The algorithm that caught what the emergency room missed 🌿

Researchers from Harvard Medical School and Beth Israel Deaconess Medical Center set out to compare an AI reasoning model against experienced emergency physicians. They published their findings in Science, one of the most prestigious journals in the world. The result was harder to dismiss than most AI health claims: OpenAI's o1 model outperformed doctors in clinical diagnostic reasoning, across multiple tests, using real patient data.

In the study's most striking benchmark, the AI achieved a perfect clinical reasoning score 98% of the time. Attending physicians scored the same benchmark 35% of the time. In a real-world ER simulation using de-identified patient records, the AI identified an exact or close diagnosis 67% of the time. Two experienced physicians given the same cases scored more than 10 percentage points lower.

The researchers are careful to note what the AI doesn't have: it can't perform a physical examination, read a patient's affect, or draw on the contextual knowledge a physician builds over a career. For now, it sees text, not humans. But that's precisely why this matters as an augmentation story, not a replacement one. Emergency rooms are under enormous pressure. Diagnostic errors in the ER are common, consequential, and often preventable. A tool that catches what an exhausted physician might miss, and does it at scale, addresses a real and urgent problem.

The next step is prospective clinical integration: using AI-assisted reasoning in live ER settings and measuring patient outcomes directly. We're not there yet. But the gap between "interesting research" and "clinical reality" just got meaningfully smaller. A study in Science, with real patient data, published by two of the most respected medical institutions in the world, is not a press release. It's a signal worth paying attention to.

Emerald AI, a 2026 BloombergNEF Pioneer award winner, partnered with NVIDIA and six major energy companies to demonstrate that AI data centres can dial down power consumption on command while maintaining full computational performance. Across five commercial demonstrations, the model held. If AI factories can act as flexible grid assets rather than just demand spikes, the energy narrative around AI changes structurally.

Source: NVIDIA Newsroom

Late-April negotiations in Brussels to delay the EU AI Act's August 2 full-application deadline collapsed. The world's most comprehensive AI regulation goes into full enforcement in three months, including risk management mandates, transparency obligations for synthetic content and deepfakes, and regulatory sandboxes opening in all 27 member states. The failed delay attempt signals political will to hold the line.

Source: PPC.land

A Philippine Institute for Development Studies policy note found that active use of AI-powered digital finance platforms is associated with a 78.5 percentage-point increase in formal account ownership. For a country with one of the world's largest unbanked populations, AI-driven fintech isn't a convenience feature; it's an on-ramp to a financial system that previously excluded millions of people entirely.

Source: Manila Bulletin

BCG Henderson Institute analysed 165 million U.S. jobs across 1,500 roles to produce what may be the most credible large-scale estimate of AI's workforce impact to date. The headline finding: 50 to 55% of American jobs will be reshaped by AI within two to three years. But reshape is doing a lot of work in that sentence, and the distinction matters enormously.

What "reshaped" actually means

Only 10 to 15% of jobs (between 16 and 25 million positions) face outright elimination over the next five years. The vast majority of workers, roles, and organisations will not disappear. They'll transform. The internal structure of jobs will shift, the tasks people perform will change, the skills they need will evolve. BCG is explicit about one thing: companies that over-cut headcount beyond what AI can genuinely absorb will lose productivity, institutional knowledge, and talent. The warning is built into the report.

The PM lens

For product managers building AI-powered tools and workflows, this is both a mandate and a calibration. The organisations that benefit most from AI won't be the ones that reduce headcount fastest, they'll be the ones that redesign work thoughtfully. That means investing in change management, building tools people actually want to use, and measuring adoption outcomes, not just deployment metrics. The 10 to 15% of roles facing elimination are largely concentrated in repetitive, rules-based work. The growth opportunity for AI product builders is in everything else: the complex, judgment-heavy, relationship-driven work that AI augments rather than replaces.

The risk lens

The BCG report surfaces a risk that often gets underestimated: premature workforce reduction based on AI capabilities that don't yet exist at scale. Companies cutting teams based on projected AI replacement, rather than demonstrated productivity gains, are taking on real operational and reputational exposure. There's also a concentration risk. If the skills that survive AI disruption are clustered in higher-income, higher-education demographics, then "most jobs survive" becomes a story of deepening inequality rather than shared progress. That's the scenario worth watching most carefully, and the one that regulation and workforce investment programs like the DOL's new AI Apprenticeship Portal are trying to prevent.

The next 12 to 24 months

The BCG findings put a clock on decisions that companies are already making. Organisations that move now to reskill workers, redesign workflows, and deploy AI where it genuinely adds value will be in a structurally different position than those treating AI as a cost-reduction lever first. The two-to-three-year reshaping window is already underway. The decisions being made right now in procurement, HR, and product strategy are the ones that will determine whether BCG's more optimistic reading, that most jobs survive and evolve, turns out to be accurate.

Most AI diagnostic tools need hundreds or thousands of labelled slides to learn a new cancer type. PRET (Pan-cancer Recognition without Example Training), developed by researchers at HKUST, Guangdong Provincial People's Hospital, and Harvard Medical School, requires as few as 1 to 8 annotated slides to diagnose across 18 cancer types, with no retraining required. In lymph node metastasis detection, it achieved a 98.71% AUC, surpassing the average performance of 11 pathologists. Validated across 23 international datasets, PRET is designed to work in hospitals where pathologist shortages are most acute and annotated training data is hardest to come by, published in Nature Cancer.

→ Read more: Medical Xpress

If the AI in your hospital consistently outdiagnosed your doctor, would you want to know? And should you have a choice?

Hit reply. We read every response.