Ara Darzi, the Imperial College London surgeon who directs the Institute of Global Health Innovation, told WIRED Health in London on April 16 that AI-powered diagnostics are now achieving above 99 percent accuracy on antibiotic-resistant infections without any added laboratory infrastructure. The claim lands against a death toll of more than 1 million people a year directly attributed to antimicrobial resistance, with nearly 5 million more deaths linked to it. A 2024 report in The Lancet projected 40 million cumulative deaths from drug-resistant infections by 2050.
Darzi framed the moment bluntly. "We're standing, right now, in 2026, at the first genuine inflection point in this crisis," he said. The pitch: machine learning has finally caught up to a problem where every hour of delay carries a 4 to 9 percent increase in the risk of death.
The clinical bottleneck is diagnostic speed. Traditional workups for resistant infections take two to three days because they require culturing bacteria from a patient sample. For sepsis patients, that window is fatal, which is why physicians today largely guess at which antibiotic to prescribe while results come back.
Key facts
- 01Antibiotic resistance causes more than 1 million deaths a year directly and contributes to nearly 5 million more.
- 02A 2024 Lancet report projects 40 million deaths from drug-resistant infections by 2050.
- 03AI-powered diagnostics now exceed 99% accuracy without extra lab infrastructure, according to Ara Darzi.
- 04An NHS-DeepMind system identified unknown resistance mechanisms in 48 hours; the same problem took Imperial College London a decade.
- 05In southeast Asia and the eastern Mediterranean, 1 in 3 reported infections were antibiotic-resistant in 2023; in Africa, 1 in 5.
Closing that gap is what Darzi argues AI is built for. "AI-powered diagnostics are achieving accuracy above 99 percent without additional laboratory infrastructure," he said, pointing to rural and remote settings as the most obvious deployment target. The World Health Organization estimates that one in three reported infections in southeast Asia and the eastern Mediterranean were resistant in 2023, and one in five in Africa.
“DeepMind's system identified previously unknown resistance mechanisms in 48 hours, cracking a problem that had taken Imperial College London researchers a decade to solve.”— Jaeden Schafer
Drug discovery is the other front. The UK's National Health Service is working with Google DeepMind on a system aimed at antibiotic resistance, and in one demonstration the model identified previously unknown mechanisms of resistance in 48 hours. Researchers at Imperial College London had spent roughly a decade chasing the same answer.
Paired with an automated laboratory, Darzi said it is now possible to run hundreds of parallel experiments around the clock. Deep learning models can screen billions of molecular structures in days, and generative models are being used to design compounds that do not exist in nature. The technical ceiling on candidate molecules has, in effect, been removed.
The economic ceiling has not. Major pharmaceutical companies have largely exited antibiotic development because the business model is broken: any genuinely new antibiotic has to be held in reserve to slow resistance, which directly cuts the high-volume sales pharma depends on. The better the drug, the less of it gets sold.
Two governments are testing a fix. In 2024 the UK launched a pilot of a so-called Netflix-style payment model, in which the government pays pharma a fixed annual subscription for access to new antibiotics rather than paying per prescription. Sweden is running a partially delinked model along similar lines, separating revenue from volume.
The skeptical read is that 99 percent accuracy in a demonstration and 99 percent accuracy in a district hospital in Lagos or Jakarta are not the same number. Regulatory approval, clinical workflow integration, and reimbursement all sit between Darzi's slide and a patient outcome, and the pharmaceutical industry's retreat from antibiotics has not yet been reversed by either the UK or Swedish pilots. The DeepMind result also remains a single demonstration, not a deployed product.
Still, the shape of the argument is hard to dismiss. AMR is one of the rare problems where the AI capability curve and the public-health curve actually intersect on a timetable that matters, and where the bottleneck is no longer scientific but procurement, pricing, and political will. Darzi's closing line was the one worth keeping. "The question that will determine the shape of medicine for the next 100 years, is not whether we have the tools to respond. We have the tools. The question is whether we have the character to take seriously what we are seeing."
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
