Two weeks ago, the magazine Science published a paper by researchers from Stanford University (Samuel H. King et al.) who used two genomic language AI models, Evo 1 and Evo 2, to create the complete genomes of 16 new viruses capable of successfully attacking the bacterium Escherichia coli C . Beyond the scientific achievement this represents, the fact that artificial intelligence is capable of generating viruses that do not exist in nature raises new alarms. The authors created bacteriophages – which do not attack animal cells – and applied strict safety measures, ensuring that their creation did not leave the laboratory. But Murphy’s law inevitability points out that if something can happen, it will happen. Biosafety needs to prepare for this new technoscientific capability.
Read more A group of intellectuals ask Sánchez and Macron for the release of Algerian journalist Ayoub Aïssiou
In a complementary text, the researchers stated that if their method were applied to design viruses capable of infecting cells instead of bacteria, the new pathogens “could be viable on a similar scale.” Because of that risk, they warned that work of that nature “should only be carried out after thorough review and deliberations with other researchers and experts in biosafety, biocontainment, and bioprotection, as well as respecting all current and future governance structures and best practices.”
Although variants of the ɸX174 virus, well known in the scientific community for laboratory study, presented minimal risk of significant ecological alteration, the researchers conducted all the work “inside a biosafety cabinet with appropriate PPE – protection – and specific equipment, which was periodically sterilized with 70% ethanol, 10% bleach, and UV – ultraviolet – treatment.” Additionally, they disposed of all waste as “biohazardous.”
Aware that they had to minimize risks, the authors of the work consulted biosafety experts. They were clear that if they had worked with viruses capable of infecting animals, they would have had to apply “a much stricter biosafety review and biocontainment protocols than those described in this study.” Furthermore, initiating such research would have required compliance with much more rigorous U.S. government regulations.
The concern about what creating a virus through AI could mean led them to apply an additional precaution: they deliberately excluded from the training data of the models all viruses capable of attacking cells, “including pathogens for humans.”
The growing access to resources makes malicious use cases increasingly possible
Accessing the data and computational resources to do so is not easy, but the authors themselves admit that “the increasing accessibility to such infrastructure makes malicious use cases increasingly possible.”
Read more Feijóo asks Sánchez to decree the lockdown in Ceuta due to «health risk»
Víctor de Lorenzo, research professor at CSIC at the National Center for Biotechnology, told Science Media Center that if this strategy were extended to designing viruses that attack cells, “opportunities would arise such as designing highly specific oncolytic viruses – against cancer cells –,” but he warned of “worrying scenarios, such as creating viruses with selective affinity for certain cell types, tissues, or even population groups.”
A gap in the European Union AI law
The enormous list of limited, regulated, or prohibited uses in the European Union AI law does not include the use of artificial intelligence for custom virus design, which can have therapeutic applications – such as creating pathogens that only attack cancer cells – but also presents risks due to its possible use as a biological weapon.
In any case, the wording can be flexible for interpretation. For example, Article 58, which regulates AI trials in isolated spaces. When the authorities of a country consider whether to authorize a trial, they must agree on “appropriate safeguards with participants, aiming to protect fundamental rights, health, and safety.”
Jordi García-Ojalvo, professor of Systems Biology at Pompeu Fabra University, believes the system is still inefficient, since they obtained only 16 viruses out of 300 genomes, and that “the danger here is less than in traditional large language AI models, since the designed genomes must be tested in the laboratory one by one.”
Along with the articles on the research, Science published a third by Thomas V. Inglesby and Moritz S. Hanke who observe that “the ability to compose viral genomes using generative AI already exists; however, the necessary governance to manage it safely does not yet exist.”
Also read
This happens because, although there are biosafety regulations, “generative genomics breaks this logic because the genetic alterations it creates are unpredictable.” “It is – they reflect – about whether society can establish an oversight system that allows its benefits to be harnessed while preventing it from causing serious harm.”
Read more The rise of used vehicles drives the import of old European cars
