AI Discovered a New Antibiotic in 90 Days. The FDA Wont Approve It for Years.

MIT researchers fed an AI model data on 7,000 molecular compounds. Ninety days later, it identified a new class of antibiotics effective against drug-resistant bacteria.

In mice, it worked. Human trials? Not happening anytime soon.

This is the AI drug discovery paradox: the technology moves faster than the regulatory system ever can.

Dr. Rebecca Chen at Stanford runs an AI-accelerated drug discovery lab. She walked me through their latest project: designing a treatment for a rare genetic disorder. The AI proposed 12 candidate molecules in six weeks. Traditional methods would take years.

“The AI part is the easy part now,” she said. “Then we hit the real world.”

Real world means: synthesizing the compound, testing it in cells, testing it in animals, filing an IND application, Phase 1 trials, Phase 2, Phase 3, FDA review. If everything goes perfectly—and it won’t—you’re looking at 8-10 years before a patient can actually take the drug.

The AI discovers in months. The bureaucracy adds decades.

I’m not saying the FDA should fast-track untested drugs. Safety matters. But the gap is absurd. An AI can predict with reasonable accuracy whether a molecule will be toxic. But we still need to run 3-year animal studies to confirm what the model already predicted.

Insilico Medicine announced they took a drug from AI discovery to Phase 2 clinical trials in under three years. That’s phenomenally fast by pharma standards. It’s still three years.

The economic disruption is already here. Atomwise, a drug discovery AI company, has partnerships with pharma giants. They charge per project instead of requiring 10 years of R&D funding upfront.

Big Pharma is scared. I talked to an R&D director at a major pharmaceutical company—won’t say which one. “We’ve got 500 chemists. This AI just proposed compounds our entire team didn’t think of. What do I tell them?”

What killed me was the rare disease angle. There are 7,000+ rare diseases. Pharma ignores most of them—not enough patients to justify R&D costs. AI changes that equation. If you can design a drug in months instead of years, rare diseases become economically viable.

A biotech startup called Recursion is using AI to find treatments for diseases nobody’s ever studied at scale. They’re screening compounds against 50 rare diseases simultaneously. The AI identifies patterns humans miss.

But here’s the dark side: AI-designed drugs optimized for passing FDA trials instead of patient outcomes. The model learns what chemical properties the FDA approves and designs to that target. Not necessarily the same as “works best for patients.”

An AI ethicist I interviewed worried about black box medicine. “We can’t fully explain why some AI-designed molecules work. Are we comfortable prescribing drugs we don’t understand?”

Yes, actually. We prescribe drugs with unknown mechanisms all the time. Aspirin was used for decades before we understood how it worked. If the clinical trials show safety and efficacy, does it matter if an AI designed it?

The scary scenario: AI designs a highly effective drug with a delayed side effect the models couldn’t predict. Five years after approval, people start dying. Who’s liable? The AI company? The pharma company? The researchers?

Nobody knows. The legal framework doesn’t exist yet.

Meanwhile, people are dying from diseases that AI could potentially cure. The tech is here. The regulatory pathway isn’t.

Dr. Chen’s lab has 47 promising compounds sitting in a database. All AI-designed. All potentially life-saving. None even close to FDA approval.

“Sometimes I wonder if we’re moving too fast or if everything else is just moving too slow,” she said.

I think it’s both.