Hey Reader,

I’ve talked before about how education and medicine are two areas where AI has huge potential.

Both are fields where almost everyone on the planet would benefit from extra help or higher-quality care. There’s basically no limit to demand, so AI improving supply is a universal win.

Today I want to show you a handful of real-world studies that give us a glimpse into a future where doctors are better, teachers are better, and both patients and students get better outcomes.

The real numbers I’ll show you today:

• An extra life saved for every 14 high-risk patients
• Kids who learned 1.7 years of math in 8 weeks

Wow. 🤩

Let’s dig into the real outcomes.

In Taiwan, 1 extra life saved for every 14 high-risk patients

The medical study ran at two hospitals in Taiwan and was published in Nature Medicine in April 2024. It covered 39 physicians and 15,965 hospitalized patients.

The setup was modest. An AI read each patient’s electrocardiogram (ECG). It didn’t diagnose anything or pick a treatment. When it spotted a pattern linked to a high risk of death, it sent the patient’s doctor an alert. That’s it.

Half the patients got the alert version. Half got usual care.

Across everyone in the trial, 4.3% of the usual-care group died within 90 days versus 3.6% in the alert group. That’s roughly one life saved for every 143 patients, which is a solid result on its own.

But the AI had flagged 1,397 patients as high-risk, and that’s where the numbers get really wild. In that group, 23% of the usual-care patients died within 90 days. With the alert, it was 16%.

That’s about one life saved – directly by the added AI alerts – for every 14 high-risk patients.

Cardiac deaths in the high-risk group dropped from 2.4% to 0.2%.

So what did the AI do to save those people? It didn’t provide any direct care – it just flagged a situation the doctors wouldn’t otherwise have noticed. The study found that flagged patients in the alert group received more intensive care than flagged patients in the control group. The alert went out, and the doctors changed what they did.

The trial is two years old, and the reason it’s on my mind is a systematic review published in May in npj Digital Medicine that pooled 32 randomized trials of AI in cardiovascular care. Across the trials measuring deaths, about 33,700 patients, AI decision support cut all-cause mortality by 16%, or one death prevented for every 32 patients. Across this review, it wasn’t that AI did a better job directly treating or diagnosing – it was that AI worked when it prompted a human to pay attention to something important that might otherwise have been overlooked.

The takeaway: AI increases the capacity of humans to do good work.

In Sierra Leone, 8 weeks of AI tutoring helped kids zoom 1.2 to 1.7 years forward in math

In May, Google DeepMind and an education nonprofit called Fab AI published results from a preregistered trial with 1,763 seventh and eighth graders across 48 math classrooms in Sierra Leone.

For eight weeks, teachers in the treatment classrooms used Gemini’s Guided Learning mode in about half of their weekly math lessons. Control classrooms taught the usual way. Then every student took a test with no AI in the room.

The AI classrooms scored 0.26 standard deviations higher. The researchers translate it as 1.2 to 1.7 years of typical learning progress in low- and middle-income countries. From eight weeks of tutoring!

Students who hit the target of 12 hours with the tutor, which was 69% of them, gained 0.38 standard deviations, an even bigger learning jump. Google describes that as moving an average kid from the middle of the class to the top third.

The topics were fractions, exponents and prime numbers. My son is 12, so I recently got a first-hand look at how he and his peers learned these topics, too. We already use a similar method at home – I’m his “tutor” and I get support from AI when I can’t properly explain something, which happens at least a few times a week. Being able to scale that up to nearly every student in the world – at very low or zero cost – is a pretty amazing prospect.

Now, the caveat: Google built the model and ran the study. A company grading its own homework is not an independent replication, so this result isn’t as strong as it would be if we saw it happen a few more times in the future.

But we do have another example – a different AI model produced nearly the same result two years earlier.

Nigeria got there first: 6 weeks, 1.5 years of learning progress

In 2024, the World Bank ran a six-week experiment in Edo State, Nigeria, with about 800 first-year senior secondary (high school) students. Twice a week after school, students worked on English grammar and writing with Microsoft Copilot, which at the time ran on GPT-4. A teacher opened each lesson, stayed in the room, and helped students catch the AI when it was wrong.

The students who got the program outperformed the control group by 0.31 standard deviations overall (0.23 on English alone). The World Bank’s estimate: 1.5 to two years of typical learning progress, in six weeks.

Different country. Different subject. Different age group, different model, different company. Same basic design: a teacher sets the lesson, an AI works one-on-one with each kid, and the teacher supervises. Nearly the same result.

Two studies aren’t proof, yet. But two studies pointing the same direction, with reputable non-profits and international groups involved, is a very impressive trend.

The pattern: AI wins where a human’s attention runs out

A hospital and an eighth-grade math class don’t have much in common, except this: in both places, there’s a person with expertise who cannot pay attention to everyone at once.

A cardiologist can’t watch 16,000 ECGs. A teacher with 50 kids can’t sit next to each one while they fight through a fraction. In both trials, the AI was dropped into the exact spot where human attention ran out. It watched the ECGs. It explained a math concept to the kid. And then the human took over.

My view is that “will AI replace doctors and teachers?” is the wrong question. The better one is: where in this system does a person run out of attention, and what happens if we put an AI there? That’s also the question I’d ask about your own business or career or personal life, before you ask whether AI can do your job.

This is real evidence of positive outcomes from AI, and it’s only going to get more impressive down the road.

Until next time,

Rob Howard
Founder, Innovating with AI

👋 Thanks for being a part of Innovating with AI. You’re in good company – some of our 175,000 readers work at Google, Apple, Microsoft and IBM… plus Canva, Home Depot and Delta.

❤️ Did you love (or hate) today’s newsletter? Reply to this email to tell me. I personally read every single message (without AI).

📬 Share this post (no paywall). If a friend forwarded you this email, subscribe here.