Should we worry AI will create deadly bioweapons? Not yet, but one day

Artificial intelligence holds the potential to revolutionize biology, enabling the creation of improved drugs, vaccines, and even engineered organisms for purposes such as decomposing waste plastic. However, there are concerns regarding its possible application for malicious activities, such as generating bioweapons that conventional detection methods may not catch until it is too late. Should we be genuinely concerned?

“Advancements in AI are driving significant progress in biology and medicine,” states Eric Horvitz, Microsoft’s chief scientific officer. “With increased power comes the responsibility to remain vigilant.”

Horvitz’s team has conducted research examining whether AI can engineer proteins to mimic the functions of known dangerous proteins, while being distinct enough to evade classification as hazardous. Though details about the proteins involved remain undisclosed, it’s likely that they explored toxins such as ricin—infamously associated with a 1978 assassination—and botulinum, the highly potent neurotoxin commonly referred to as Botox.

To produce ample quantities of a protein like botulinum, one needs its genetic blueprint—the DNA encoding that protein. Typically, biologists source specific DNA sequences from companies that specialize in providing customized DNA.

Fears over potential bioterrorists acquiring DNA sequences to produce bioweapons have led some DNA synthesis firms to voluntarily screen their orders to identify attempts to produce hazardous materials. This screening targets whether any synthesized amino acid sequence corresponds with known “sequences of concern,” or potential bioweapons.

However, AI may enable the design of protein variants with altered amino acid sequences that retain the same functionality. Horvitz and his colleagues explored this with 72 potentially dangerous proteins, revealing that existing screening methods frequently overlook these alternative variants.

This scenario is not as alarming as it may appear. For one, the researchers did not actually synthesize the redesigned proteins, which carries clear safety implications. Additionally, an earlier study this year tested redesigned versions of benign proteins and discovered they were ineffective.

Moreover, although there have been attempted bioterrorist actions—though relatively few—there’s scant evidence to suggest these attempts are due to shortcomings in current voluntary screening measures. There are numerous alternative avenues for bypassing these controls, such as acquiring ricin from castor oil plants commonly found in gardens. This study is akin to sounding the alarm about a sophisticated heist while the vault remains unlocked.

Importantly, excluding state actors, no bioterrorism incidents targeting individuals with protein-based bioweapons have been successful. The Aum Shinrikyo cult in Japan aimed to use botulinum against people but only managed to employ chemical agents effectively. The ricin-laced letters delivered to the White House resulted in no fatalities. Statistically, firearms and explosives pose a far greater danger than biotoxins.

Should we then disregard concerns about AI-driven bioweapons? Not exactly. While Horvitz’s research focused solely on proteins, the real anxiety lies with viruses, and AI is already being utilized to modify entire viruses.

Recently, researchers at Stanford University unveiled findings from their efforts to redesign a virus that targets the bacterium E. coli. Similar to the protein redesigns, outcomes proved to be underwhelming—out of 302 AI-engineered viruses, only 16 succeeded in infecting E. coli. Yet, this is merely the beginning.

Discussing AI-engineered viruses, James Diggans from Twist Bioscience, which is a partner in Horvitz’s team, stated that DNA-encoding viruses of concern can be more efficiently detected than proteins of concern. “Synthesis screening benefits from abundant data. At the genomic scale, that information is incredibly valuable.”

Nonetheless, not every DNA fabrication company employs these screening protocols, and compact DNA synthesizers are becoming accessible. Conversations are ongoing regarding the development of AI tools capable of refusing to engineer hazardous viruses or identifying malicious intent, but people have historically found ways to circumvent safeguards designed to, for example, prevent AIs from providing bomb-making blueprints.

To be clear, historical evidence indicates that the threats posed by “wild” viruses far outweigh those from bioterrorism. Contrary to claims from the current US administration, evidence suggests that SARS-CoV-2 emerged following a bat virus jumping to other wild species, subsequently infecting humans at a market, without laboratory involvement.

Furthermore, a would-be bioterrorist could inflict considerable harm simply by releasing a known virus like smallpox. Given the substantial vulnerabilities in bioweapon controls currently in place, there’s little incentive for resorting to AI-related schemes to bypass them.

For all these reasons, the likelihood of an AI-generated virus being unleashed in the near future is probably minimal. However, this risk is poised to increase as technology continues to advance—the COVID-19 pandemic illustrated the significant disruption even a moderately pathogenic virus could cause. Concerns about bioweapons will likely become more pertinent as developments progress.