Rather than training the algorithm on content scraped from the internet, scientists trained the AI on nearly three million genomes.