Artificial intelligence already exceeds human limits in several important ways. It can read more text than any scientist could absorb in a lifetime, search enormous spaces of possibilities, perform calculations at machine speed, compare thousands of alternatives, and repeat a task without fatigue.
Yet today's most capable AI systems face a basic limitation.
Their model of the physical world is still built largely from human knowledge about the physical world: books, papers, databases, images, measurements, laboratory records, and other observations that people have already collected and interpreted.
In other words, AI can reason over an extraordinary amount of knowledge—but much of that knowledge has first passed through us.
This helps explain why AI has already shown some of its most dramatic successes in mathematics and computer programming. In these domains, the essential objects already live in an information world. A proof can be expressed symbolically. A program can be executed. A candidate answer can often be checked quickly and automatically. The AI does not need to wait for nature to respond.
Science is different.
If an AI system is to discover something genuinely new about a molecule, a material, a cell, or a physical system, it ultimately needs information from the physical world. If it can only reason from observations that humans have already made, then its scientific horizon remains bounded by the observations available to it.
The next milestone may therefore be much larger than simply building a better reasoning model.
It may be giving AI a way to observe the physical world directly.
Imagine an AI system that does not merely read yesterday's experiment, but decides what to measure next. It sends instructions to an instrument or robotic laboratory, receives the new data directly, revises its model, and designs the next experiment.
At that point, AI stops being only a consumer of human scientific knowledge.
It begins to participate in the process by which scientific knowledge is created.
That transition could become a milestone not only for AI, but for science itself.
Prefer to watch? Continue with the video below, or keep reading.
AI Has Been in Science for a Long Time
There is an obvious objection to all the excitement around AI for Science:
Haven't scientists been using computers and machine learning for years?
Of course they have.
Machine learning has long been used to classify astronomical objects, predict molecular properties, analyze medical images, identify patterns in genomic data, optimize experiments, and approximate difficult physical simulations.
Scientific computing is even older. For decades, computers have solved differential equations, simulated molecules, processed telescope data, and searched enormous parameter spaces.
So the important question is not:
When did scientists begin using computers?
Nor even:
When did scientists begin using machine learning?
The more interesting question is:
When did the computer begin to move from being a tool inside the scientific workflow to becoming an active participant in the workflow itself?
That transition is much more recent.
A Hint of What Was Coming
Protein structure prediction provides one of the clearest early examples.
For decades, scientists knew that a protein's three-dimensional shape is central to understanding what it does. The problem was extraordinarily difficult: a protein may contain hundreds or thousands of amino acids, and the number of possible conformations is immense.
AlphaFold changed the practical landscape of that problem. It showed that AI could do something scientifically consequential that had previously required an enormous amount of specialized human effort.
But AlphaFold was still, in an important sense, a specialized scientific instrument. Give it a particular kind of input and it produces a particular kind of scientific output.
The next step is more ambitious.
What if AI could help decide what problem to investigate next?
From Models to Agents
For much of the recent AI era, we interacted with models one prompt at a time.
Ask a question. Get an answer.
Then something changed.
Models became much better at writing code. They gained tool use. They could search documents, call software, run calculations, inspect the result, notice failures, and try again.
Early scientific agents show that this loop is beginning to reach the laboratory. ChemCrow paired a large language model with chemistry tools for planning and analysis. Coscientist, reported in Nature in 2023, went further: it could search the web and technical documentation, execute code, plan chemistry experiments, and control liquid-handling hardware through experimental APIs.
These were constrained demonstrations rather than general autonomous scientists, but they showed a crucial transition from answering scientific questions to acting inside a scientific workflow.
In other words, the model began to sit inside a loop.
That loop matters enormously for science.
A useful scientific system does not merely need to know facts. It needs to do things in sequence:
\[ \text{observe} \rightarrow \text{reason} \rightarrow \text{propose} \rightarrow \text{test} \rightarrow \text{evaluate} \rightarrow \text{revise}. \]
That looks remarkably similar to scientific work.
The important change is therefore not simply that models have become larger or more knowledgeable.
It is that they can increasingly act, receive feedback, and change what they do next.
Why Mathematics and Programming Came First
Mathematics and computer programming provide unusually favorable environments for AI.
In programming, an AI can write code, execute it, run tests, inspect errors, revise the program, and try again. Feedback can arrive almost immediately.
Mathematics can sometimes provide a similar environment. If a result can be reduced to a formal proof or another objective criterion, a candidate solution can be checked without waiting for the physical world.
These domains therefore support a rapid loop:
\[ \text{generate} \rightarrow \text{verify} \rightarrow \text{revise} \rightarrow \text{generate again}. \]
This matters because intelligence without feedback is limited.
Suppose an AI proposes ten ideas but has no way to tell which one is better. It can produce more ideas, but it cannot reliably improve.
Now suppose every proposal can be tested automatically. The system can generate ten candidates, keep the promising ones, modify them, test again, and repeat.
That is no longer just generation.
It is an engine of improvement.
Science Has a Physical-World Bottleneck
Real scientific discovery is harder because nature does not always answer quickly.
An AI can suggest a drug molecule in seconds. Learning whether that molecule is safe and effective may take months or years.
