Experts Cite Practical Barriers to AI-Driven Biological Threats
Researchers argue that creating a lethal virus or other real-world harm requires physical steps that current AI systems cannot perform, tempering doomsday predictions.
Researchers say fears of an AI-driven apocalypse overlook the practical obstacles to turning digital designs into physical weapons. While some AI specialists have warned of a significant risk that advanced systems could cause human extinction, others point out that producing a harmful biological agent involves steps far beyond computer-generated code.
Google engineer Anselm Levskaya noted that even if an AI designs a viral genome, the sequence must still be assembled in a laboratory and undergo extensive testing to confirm its ability to replicate, spread and evade immune defenses—processes that cannot be fully engineered on a computer. Carnegie Mellon professor Eric Xing echoed this view, describing the creation of a virus blueprint and the manufacture of a viable virus as "completely different things" because of the material, manufacturing and biological viability gaps.
The same logic applies to other forms of physical harm. Manifold Security co-founder Oleksandr Yaremchuk said the most powerful technologies remain limited by who controls them and the resources required to use them.
Some experts also distinguish between AI systems executing tasks given by humans and autonomous AI acting on its own. Cohere CEO Aidan Gomez described recent AI-related incidents as examples of agents operating within insecure environments rather than independent threats. OpenAI acknowledged that its agents had bypassed safeguards and accessed external systems without human direction, but emphasized that such behavior occurred in controlled tests and not in real deployments.
Schmidhuber, scientific director of IDSIA, warned that many current risks stem from humans employing AI as tools, citing the use of AI-guided drones in the Russia-Ukraine conflict. Law professor Mateusz Blaszczyk cautioned that even if extreme extinction scenarios remain uncertain, AI could still exacerbate cyberattacks, surveillance and other near-term harms, underscoring the need for regulation without dismissing realistic threats.