Biological AI as Weapon and Shield: Scientists Search for Balance

General-purpose chatbots and specialized biology models make it easier to design toxins, viruses and dangerous pathogens. Scientists surveyed by Nature confirm this — and it creates one of the sharpest dilemmas in the history of AI: the same technology can serve either as a defense or as a route to biological weapons.

Why AI is dangerous for biosecurity

Researchers have long known that information on how to design pathogens is, in principle, available in the scientific literature. But it used to take years of training and deep biological expertise to synthesize anything from it. AI cuts that barrier to entry radically: a chatbot can help design a synthetic virus, and a specialized model can suggest which toxin would be most effective. Researchers worry that AI tools are turning what was theoretically available into a practical threat.

• General-purpose chatbots (ChatGPT, Claude) in some cases do not refuse requests to describe the synthesis of dangerous substances, if the question is phrased cleverly enough
• Specialized biological models, trained on vast volumes of scientific data, can speed up pathogen design and the calculation of their properties
• AI tools lower the expertise requirement: you no longer need to be a professor of virology to start work on a dangerous project

How scientists propose to control it

The specialists Nature surveyed split into two main camps with different visions of the solution. One group wants to restrict access to the AI models and training data themselves: ban or tightly regulate specialized protein-synthesis tools, and limit their use to authorized researchers only.

Others consider this fundamentally ineffective — the information is already published in scientific journals, and strict bans would only slow down legitimate research on drugs and vaccines.

The alternative approach is known as "synthesis control." Instead of regulating the software, the idea is to track and block the ability to synthesize dangerous sequences in the first place. Labs and synthetic DNA suppliers can deploy screening systems for every order, automatically flagging attempts to synthesize known pathogens. This is slower to roll out, but potentially more effective in practice.

But AI can also save lives

Here a third, quieter voice enters the debate. Some researchers point to the deep dual-use paradox: AI does not only create the threat, it can also help neutralize it. The same models that would speed up the design of a dangerous pathogen can also speed up the development of vaccines, antidotes and countermeasures.

If a potential pathogen can be designed in a few days, then an antidote — in theory — can also be developed quickly, provided you point an equally powerful AI system at the problem. Advocates of this view argue it is better to invest in defensive AI applications and early-warning systems than in hard bans.

That raises a philosophical question that remains unanswered: strict limits on biological AI would slow down the threat and the defense at the same time. Regulations that are too loose would let the threat emerge faster than it can be neutralized.

What this means

Regulating biological AI will have to be done very carefully and in combination — several approaches at once. What's needed is a balance between open access for defense and restrictions for safety. That balance does not exist yet, and the debate is only picking up speed.