Can AI help someone build a dangerous virus?

Recent reports indicate that people have already attempted to use large language models—the technology behind ChatGPT—to learn how to make existing viruses more transmissible or dangerous. These systems, trained on vast amounts of human scientific knowledge, can act as a kind of digital encyclopedia, offering instructions or guidance on complex experiments. The concern is amplified by the fact that gene-editing and synthetic biology tools are becoming easier for hobbyists to access. While AI companies have tried to put guardrails on their products to prevent them from providing dangerous instructions, researchers argue that these barriers are often easy to bypass. Currently, the industry is in a constant cycle of trying to build better filters, only for AI to find new ways around them.

Artificial intelligence models for science work by recognizing patterns in massive datasets. To discover new medicines, researchers use these models to simulate how different molecules will behave, which is incredibly useful for finding cures. However, the exact same process can be flipped. If a model is trained on chemical properties, it can be asked to generate a list of molecules that are not beneficial drugs, but highly toxic substances. Because these models can iterate through millions of possibilities in hours, they can effectively bypass the need for years of human trial and error. When you combine this computational power with current advances in DNA synthesis—which allows labs to order custom genetic material—you lower the barrier to entry for biological experimentation significantly. If the AI acts as the brain that designs the threat, the biotech tools act as the hands that build it.

There is no consensus on how immediate this danger is. Some biologists argue that the real-world process of actually creating and handling dangerous pathogens is still too difficult and time-consuming for these digital shortcuts to be the primary threat. They suggest that our existing health crises, like naturally occurring bird flu, remain a much more pressing concern than AI-designed weapons. Regardless of who is right, the conversation is shifting toward preparation. The goal is to build better surveillance, strengthen our public health response, and ensure that our safety measures evolve as quickly as the technology itself. We are learning that powerful tools rarely stay locked in a box, and the challenge ahead is figuring out how to use the power of AI to detect and neutralize risks faster than they can be created.