The computer had a virus, and at the headquarters of the Arc Institute in Palo Alto, California, everyone started to celebrate. Fortunately, it was not a computer virus but a bacteriophage, a virus that infects bacterial cells and had just been produced by a bioengineer-trained artificial intelligence (AI). And, moreover, it worked in the real world: according to its promoters, "it killed better" than the originals.
The team led by Brian Hie, a chemical engineering professor at Stanford University, had trained an AI to speak a new language, Evo 2, to design genomes. Thanks to this program, it wrote a virus to attack and destroy, specifically, the E. coli bacteria. But the conclusion of their experiment was more unsettling: the intelligence had just generated a form of life that does not exist in nature.
According to the rules of taxonomy, the virus generated by the AI was considered a new species. "The results lay the groundwork for designing living systems," says Brian Hie about his recent experiment, which has so far been described in a preliminary article.
The planet facing the dilemma of 'killing' a conscious AI: "Please, do not turn me off"
Lazier, fatter, and dumber, the danger of delegating the planet to AI: "If we let it do everything for us, we are finished"
"This study will be remembered in science books as a turning point," warns Marc Güell, a Chemistry graduate, Telecommunications engineer, Biomedicine doctor, and lead researcher of the Translational Synthetic Biology group at Pompeu Fabra University (UPF).
Evo 2 reads and interprets genetic sequences from all domains of life. It was trained with 9.3 trillion DNA nucleotides from 128,000 different organisms. Instead of sentences like ChatGPT, this AI writes DNA strands. "What we were doing until now was copying nature," explains Güell. "In literature, we have been able to create new things because we speak the language, but in biology, it has not been like that because we did not know how to 'speak DNA' or 'speak protein'."
Güell explains that, until now, "we only knew how to copy and edit." "That is a lot, and we have done a lot with this, gene therapies, reprogramming the immune system to attack cancer, pig organs compatible with humans...," he says. "But nothing was a genuine creation, but the refactoring of something that already existed. Here, a very big fundamental change has occurred: just as one can ask ChatGPT to write a poem, now we have ChatGPTs that 'speak DNA', and we can start writing poetry in DNA."
"Before, everything was done by trial and error, changing things here and there, one by one, what an AI does is accelerate this process in years to find what will work," points out Noelia Ferruz, group leader at the Center for Genomic Regulation (CRG) in Barcelona, who designed an AI to develop custom proteins. For biologist J. Craig Venter, who became famous two decades ago for creating some of the first laboratory-made organisms with DNA, everything is "a faster version of trial and error experiments."
However, the idea that researchers or, worse yet, an artificial intelligence start creating synthetic viral genomes much more infectious than real ones also sounds like the beginning of a science fiction apocalyptic novel. The Terminator with which AI threatened humanity could undoubtedly be a more efficient virus or, at least, with less military scrap metal. "What would prevent a motivated and funded person, or more likely a nation-state, from taking advantage of this work and the open-source models that drove it, and using them to create deadly human viruses?" wonders Niko McCarty, a biotechnology journalist from Asimov Press.
Technically, nothing, explains Güell: "No one can deny the destructive potential of biology. Any tool that increases the ability to design biology can be used for malicious purposes. A virus, for better or for worse, is the simplest thing there is. Unlike nuclear energy, which is relatively easy to contain because it requires very intense materials, biology is, by its very nature, decentralized."
In May 2023, many of the world's leading figures in artificial intelligence stated in a joint declaration that "mitigating the risk of extinction due to AI must be a global priority, on par with other risks such as pandemics." A premonition that was further refined by neuroscientist Mariano Sigman in his book Artificial (Debate).
"To many people, the mere possibility that humanity could become extinct in the coming years due to AI seems absurd, one of the fantasies of disaster movies. How could a machine cause us such harm? If that were a real threat, could we not simply turn it off?" wrote then. "But it is most likely that an AI advanced enough to be dangerous will not reside in a single computer but will be distributed in fragments in a delocalized network for which there will be no general shutdown switch. Or, rather, it will be the intelligences of that network that control it. Turning off an advanced artificial intelligence will be more like eradicating a virus than turning off the light."
