AlphaFold is the example most often cited, and for good reason. Predicting how a protein folds was a 50-year-old open problem in biology. AlphaFold solved it well enough that its public database now lists predicted structures for nearly every known protein — over 200 million entries, free for any researcher to use.

For a working biologist, this is the difference between a multi-year experimental effort and a query that returns in seconds. Researchers studying malaria, antibiotic resistance, and dozens of neglected tropical diseases now start with AlphaFold's prediction and work from there.

It is also a useful reality check: AlphaFold is brilliant at one narrow problem. It cannot diagnose your patient, design your drug, or run your trial. The hype around AI sometimes blurs that line — the actual achievement is sharper and more specific.

Most large hospitals in wealthy countries now use AI somewhere in the workflow — most often in radiology (reading scans), pathology (analysing tissue samples), or scheduling and triage. The gains are usually in throughput and consistency rather than in any single dramatic diagnosis.

The more interesting impact is in places where specialists are scarce. AI-based diabetic retinopathy screening is now operating in rural primary-care clinics in India and Thailand, flagging cases that need a referral days or weeks earlier than they would have been caught otherwise. That is meaningful — these are preventable causes of blindness — and it is a different model from the AI-replaces-doctor narrative.

The trade-off worth naming: when an AI system is wrong about your scan, the chain of accountability is murky. Was it the developer? The hospital? The doctor who signed off? Health systems are still working out how to handle this.

The near future is integration rather than revolution. Genomics, wearable data, electronic health records, and imaging will increasingly be analysed together rather than in separate silos. The doctor's dashboard becomes one combined view rather than five different ones.

The deeper bet is on prevention. If a model can detect early signs of cardiovascular disease, kidney decline, or cognitive decline years before symptoms appear, the question shifts from "how do we treat this?" to "do we want to know yet?" — a genuinely new ethical territory for medicine.

The technology will arrive faster than the policy. How insurance, privacy, and patient consent adapt to predictive medicine is one of the most consequential debates of the next decade.

Google's flood forecasting model is the example worth knowing. It uses satellite imagery and river-gauge data to predict floods up to seven days in advance, in regions that often have no national weather service of their own. As of recent deployments it covers more than 80 countries.

The communities most exposed to climate disasters are typically the ones with the least monitoring infrastructure. A working flood warning, sent to a phone with five days of notice, is the kind of intervention that meaningfully changes outcomes — fewer drownings, fewer livelihoods destroyed, more livestock saved.

It is also a good example of the public-good model: open API, free to use, partnered with local agencies. Not every climate AI application is structured this way, and the difference matters.

AI tools are now embedded in over 120 national sustainability programmes — wildfire prediction, illegal-fishing detection, wildlife population tracking, renewable-energy forecasting, building energy optimisation. The aggregate effect is a step change in how much the planet is being measured at any given moment.

That is a real gain, with one caveat: measuring a problem is not the same as solving it. AI does not stop a fire, replant a forest, or reduce emissions. Humans still have to act on what AI shows them, and political will remains the bottleneck.

The energy cost of AI itself is becoming a serious issue. Hyperscale data centres are now significant consumers of electricity and water, and the people building them are increasingly the same people promising AI will help solve climate change. That tension is worth holding onto.

Expect AI-driven sustainability dashboards to become standard at the city and national level. Real-time deforestation alerts, emissions tracking by industrial site, ocean-temperature anomaly detection — all of this will arrive faster than the regulatory machinery to act on it.

Materials science is the area where AI may make the biggest long-term climate contribution. Models that screen candidate chemistries for better batteries, more efficient solar cells, or carbon-capture catalysts can collapse years of lab work into weeks. None of this is a silver bullet — but cumulatively, it matters.

The future worth wanting here is one where AI's environmental benefits clearly outweigh its environmental costs. Whether we get there depends less on the technology and more on who is funding it and what they are pointing it at.

Khan Academy's Khanmigo is one of the more carefully designed examples. Rather than answer student questions directly, it is built to ask Socratic follow-ups — "what have you tried so far?", "where do you think you got stuck?". The model deliberately withholds the answer until the student has shown their reasoning.

That design choice matters. A tutor that gives away answers is useful for getting through homework; a tutor that helps you find answers builds the underlying skill. The technology is the same in both cases; the framing around it makes the difference.

Another worth knowing: Duolingo uses deep learning to analyse pronunciation and grammar, calibrating the difficulty of each session to the individual learner. For languages where human tutors are expensive or unavailable, this works.

Tens of millions of learners now use AI-based tutoring or assessment tools globally. The picture is uneven. In well-resourced schools, AI tends to add another layer to existing instruction. In under-resourced contexts, it is sometimes the closest thing to a tutor a student has ever had.

Accessibility is one of the clearest wins. Real-time captioning for deaf students, text-to-speech for dyslexic readers, translation between dozens of languages, image description for blind learners — these are not future features, they are running today and they are quietly transforming what "inclusive education" can mean.

The cost is harder to measure: when students rely on AI to draft, summarise, and explain, what happens to the underlying skills those activities used to build? Educators are still figuring this out, and the policies vary wildly between schools, even between teachers in the same school.

The classroom of the next decade will probably blend a human teacher, an AI tutor, and AR-style simulations — each doing what it is best at. The teacher handles motivation, judgment, and relationships. The AI handles practice, pacing, and instant feedback. The simulations handle the things books cannot show.

Assessment will change more than instruction. If a model can write a passable essay, "write me an essay" stops being a useful assignment. Teachers are already moving toward oral defenses, process documentation, and in-person work — assessing thinking, not just output.

The deeper question is structural. If AI tutors really do work well, the role of the human teacher shifts — and that is a political question, not a technical one. Who pays for that transition, and what happens to teachers who were not part of the design, will shape education in ways the technology alone cannot.