It can suggest a new material in seconds. Someone still has to make it.
It can propose a new physical mechanism. An experiment may need to be designed and built before anyone knows whether the idea survives contact with reality.
This reveals one of the central asymmetries of AI for Science:
Generating hypotheses may become cheap. Obtaining trustworthy evidence may remain expensive.
That is why the central challenge is not simply building a model that knows more science.
It is building a system that can connect reasoning to observation.
From Human Observations to Machine Observation
Today's AI systems already learn from enormous collections of scientific observations.
But most of those observations were collected, selected, labeled, summarized, or published by humans.
A paper reports the experiment that was performed, not the millions of experiments that were never attempted.
A database contains measurements someone chose to record.
A textbook reflects concepts that generations of scientists have already decided are useful.
This gives AI a powerful but inherited picture of the world.
The next step is different.
Suppose an AI system can directly request a measurement.
It notices uncertainty in its current model and asks an instrument for data in exactly the region where that uncertainty matters most.
The measurement arrives.
The result does not fit the prediction.
The AI revises the hypothesis and requests another experiment.
Now the model is not merely learning from a static archive of human observations.
It is helping decide which observations should exist next.
That changes the role of AI fundamentally.
The Scientific Loop
A useful way to think about AI for Science is as a closed loop:
\[ \text{Question} \rightarrow \text{Hypothesis} \rightarrow \text{Prediction} \rightarrow \text{Experiment} \rightarrow \text{Data} \rightarrow \text{Evaluation} \rightarrow \text{New Hypothesis}. \]
Traditional scientific computing usually occupies one or two parts of this loop.
A simulation predicts what may happen.
A machine-learning model analyzes experimental data.
A database stores known results.
AI4S becomes much more ambitious when the same system begins connecting several steps.
It may read the literature, propose a hypothesis, write the code needed for simulation, analyze the result, select an experiment, interpret the returned data, and decide what to try next.
The closer we get to closing that loop, the more AI begins to look less like a scientific calculator and more like a scientific collaborator.
The Laboratory Starts to Become Programmable
This would not be possible through AI models alone.
The physical laboratory is also changing.
Many scientific instruments are already controlled by software. Robotic systems can perform standardized procedures. Automated platforms can prepare samples, vary parameters, make measurements, and return data with limited human intervention.
In chemistry and materials science, researchers increasingly speak of self-driving laboratories: systems in which models, automation, experiments, and feedback are integrated into a repeated design–make–measure–learn cycle.
This creates an important feedback mechanism.
Better AI makes automated laboratories more useful.
Better laboratory automation gives AI faster experimental feedback.
Faster feedback permits more iterations.
More iterations create better data.
And better data can improve the next round of models.
The result is not simply another scientific software package.
It is the beginning of a possible closed research loop.
But Is That Really Discovery?
This is where the story becomes more interesting.
Suppose an AI reads every relevant paper, proposes a promising molecule, predicts its properties, and sends instructions to a robotic laboratory.
The experiment succeeds.
Has the AI made a scientific discovery?
Perhaps.
But now suppose it cannot explain why the molecule works.
Has it discovered a mechanism—or merely found a successful answer?
Or imagine that the system searches a billion candidate hypotheses and identifies one that predicts the data perfectly.
Did it understand anything?
And who decided that this particular question was worth asking in the first place?
These are not philosophical decorations around AI4S.
They point directly to its technical limits.
Science is more than prediction.
It involves deciding what matters, identifying anomalies, forming concepts, proposing explanations, testing causes, and connecting individual results into larger theories.
Some of those tasks may turn out to be much harder than generating an accurate prediction.
The Scientist's New Colleague
It is tempting to frame AI4S as a competition between scientists and machines.
That may be the least useful way to think about it.
A more interesting possibility is that the unit of scientific work changes.
For centuries, scientific progress depended on individuals, small groups, laboratories, and eventually large international collaborations.
Now another kind of collaborator is entering that system.
It can read faster than any person. It can explore many alternatives in parallel. It does not become tired of repetitive calculations. It can write code, search databases, compare models, and monitor experiments continuously.
But it also lacks things that scientists often take for granted: physical intuition, scientific taste, a reliable sense of what is important, and perhaps the ability to form genuinely new concepts.
The emerging question is therefore not simply:
Will AI replace scientists?
A better question is:
What happens to science when scientists can work with machines that participate in observation, reasoning, experimentation, and discovery?
That is the question this series will explore.
We will look at how AI chooses experiments, why verification becomes the central bottleneck, whether prediction should count as discovery, whether machines can form new scientific concepts, and what may remain uniquely important about human scientific judgment.
Then we will turn to one of the most demanding laboratories for these ideas: the quantum world.
Quantum science adds an unusual twist.
The object being studied is itself information-rich in a way that ordinary classical data are not. Measuring it changes it. Some properties cannot be simultaneously observed. And sometimes converting a quantum experiment into classical data may throw away exactly the information we hoped to learn.
That will eventually lead us to a very different possibility:
not merely AI helping quantum science, but a quantum machine learning directly from quantum nature.
And there, the meaning of “AI for Science” may change once again.
0 comments