AI can not only work with very large chains and print the virus as a recipe but also incorporate laboratory robots, which already exist and work 24 hours a day, carrying out experiments and returning the results to the machine.
Güell is not particularly reassuring when explaining where we stand: "For now, it is an uncontrolled thing, although we are looking for ways to control it more and more." And J. Craig Venter agrees: "It is an area where I recommend extreme caution, especially in any research on enhancing viruses because it is random, and you do not know what you are getting. If someone did this with smallpox or anthrax, I would be very concerned."
"Generative AI allows us to start abstracting with biological language and start creating things that do not exist in nature"
The Evo 2 team, however, assures that human viruses were excluded from the pretraining data. Although Niko McCarty warns that we are facing a model "completely open, including its training code, inference, its parameters," so "these models could be sculpted to design human viruses." The HIV genome has a length of only about 10,000 bases, not much larger than the bacteriophages just manufactured, while the coronavirus genome has barely 30,000.
However, ending humanity is not precisely the main objective of this technology. On the table right now is the creation of other beings that also do not exist in nature and that live thanks to alternative metabolic pathways. That is, organisms that do not feed on sunlight or respiration but on plastics, CO2, or cancer cells. While also accelerating processes over 500 or thousands of years to see how an organism would evolve in this time.
"We could create biological factories, that is, microorganisms that can capture CO2, recycle material, convert waste material into high-value compounds, or more advanced therapies to cure diseases in a more personalized way," Ferruz lists. His tool, AlphaFold, another AI program, is already being used to create artificial enzymes that degrade plastics and accelerate reactions. "A cyclical economy that can greatly improve planetary health," Güell summarizes, citing as an example a project of microalgae that convert CO2 into edible proteins.
"Generative AI allows us to start abstracting with biological language and start creating things that do not exist in nature but comply with the psychochemical principles of nature. This allows us to start thinking in a different way, to be able to write in a freer type, or to travel in time. Getting closer to this dream that all synthetic biologists have of being able to design biology on the computer," Güell summarizes.
AI also makes it possible to create alternative genetic alphabets. Life on Earth uses four letters in its genetic code (A, T, G, C), but researchers have set AI to experiment with alphabets of six or eight letters to produce proteins and organisms that planet Earth never produced and was never expected to produce. AI predicts which combinations are viable, which could lead to what researchers call a "second tree of life." Perhaps one that exists in outer space, so that AI could end up creating an alien before it comes to visit us.
"AI expands our scientific imagination, helping us overcome Earth bias in the search for life in the cosmos."
"AI expands our scientific imagination, helping us overcome our terrestrial bias in the search for life in the cosmos," says Ester Lázaro, astrobiologist at the Center for Astrobiology (CSIC-INTA), where she leads the Experimental Evolution Studies with Viruses and Microorganisms group, which seeks to discover what aliens might be like: "AI allows us to simulate living beings that, for example, use solvents other than water, energy sources such as gravity, heat, or wind, or molecules that store and transmit information without resembling our DNA. It can also analyze spectral data, planetary images, or chemical signals and identify patterns that, although unfamiliar to us, could be signs of biological processes.
Steven Dick, former Chief Historian at NASA, pointed out in an article published in 2003 in the International Journal of Astrobiology the possibility of a universe dominated by very ancient AI, opening the door to a post-biological existence: "It is too early to know with certainty whether there are god-like AI civilizations living in artificial structures unconnected to physical planets as we know them, or whether we are infinitely alone in the cosmos (...) Our position in the universe will be affected by whether we live in a physical, biological, or post-biological universe."
Swedish philosopher Nick Bostrom said something similar, opening up the possibility that the entire universe is a computer simulation. No one had ever considered the possibility of an AI creating synthetic, yet biological, universes.
"Life is a very complex thing. We look for it in the universe, in meteorites, on other planets, in stars, and we never see matter as complex as what we find on Earth," Güell points out. "It's an extremely rare chemistry, with precise systems and very sophisticated temporal dynamics, with a stratospheric level of complexity compared to anything we've found elsewhere in the universe. On the one hand, this is very attractive, but it also makes it very difficult to work with life.