Beyond the “Super Virus”: How to Build Trust in AI-Designed Biology

When a recent Science paper reported the first generative AI design of complete, viable bacteriophage genomes, public discussion quickly converged on a more dramatic question: Can artificial intelligence now create a “super virus”?¹ The concern is understandable, but it misses the more consequential milestone.

The researchers used genome language models, Evo 1 and Evo 2, to generate bacteriophage genomes based on the well-characterized ΦX174 phage, which infects bacteria rather than humans. They synthesized and assembled 285 generated genomes and identified 16 viable phages. A cocktail of generated phages overcame bacterial resistance that challenged comparable natural phages. This is not evidence that an AI system can simply design a dangerous human pathogen on demand. It is evidence that AI is beginning to move from reading biology to helping write it.

In the use of AI in healthcare, trust has become the currency of adoption. A model can be highly accurate and still fail to earn trust if its behavior cannot be examined, audited, validated, or governed. Generative AI in biology raises the same problem with an added dimension: Outputs can move from a digital model into the physical world. Trustworthy AI in biology therefore cannot be treated as a final compliance check. Trust has to become its infrastructure.

Building Layers of Trust in AI for Biology

A useful way to think about building trust into AI models is as a layered “trust stack.” At the model layer, developers need rigorous capability evaluations, carefully governed training data, risk-appropriate access controls, and records that support accountability. The paper also points to model-level safeguards from prior Evo 2 work, including training-data exclusions designed to limit certain viral design capabilities.¹

At the design layer, AI-generated biological proposals should carry provenance information and auditability. Scientists and reviewers should be able to determine what system produced a design, under what constraints, and what evaluations preceded laboratory work. This does not mean that every biological model must be fully interpretable in a simplistic sense. It means that consequential decisions should not become untraceable simply because a generative model sits upstream.

At the synthesis layer, existing safeguards need to evolve. Many commercial DNA synthesis providers already use sequence and customer screening, but AI-generated designs increasingly challenge approaches that rely mainly on similarity to known sequences. Recent work has argued for screening that considers predicted biological function, not only sequence resemblance.² That shift matters because future models may generate functionally meaningful sequences that look increasingly unfamiliar.

At the laboratory layer, biosafety, biocontainment, and independent oversight remain essential. Digital safeguards cannot substitute for responsible experimental practice. The original authors emphasized this point, and a related Science commentary warned that the ability to compose viral genomes with generative AI creates urgent biosafety and biosecurity questions.³

Governance must connect these layers rather than regulate each in isolation. Biological design, cloud computing, model access, DNA synthesis, and laboratory work can cross institutional and national boundaries. Global screening practices remain uneven, and researchers have called for layered defenses rather than dependence on any single control.⁴ The timing of this new study is notable: The National Institutes of Health is currently seeking public comment on a draft biosafety policy intended to modernize oversight as biological risks evolve.

The final layer is literacy. Public legitimacy will depend on whether institutions can explain what is being done, why, by whom, and under what accountability. Scientists, AI developers, policymakers, journalists, and the public all need enough shared literacy to distinguish capability from consequence.

Public discussion cannot collapse “AI generated a DNA sequence,” “scientists synthesized it,” “a viable biological system resulted,” and “a dangerous human pathogen was created” into the same statement. These are not equivalent events.

Fear is a poor governance strategy. So is complacency. If every advance in generative biology is described as a pathway to a super virus, public trust will erode, and beneficial research may become harder to discuss rationally. If legitimate dual-use risks are dismissed because current experiments use nonpathogenic systems, the scientific community may lose the opportunity to build safeguards before capabilities expand.

The more productive position is to treat governance as part of innovation itself. Safety evaluations, auditability, synthesis screening, laboratory oversight, and public education are not barriers placed in front of progress. They are the conditions that allow powerful technologies to scale responsibly.

The significance of this Science paper is therefore not that AI has crossed some cinematic threshold into creating super viruses. It is that generative models are becoming participants in biological design. That is a remarkable scientific opportunity.