AI can now steer a living organism with plain English - no new lab work required.
Researchers built a natural-language interface for xenobots, synthetic multicellular constructs built from frog cells that have no nervous system. Instead of running fresh experiments to learn which instruction produces which biological behavior, they treated an existing archive of past interventions and outcomes as a fixed dataset. A vision-language model then judged, without any new wet-lab work or human review, whether an archived outcome matched a written description, and that judgment became the only training signal. The resulting system mapped new instructions to the archived intervention already known to produce the described behavior, hitting 80.0% accuracy on held-out data versus a 66.7% chance baseline - matching a model trained directly on ground-truth labels.
The bottleneck in applying AI to biology has never been clever architectures, it's data: every language-intervention-outcome triple usually costs its own wet-lab experiment. This work suggests old archives can substitute for new experiments, which matters for scaling language control beyond xenobots to organoids and other lab-grown biobots.
A xenobot has no nervous system and can't refuse an instruction, so "controlling" it with language is a lower bar than it sounds - the harder test comes when someone tries this on tissue that can actually push back.