Sixteen Out of 285 - Why That Ratio Should Worry You More Than Reassure You

Sixteen Out of 285 Is the Wrong Number to Focus On

Here's what's actually happening: researchers didn't ask an AI to invent a virus from scratch. They trained a genome-language model called Evo on thousands of known bacteriophage genomes, then asked it to propose new, never-before-seen genomes within that same biological family.

Scientists physically built 285 of the AI's proposed genetic sequences and tested them. Sixteen worked as living, functioning viruses - some even solved a problem human engineers had struggled with: overcoming bacterial resistance to the original virus.

The Mechanism: A Language Model, But For DNA

Evo works exactly like a large language model, except instead of learning patterns in words, it learns patterns in the four-letter genetic alphabet - A, C, G and T - across enormous numbers of real genomes. Ask a text-based AI to write a paragraph and it predicts plausible next words; ask Evo to design a genome and it predicts a plausible, biologically coherent DNA sequence.

The scientists then had to physically synthesise that sequence and insert it into bacteria to see whether the genetic instructions actually produced a working virus. Most didn't. Sixteen did.

Why 16-of-285 Is Reassuring Only Once It's tempting to read a roughly 6% success rate as evidence this technology isn't dangerous yet.

That reading misses the actual risk. A low success rate matters only when the number of attempts stays small - and a computer can generate thousands of candidate genomes in the time it takes a human lab to test one. This is what biosecurity researchers call capability amplification: AI doesn't need to invent new virology from nothing, it just needs to integrate everything virologists already know about receptor binding, transmission and immune escape and explore combinations far faster than any human team could test manually.

Where India Stands Globally

• And Why That's the Real Stake India is investing seriously in sovereign AI capability through the IndiaAI Mission, but this story's genuine UPSC-relevant point isn't the Evo experiment itself - it's the choice India now faces. If frontier biomedical AI models remain concentrated with a handful of companies and countries, Indian researchers working on antimicrobial resistance, vaccines or cancer therapeutics could be stuck using smaller, weaker tools not because they're sufficient, but because nothing stronger is available to them. That's a scientific dependency problem as much as a security one. Building biosecurity screening into India's AI mission now - rather than after a crisis forces the issue - is the actual policy choice on the table. For the exam, the insight worth carrying is that generative biology's real danger isn't a rogue AI inventing a new pathogen unprompted - it's a knowledgeable, well-resourced lab becoming dramatically faster and more capable at doing what it could already, in principle, do. Governing that acceleration, not preventing some imagined spontaneous invention, is the actual regulatory challenge.