Scientifically Speaking: AI is already changing biology and medicine, but guardrails are necessary

Scientists have demonstrated how artificial intelligence (AI) can be used to generate genetic instructions for viruses in a groundbreaking study. The researchers at Stanford University and Arc Institute created 285 virus designs and only 16 of them actually worked. This study, initially shared in September, was later published in the journal Science. The sequence of events involved creating DNA from AI suggestions, inserting it into bacteria, and observing if the bacteria produced new viruses.

The viruses used in this experiment were chosen from bacteria, and they were unable to infect humans. The researchers trained AI models on vast amounts of DNA data, specifically focusing on a virus called phi X174 and its relatives. After the model was trained, it was able to propose new genetic sequences which were then turned into DNA molecules and inserted into bacteria. Despite most designs failing, the successful ones produced viruses that multiplied and infected their hosts.

Interestingly, one of the successful AI-generated viruses contained a protein from a more distant relative, a change that previous human engineering attempts had failed to incorporate. This showed that AI can find biological solutions that conventional engineering could not.

The implications of this study extend beyond the realm of virology. For instance, viruses that target bacteria can be used to treat difficult-to-treat bacterial infections. Moreover, viruses can also be used as delivery vehicles for inserting therapeutic genes into patient cells. Therefore, better design methods for these viruses could have significant benefits for medicine and laboratory research.

However, it is crucial to introduce "guardrails" in the use of AI, just as we do in other areas of technology. The researchers in this study deliberately chose viruses that could not infect humans and tested their designs under controlled conditions, which contributed to the success of their experiment. As AI becomes more capable, human oversight becomes even more vital.

The hospital examples of AI helping to identify patients at risk of sepsis demonstrate the value of human assessment of AI findings before acting. Ultimately, while AI can help write the instructions, humans must remain responsible for the outputs and their consequences.