Scientific Discovery Is Already Becoming a Human-AI Team Sport.

The Useful Question Is No Longer "Can AI Do Science?"

That question is already outdated.

The better question is:

What kinds of science can humans and AI now do together, and how quickly does that change the pace of discovery?

Because the answer is no longer hypothetical.

AI is already helping:

predict protein structures
design new proteins
generate candidate materials
simulate molecular interactions
forecast weather and extreme conditions
compress literature review and hypothesis generation

We do not yet have fully autonomous Nobel-winning robot scientists.

But we absolutely have the early shape of human-AI scientific collaboration.

The Most Famous Example Already Won a Nobel

The most obvious case is AlphaFold.

In October 2024, the Nobel Prize in Chemistry was awarded in part to Demis Hassabis and John Jumper for protein structure prediction, alongside David Baker for computational protein design. The Nobel committee explicitly said AlphaFold helped solve a 50-year-old problem in biology.

That is not a minor workflow improvement.

That is AI becoming part of the historical record of scientific discovery.

According to the Nobel Foundation, AlphaFold2 has been used by more than two million people from 190 countries, and has enabled predictions for virtually all 200 million known proteins.

That changes the economics of biology.

Tasks that once required years of experimental work can now often begin with a strong computational prior. Scientists still need experiments, judgment and domain knowledge. But they are no longer walking into the dark with the same level of ignorance.

That matters enormously.

Drug Discovery Is the Next Obvious Frontier

AlphaFold did not end at protein structure.

Google DeepMind and Isomorphic Labs pushed further with AlphaFold 3, which predicts not just protein structures but interactions among proteins, DNA, RNA, ligands and more.

That is where the scientific value deepens.

Predicting a structure is powerful. Predicting interactions is closer to actually designing interventions.

And that is why the follow-on story has become drug discovery.

Even Nature's February 2026 coverage of Isomorphic Labs' newer proprietary system framed it as a major advance in AI-based drug design - effectively the next wave after AlphaFold 3.

The implication is clear:

AI is moving from understanding biology to helping engineer therapies.

We are not yet at "click a button, cure disease." But we are far past "AI writes a cute summary of a paper."

Materials Science Is Quietly Being Rewired Too

Biology gets the headlines. Materials science may be just as important.

Microsoft Research's MatterGen is a good example of why.

The old workflow for materials discovery often involved screening huge spaces of possible compounds and hoping something useful emerged after expensive simulation and validation.

MatterGen flips that logic.

Instead of searching candidate by candidate, it tries to generate novel materials directly given desired properties.

That is not just a speed-up. It is a different scientific workflow.

Microsoft positions MatterGen and MatterSim as a pair:

one system proposes candidate materials
another simulates whether those candidates behave the way researchers want

That is the larger pattern you should watch in AI science:

not one giant omniscient model, but loops of proposal, simulation, ranking and experimental validation.

That is much closer to how real discovery works.

Weather Science Is Becoming Faster, Cheaper, and More Accessible

The same thing is happening in Earth science.

Google DeepMind's GenCast and WeatherNext 2, Microsoft's Aurora, and the Aardvark Weather work coming out of Cambridge, the Alan Turing Institute, Microsoft Research and ECMWF all point in the same direction:

AI is not just making forecasts a bit prettier.

It is changing the cost structure and speed of scientific prediction.

DeepMind says GenCast beats ECMWF's top operational ensemble system on weather and extreme-condition forecasting up to 15 days in advance.

Microsoft describes Aurora as a foundation model for the Earth system that can match or exceed traditional methods across multiple forecasting tasks while being far faster and cheaper than classic numerical systems.

And the Nature paper on Aardvark Weather goes even further, arguing that an end-to-end data-driven weather system can replace the full numerical weather prediction pipeline at deployment time, while using orders of magnitude less computation.

That is a profound scientific shift.

Why?

Because it means more scientific organizations - including smaller and less wealthy ones - may eventually gain access to forecasting and modeling capabilities that used to require giant institutional compute stacks.

That is how scientific leverage spreads.

What AI Is Really Doing to Science

The easiest way to misunderstand all of this is to imagine AI as a substitute scientist.

That is still too simplistic.

What AI is really doing, right now, is changing the shape of the scientific loop:

reading faster
searching wider
predicting earlier
simulating cheaper
proposing more candidates
compressing dead ends

The human still matters.

The scientist still frames the problem. Chooses the constraints. Decides what matters. Interprets ambiguous results. Designs experiments. Checks reality against model seduction.

But the loop is changing.

A strong human scientist with strong AI tools can already do materially different work from a similarly skilled scientist without them.

That gap will likely widen.

So What Can We Expect Next?

This next part is partly inference, but it is grounded in the systems already shipping.

I think the next stage of AI-assisted science will look like this:

1. Smaller teams with bigger reach

A lab of ten will begin to behave like a lab of fifty in terms of reading, hypothesis generation, simulation, and candidate prioritization.

2. More science becomes software-shaped

As more of discovery becomes prediction, ranking, design, and iteration, the boundary between research and compute tightens.

3. The bottleneck shifts from ideas to verification

If models can generate promising hypotheses cheaply, the new bottleneck becomes:

wet lab capacity
clinical trials
instrumentation
reproducibility
access to real-world data

4. The scientist class expands

This one matters most culturally.

When AI lowers the cost of doing serious reasoning, more people begin to feel like they can participate in science.

Not because the hard parts disappear. But because the entry threshold falls.

That does not make everyone a great scientist. It does increase the number of people who can contribute meaningfully.

What AI Still Cannot Do

It is important not to romanticize too quickly.

AI still struggles with:

experimental grounding
causal uncertainty
data quality
hidden confounders
novelty outside training priors
tacit lab knowledge

And scientific discovery is full of those things.

A model can help propose a molecule, but reality still gets the final vote. A model can forecast weather better, but deployment still depends on institutions and decision-making. A model can generate candidate materials, but actual synthesis and validation still take work.

So no, we are not at autonomous science-as-a-service.

But we are clearly entering something else:

science as a high-speed collaboration between human judgment and machine prediction.

The Bigger Civilizational Point

This is why I think the scientific impact of AI is ultimately more important than the chatbot impact.

Chatbots change interfaces. Science changes civilization.

If AI helps discover better drugs, better proteins, better materials, better forecasts, better energy systems, and faster experimental cycles, then the effect compounds outward into:

medicine
climate adaptation
agriculture
manufacturing
energy
resilience

That is a much bigger story than "AI saves me 20 minutes writing emails."

The Bottom Line

AI is already doing science with us.

Not perfectly. Not autonomously. Not in a way that makes human scientists obsolete.

But very clearly, materially, and already.

The right mental model is not "robot scientist replaces lab."

It is:

the human-AI discovery loop is getting tighter, faster, and more powerful.

And if that loop keeps compounding, the next decade of science may feel less like a sequence of isolated breakthroughs and more like a permanent acceleration.