Summary of Key Points
Stanford University used generative AI to design 16 phages that do not exist in nature—viruses that specifically target bacteria—which has raised concerns about whether AI could create deadly viruses. However, the actual purpose of this research was to address the challenges associated with using phages to treat bacterial infections: the difficulty in finding the right phages and the rapid development of bacterial resistance. The AI-generated phages are merely highly similar imitations of existing ones and are far from being capable of causing a pandemic. They have a low success rate, simple structures, and fail to overcome the core issues in virus design, such as originality and complex functionality.
1. The true purpose of the research: to save lives where antibiotics fail
Many people think this is about AI creating viruses, but the team’s initial goal was to solve the problem of antibiotic resistance. For example, in 2015, a professor was infected with the superbug Acinetobacter baumannii, for which all antibiotics were ineffective; he was only saved by phages. Phage therapy has not become widely used because finding the right phage for a specific bacterium is like hitting a lucky shot, and bacteria quickly develop resistance. The Stanford team wanted to use AI to help design phages that can kill specific bacteria, eliminating the need for a lengthy search in nature.
2. Why hasn’t phage therapy become popular? Finding the right phages is too difficult, and bacteria can fight back
Phage therapy has a longer history than antibiotics (it was popular in the 1920s and 1930s), but it fell out of favor with the advent of penicillin:
- Finding phages is luck-based: Although there are approximately 10³¹ phages worldwide, finding one that can kill a particular bacterium is like searching for a specific grain of sand in the desert.
- Bacterial resistance: Even if a useful phage is found, bacteria can quickly evolve to become resistant, requiring the search for new phages, making large-scale use impractical.
It’s like finding a key that fits a lock only to have the lock changed, forcing you to start over—very troublesome.
3. How AI generates phages: not from scratch, but by mimicking existing viruses
AI does not create phages from nothing; instead, it mimics training data, similar to how ChatGPT writes articles:
- The team trained the AI using 2 million phage genomes and specifically focused on sequences similar to Phage 174 (because E. coli is resistant to it), hoping the AI could generate a version that overcomes this resistance.
- Of the 302 sequences generated by the AI, only 16 were able to kill E. coli, and these new phages had 97% genetic similarity to Phage 174—essentially, they were copied with minor modifications, not entirely new viruses.
4. Why shouldn’t we worry about AI creating deadly viruses?
There’s no need to fear that AI will create a pandemic virus:
- Low success rate: Only 16 out of 302 candidates were effective, and phages are among the simplest viruses (with just 5,000 base pairs of DNA), compared to the 30,000 base pairs in the COVID-19 virus and the 14,000 base pairs in the flu virus, which have much more complex structures.
- Complex functionality: Deadly viruses need to evade the human immune system while being highly contagious and toxic, capabilities that AI’s imitations cannot achieve. It’s like ChatGPT can write articles but not create masterpieces like “Dream of the Red Chamber.”
- Misaligned goals: AI is good at creating large numbers of similar versions, but designing dangerous viruses requires originality, which AI currently cannot accomplish.
5. The significance of this research: proving AI’s potential, but not solving core problems
The value of this research lies in demonstrating that AI can generate usable phage sequences. However, it does not address the fundamental issues in phage therapy:
- Just as AI can write 1,000 articles, it may still not produce one that is effective in treating diseases.
- In fact, designing phages is even more challenging than designing antibiotics, which have simpler structures compared to complete viruses.
Therefore, this is just a small step forward for AI in the biological field and far from the beginning of creating deadly viruses.
Conclusion
This research is essentially using AI to help humans find the right “keys” (effective phages), not to create “bombs” (deadly viruses). Concerns about AI causing a pandemic are premature—AI has yet to produce even decent imitations, let alone original, dangerous viruses. Instead of worrying, we should focus on how AI can help solve the real problem of antibiotic resistance.