AI-designed phages work. The rulebooks that should cover them were written for a different world.

A recent announcement of novel, viable viruses created by artificial intelligence was celebrated as an advance against antibiotic resistance, and on the science it earned that: the AI-designed bacteriophages kill only bacteria, and they rapidly overcame antibiotic-resistant strains of E. coli ^[1][2]. The more durable consequence is administrative, because that capability was assembled from technologies regulators have traditionally considered separately, while the laws themselves are organised around particular technologies, organisms or activities ^[4].

That is the argument in an analysis published by phys.org, and it is worth taking seriously because it declines the easy framing. This is not merely an AI safety problem, the authors write; the more fundamental question is what happens when computational technologies become part of biological design ^[17]. Three layers now stack. CRISPR made targeted edits to DNA routine, and has already produced approved medicine, including a CRISPR-based therapy for sickle cell disease ^[5][18]. Synthetic biology let researchers design and construct biological systems ^[6]. AI adds analysis of biological information, pattern-finding, and the generation of novel genomes ^[7]. The result is that the boundary between computational design and biotechnology is less clear at precisely the point where a statute needs a defined object to attach to ^[3][4].

Note what the analysis does not claim: it does not say any regulator has disclaimed the field. The problem it describes is structural. A gene technology regime that classifies organisms and activities has a natural handle on the wet-lab step and a weaker one on the design step, and an AI-safety lens has the reverse problem ^[4][17]. The capability lives across the seam.

The dual-use point is not decoration here. The researchers were pursuing a useful goal, demonstrating that AI could help design novel bacteriophages against antibiotic-resistant pathogens ^[9]. The same work could be misused, and some of the resulting risks may be hard to predict, detect or contain ^[10]. That is the standard shape of dual use: the technology that solves a problem can create new ones ^[8]. For regulators, the task is therefore harder than labelling a technology safe or dangerous; it is supporting useful research while managing the risks that can reasonably be anticipated ^[19].

Existing rules are not worthless. They provide safeguards for research, manage known hazards, and let governments distinguish activities that pose different levels of harm ^[11]. But some were written for a very different scientific landscape, and the analysis argues the rules are not keeping pace in New Zealand and elsewhere ^[12]. New Zealand's Gene Technology Bill is stalled in Cabinet over disagreements between coalition parties ^[13]. It would replace parts of the 30-year-old framework developed under the Hazardous Substances and New Organisms Act, so the rules currently in force predate the phage result by roughly three decades ^[14][20]. The proposed system is tiered rather than uniform: some lower-risk gene-editing technologies would be exempt outright, while activities judged higher-risk stay under controls ^[15]. On the analysis's reading, that shows governance need not restrict innovation ^[21].

Watch two things. First, whether the New Zealand bill clears Cabinet, and whether its exemption tier is drawn by the risk of the resulting organism or by the tool used to get there, because an AI-generated genome sorts very differently under those two tests ^[15][13]. Second, Australia, where the Gene Technology Act 2000 already runs a risk framework through a gene technology regulator ^[16].