Creating deadly human viruses will get easier with AI | The Economist

Emerging Threat Landscape

• Governments fear that advances in synthetic biology, now accelerated by AI, significantly lower the barrier to creating biological weapons.
• Leading language models have surpassed human expert virologists in complex tasks such as bioinformatics and troubleshooting difficult experiments.
• The primary concern is "uplift," where AI provides novices or semi-skilled individuals with capabilities previously restricted to a small number of state actors.

User Capabilities and Risks

• AI acts as a tireless tutor, guiding users through complex experimental steps with access to the entirety of published scientific literature.
• Research indicates that novices with no laboratory experience derive minimal benefit from AI for sophisticated virological tasks.
• Individuals with existing expertise (e.g., a PhD in molecular biology) stand to gain the most, as AI mimics large teams of experts by troubleshooting and offering diverse analytical angles.
• The principal bottleneck for bioterrorism—access to extensive expert collaboration—is effectively removed by AI assistance.

Feasibility and Current Limitations

• A major study from last year found that using AI to design a novel pathogen with no known defenses currently requires data sets that do not yet exist.
• Expert hands can currently use AI to modify existing viruses to exhibit new characteristics or capabilities they do not naturally possess.
• Theoretical scenarios include a lone actor developing a respiratory virus, accidentally infecting themselves, and triggering a pandemic via public transmission.
• While biology remains significantly harder to weaponize than other methods of causing mass harm, the balance of risk could shift rapidly as AI capabilities evolve.

Mitigation and Regulatory Strategies

• Developers are improving model refusal mechanisms to deny requests regarding sensitive information on virus recovery and propagation.
• Current refusal protocols are vulnerable to being easily bypassed by motivated actors seeking to generate dangerous biological information.
• Proposed technical countermeasures include omitting sensitive data from training sets or restricting access to the models entirely.
• Governments can preemptively regulate by assessing model capabilities prior to public release to establish usage guardrails.
• Regulatory frameworks may include limiting who can access specific models and defining permissible use cases to mitigate risks.