma’s machine-learning algorithm is spotting interesting signals that otherwise would not have been detected. 

“The top eight candidates [were] detected coming from noisy parts of the spectrum, in the range 1.4–1.6GHz,” Ma tells Supercluster. “The algorithm seems to perform better in certain regards, in that it discovered more interesting signals that the traditional algorithm failed to detect.”

The algorithm strikes a balance between supervised and unsupervised machine learning. Imagine you were training the algorithm to recognize pictures of cats and dogs, says astronomer Steve Croft of Breakthrough Listen at the University of California, Berkeley. In supervised machine learning, you label the pictures ‘cat’, or ‘dog’ to help the algorithm. In unsupervised learning, the pictures are not labelled, and the AI has to figure out which are cats and which are dogs on its own.

“Then there’s in-between approaches like what Peter is doing,” Croft tells me. “He’s simulating some signals and then training an auto-encoder to reproduce the input, and then he’s feeding that into a random forest classifier that actually does the figuring out of what things are.”

Given how pervasive A.I. is becoming in our lives, we’re all having to learn the new language of artificial intelligence, only in this case it has taken on a distinctly arboreal dialect. A ‘random forest classifier’ is a collection of ‘decision trees’, which are a means to allow an algorithm to sort data by making decisions according to a tree-like model, with the branches being different outcomes, such as, is this animal a cat or a dog?

Though it sounds complicated, it’s really just pattern recognition, and in the context of SETI it becomes a kind of cosmic ‘Where’s Waldo?’ 

“The ultimate goal is a universal anomaly detector, where it figures out what all the human interference looks like and then says, here’s something that looks different,” says Croft.

BLC-1 was something of a watershed moment for SETI, where an encroaching bit of RFI — specifically the radio emission from a common oscillator widely used in electronics — was able to sneak through all the usual safeguards and filters and, for a short time, look like a convincing alien broadcast. When I asked Ma and Croft whether their algorithm would have recognized BLC-1 for what it was, they were unsure. BLC-1 was detected by a different telescope, in a different dataset to the one their machine-learning algorithm was trained on. The algorithm isn’t yet adaptable enough to be able to switch between such diverse datasets, but maybe there’s a quicker workaround.

“Currently we are working on tools that will help vet signals like BLC-1 by enabling us to better scan through datasets for ‘lookalikes’,” says Ma.

To do so, Croft hit upon the idea of a SETI analog to Google’s ‘reverse image search’, whereby you take an image and upload it to Google, and the search engine finds all the images that look like it.

“That would be great for [signals like] BLC-1,” says Croft. “If we can feed that signal in and ask the algorithm if there are any other signals like it in the data, it would help with us vetting it.” 

Croft remarks on how he idly mused about the idea of a reverse search function in his last meeting with Ma before Christmas 2022, and by the time they returned in the New Year, Ma had taken the initiative and already implemented the idea.

“Specifically, the reverse search algorithm would answer the question of, ’where else does this candidate appear in our data?’,” says Ma. ”If the signal is RFI then in the past we should have picked up similar looking hits. If it was truly a technosignature, then we should’ve seen only this kind of signal in observations of this particular target.”

Efforts to re-acquire the eight signals identified by the algorithm with Green Bank have so far failed. That doesn’t mean anything. All it tells us is that whatever the origin of the signals, they are not persistently switched on. It’s possible they will forever remain a mystery, like the Wow! signal of 1977. Either way, they’ll soon be joined by more: if the algorithm can detect eight signals in a relatively small dataset, just wait for what it will find when it’s let loose on data from one of the big next-generation radio telescopes, such as the ngVLA (Next Generation Very Large Array) or the SKA (Square Kilometer Array).

At the moment, A.I. seems really smart because it’s able to search through data quickly, but it’s only able to identify signals within the parameters that we give it. However, given the explosive acceleration in A.I. research, how soon before it is able to go beyond what we ask it to look for, and start to come up with its own ideas about how to detect extraterrestrial life?

In a field that is advancing at a rate governed by Moore’s Law, it’s difficult to predict where research into artificial intelligence will take us next. When I asked what A.I. might be able to do for SETI in, say, 10 years’ time, both Ma and Croft demurred. Instead, I chose to ask an actual artificial intelligence — ChatGPT — and put to it the question, will A.I. be able to think of new technosignatures that humans haven’t thought of?

