The article separates AI discoveries into interpolation within training data, extrapolation beyond it, and representation discovery, where a model learns new structures in the data; the 2019 halicin case involved extrapolation and, especially, representation discovery.

In that research effort, an AI model identified halicin, a potent antibiotic, by learning molecular relationships linked to antibiotic activity that chemists had not recognized.

The case illustrates how models can make important discoveries incidentally, by learning patterns their creators neither intended nor anticipated.

More broadly, people may design the search while a model supplies the learned structure that makes a finding possible, whether it is a molecule, a connection in scientific literature, or a novel game strategy.

The article argues that calling AI merely a tool can understate its role when it generates representations and hypotheses humans could not design, raising questions about authorship and credit.

It recommends distinguishing human contributions to search design from the model’s contribution to representation, and crediting the model for what it learned—not just reporting the discovery.

The article also calls for more transparency into model representations through methods such as sparse autoencoders and attribution, while recognizing that learned structures may remain only partly understood.