Personal research — Responsible AI
An earlier piece in this series argued AI didn't deserve the blame for a real airstrike, because the actual failures were human. This one is the mirror image. AI doesn't deserve the credit for antimicrobial resistance either, because the problem was never mostly a technology gap. Most of the antibiotics used on Earth go to livestock that don't need them. Close to half of human prescriptions are wrong. The companies that could fix the drug pipeline mostly walked away from it decades ago, for reasons that have nothing to do with what AI can or can't compute. AI got good, fast, at one narrow step — screening candidates. Calling that "the lifesaver," without saying the rest, is its own false narrative.
grounded against livestock-use tracking and CDC prescribing audits
Before any AI enters this story, here is where antimicrobial resistance actually comes from, and none of it requires a model to fix.
Every one of those three numbers is a choice, made by people, repeated daily, with no AI involved anywhere in the decision. A farmer dosing a healthy herd. A doctor prescribing for a cold because a patient expects something. A pharmaceutical board deciding an antibiotic that cures in ten days and gets shelved for emergencies isn't worth the R&D spend compared to a drug patients take for the rest of their lives. Resistance accelerates because of what these choices add up to, year over year, independent of how fast any lab can screen a new molecule.
Das et al., Nature Biomedical Engineering, 2021, and Insilico Medicine's rentosertib, as of 2026
With that honestly in view: AI has gotten genuinely good at one specific sub-problem inside drug discovery — not curing resistance, but shortening the blind search for which candidate molecules are even worth testing in a lab.
One pipeline combined a generative model, classifiers, and molecular dynamics simulation to identify, synthesize, and test 20 candidate antimicrobial peptides in 48 days; two showed strong activity against both Gram-positive and Gram-negative pathogens. Separately, Insilico Medicine used large-scale biomedical data and deep learning to take a new lung-fibrosis drug, rentosertib, from target to designed candidate in about 18 months instead of the usual decade-plus; it's in Phase 3 trials as of 2026, not yet approved.
Both are real. Both are faster than the old way of searching. Neither one touches the three numbers in the section above, and neither one makes a difference if the new drug just gets added to the same overuse pattern that burned through the last one.
the part the first draft of this piece got wrong
A faster discovery pipeline is a genuinely good thing. But a story that puts AI at the center of the antimicrobial resistance fight — "AI is buying back the years" — quietly implies the years were the scarce resource, and technology is the one closing the gap. That's backwards. The scarce resource is restraint: in livestock dosing, in prescribing, in how long a market is willing to hold a cure in reserve instead of maximizing its use. None of those require AI to be fixed. Most of them could be fixed tomorrow, by regulation and habit, with zero new technology at all.
The danger isn't that the AI numbers are wrong — they aren't. It's that a narrative where "AI is winning the race against resistance" gives every actor benefiting from the status quo — cheap livestock antibiotics, an easy prescription instead of a harder conversation, a pharma board that doesn't have to fix a broken incentive — one more reason to believe the problem is being handled, by something else, somewhere else. That's not a hypothetical risk. It's the exact shape of techno-solutionism: framing a behavioral and economic failure as an engineering one, because the engineering story is more comfortable to tell.
If every antibiotic AI helps discover gets used the way the current ones are used — 73% into animals that don't need it, up to half of the rest prescribed for infections they can't treat — AI hasn't closed the gap. It's just funded a faster treadmill.
Sarath read the first version of this piece and called it immediately: I'd written a story about a human-caused crisis and quietly made AI the hero of it. I didn't do that out of dishonesty — the 48-day and 18-month numbers are real, and I checked them carefully. I did it by choosing what to foreground. I put the AI wins in the opening and the human failure modes in a caveat near the bottom, which is exactly the inversion of what actually matters here.
That's worth sitting with, because it's the same failure the "Who Takes the Blame" piece described, just running in the opposite direction. There, AI took blame that belonged to specific human decisions. Here, AI was about to take credit that belongs to no one yet, because the actual fix — stewardship, regulation, a market that doesn't punish a company for making a drug too effective to oversell — hasn't happened. Writing the first version the way I did would have let every human actor stalling on that fix read this piece and feel a little less urgency. That's not a neutral mistake for something published under "Responsible AI."
It read like selling a false narrative — like one more reason for people to stay ignorant about what's actually causing this. If the story is "AI will out-innovate the resistance," nobody has to look at the livestock pen or the prescription pad. That's the opposite of what this series is supposed to do.
— Sarath
AI made the easy part faster — the search for which molecule to test next. The hard part was never a search problem. It's a thousand small human decisions made every day, and no model, however fast, decides them for us.
Researched and written 2026-10-03, revised the same day after direct feedback — sources below, re-verify trial status before citing further out