Interview
Creating deadly human viruses will get easier with AI | The Economist
- Governments express concern that advances in synthetic biology, further accelerated by AI, will lower barriers to developing biological weapons.
- Leading language models are predicted to surpass human virologists in bioinformatics and complex experiment troubleshooting, potentially providing significant capability uplift to PhD holders in molecular biology while offering limited lab assistance to novices.
- While AI access may allow non-experts to draw on guidance from an "infinite number of experts," the balance of risks is expected to shift rapidly as capabilities evolve, creating scenarios where novice users could reach levels of capability previously restricted to state actors.
- Models are anticipated to improve at refusing dangerous biological requests, yet motivated actors are predicted to bypass these refusal mechanisms very easily despite theoretical government guardrails established through pre-release capability assessments.
- Material outcomes include the hope that no actor will release a deadly respiratory virus, though current assessments suggest it is not yet easy to destroy humanity.