Dual use isn't unique to AI — nuclear physics, biology, and cryptography all face it. What makes AI different is the speed of proliferation: a dangerous biological technique requires a lab; a dangerous AI technique requires only a computer. This means traditional dual-use governance (export controls, lab safety regulations) translates imperfectly to AI, where the "lab" is a laptop and the "materials" are open-source code.
Leading AI labs evaluate models for dangerous capabilities before release: Can it provide detailed instructions for bioweapons? Can it help with cyberattacks? Can it generate convincing disinformation at scale? These "dangerous capability evaluations" determine what safety measures are needed. Models that show elevated risk in specific areas receive additional guardrails, and capabilities are sometimes removed or restricted.
Dual use creates acute tension around open-weight model releases. Open models (Llama, Mistral) can be freely modified to remove safety guardrails, enabling misuse. But they also enable security research, academic study, privacy-preserving applications, and innovation that proprietary models don't allow. The debate has no easy resolution — both sides have legitimate arguments, and the optimal policy likely evolves as capabilities and risks change.