“Yes, it’s possible that artificial intelligence (A.I.) could help us identify new technosignatures that humans haven’t thought of,” the chatbot told me. “A.I. has the potential to analyze data and patterns at a scale and speed that would be impossible for humans to do on their own, and it can also identify patterns and connections that may be difficult for humans to detect.”

Interviewing an artificial intelligence for an article was certainly a first for me. Of course, ChatGPT isn’t really sentient; it’s a smart algorithm that can search through and gather human information on the Internet really quickly to provide a response. 

“The thing that’s been blowing everybody’s mind, mine included, in the past 12 months or so has been the generative algorithms,” says Croft. “I’ve seen creative things coming out of ChatGPT where it is synthesizing all of human knowledge and then trying to move that forward to the next step. So I hope this evolves to the stage where we can ask it to take the blinders off and imagine, based on everything it knows about physics, biology, chemistry, exoplanets, technology, all of these things that the A.I. has learned, and to tell us what it thinks ET might be doing. It will probably come up with some good ideas.”

If A.I. does eventually find ET, there could be some symmetry to the discovery. Several prominent scientists with an interest in SETI, including most recently Cambridge’s Martin Rees and the SETI Institute’s Seth Shostak, have suggested that alien life, if it is technologically millions or even billions of years ahead of us, might no longer be biological, but artificial.

I talked about this with Shostak a few years ago. “I’ve been saying this for a long time,” he told me with his usual passion for the subject, imploring us to look at the timescales for technological development. One minute we’re inventing the motorcar or radio, and the next we’re developing artificial intelligence, and progress in that field is only going to accelerate to the point where we have artificial general intelligence, in the form of machines that can think for themselves.

“Thinking machines can evolve much more quickly than biology can, so it seems to me very straightforward to argue that the majority of intelligence in the cosmos is not biological, but is synthetic machine intelligence,” said Shostak.

Here’s the tricky part: there’s no need for an A.I. to live on a planet with oceans and a breathable atmosphere and a stable climate and all those things astronomers look for when thinking about life. A machine might be just as happy in the cold vacuum of space.

“Since machine intelligence might be the dominant form of intelligence in the Universe, then the obvious question is, how do we find it?” asked Shostak rhetorically. “It may not be on a planet — why would it be on a planet? It doesn’t need a planetary surface, it needs different things, it needs an energy source, it may need a heat sink, it may need this, it may need that, and we really have no idea what may be of interest to these machines. It’s like earthworms trying to figure out the behavior of humans — they’re not going to have very good ideas.”

We may be the earthworms to the Universe’s super-intelligences, but perhaps A.I. might see kindred spirits in these alien machines, and be able to make better educated guesses about what the aliens might be up to, where they might be and, more to the point, how to find them. It might take our A.I. to find an alien A.I.

As we move into the realms of full-blown speculation, here’s another thing to mull over. SETI research is beginning to consider the possibility that there could be visitors to the Solar System even now, probes sent to explore, just like we send robotic emissaries to the other planets of the Solar System. It’s a long shot, but if there were an alien probe here, so far from home, it’s likely able to act autonomously with some form of artificial general intelligence.

The prospect of sending A.I., possibly in some kind of melding with biological intelligence to explore the Universe makes sense, according to Martin Rees. In 2013 I was able to sit down for a chat with him at a 100 Year Starship conference at the Royal Society in London. There he talked about how the environment of space, coupled with the huge distances between stars and the time-spans involved in interstellar journeys, means that machines are better suited to exploring space than biological life.

“The motive for interstellar travel is rather weak so long as we are humans with human limitations,” Rees told me. “I think it would be a more attractive prospect if we evolve into some new species that lives for much longer, or if we can develop human–machine hybrids that can have longer effective lifespans, because it is going to take a long time [to reach the stars] and that timespan is not daunting for entities that can transcend human limitations.”

So, if aliens want to come to us, they may well send their artificial intelligences instead. If we were to discover an active probe in our Solar System, how could we communicate with it? Again, having a home-grown artificial general intelligence could give us an edge in reaching out to and talking to an intelligent alien machine.

For now this is all science fiction, but with the way things are progressing, it may not remain science fiction for too much longer. Astrobiologists often look to understand the origin of organic life on Earth to better understand what alien life might be like. Perhaps we should also start to chronicle the origin of artificial intelligence on Earth for the same reason.