The computer hadn’t changed science because it could make faster calculations, it changed science because it could make possible calculations that couldn’t be done before. AI is changing something more diffuse but much more important: it’s changing what scientists can potentially look for. This is a quite different matter and promise. The advantages of artificial intelligence for research are not only those of automation. We are witnessing a large scale transformation of how scientific understanding is produced, verified, and expanded.
The Scientific Method Is Being Rewritten
Throughout history, the scientific method has generally followed an observable pattern: observation, hypothesis formulation, testing, and eventually modification. While AI doesn’t work outside the realm of this pattern, it does shorten and partially reverse some of the most crucial stages in this process.
Classic research is responsive, you observe something existing, make a hypothesis about it, and develop experiments to verify or dismiss that hypothesis. AI becomes an option for something different, generative simulation. Instead of testing what you see in front of you, you develop what could potentially exist, in a computational sense before a single physical experiment is ever run. This is what is intended when researchers speak of in silico experimentation.
What this means is easy to understand. A hypothesis that would have taken months to prepare wet-lab work for can instead be tested against millions of simulated scenarios in a fraction of the time. The scientist is not excluded. They’re simply involved on a more meta-level.
When Classical Computing Isn’t Enough
Classical computing has its limits. At the molecular level, quantum chemistry – the fundamental theory of how electrons and atoms interact at the scale of the quantum – has computational demands that grow exponentially with the size and complexity of the molecule. That might sound academic. But the upshot is stark: with today’s technology, it is impossible to compute to the precision necessary for many truly interesting molecules. Traditional supercomputers simply can’t scale that far.
For the problems classical computers can’t tackle alone, the solution is likely to be a quantum-classical hybrid approach. Specifically, to solve the hardest problems that classically based machine learning can’t, like electron correlation for complex materials design, you need a synthesis of AI with quantum-informed physics simulation. Most of the interesting molecules mentioned above have these challenges in spades.
That might all sound highly theoretical. But the reality is that people are actively working on these concepts – and have working simulations which marry quantum and machine learning for problems in molecular modeling that classically based hardware can’t resolve. Organizations like https://www.sandboxaq.com are finding that these techniques are making a real difference in fields like materials science, for example, where building the best model of a crystalline structure for the next generation of battery or superconductor requires you to do so at the quantum level.
The Drug Discovery Bottleneck
Eroom’s Law highlights a common pharmaceutical problem where, in contrast to Moore’s Law, over time R&D has become slower and more expensive, while computing power has improved. More investment, longer cycles, and fewer drugs approved. It’s a problem AI can tackle and is one of the highest-impact areas for the application of this technology. The traditional method for searching for and designing drugs was based on high-throughput screening, and it was quite inefficient. It’s very expensive, slow, and error-based. With some mathematical models, you could search around 1 in 1,000, or perhaps 1 in 1,000,000, but never 1 in 10^60 for all the molecules that could possibly exist in the lab. That’s where AI comes in. It can scan a wider range of solutions applying the models it has been trained with. In other words, the brute force you need for exploring such a vast solution space. This restructuring of the process might also bring incremental improvements or even breakthrough solutions. That’s where you want to use AI, where it becomes structural. This new capacity enabled by AI could change the entire industry and the way things are being done.
The Scientist’s Role Doesn’t Shrink – It Shifts
A valid concern about AI in research could be that it replaces human judgment. The evidence does not support that concern, but rather the notion that the human role transforms.
Currently, large language models are trained on scientific literature enabling the synthesis of thousands of papers within minutes, highlighting non-trivial connections between results, and suggesting testable hypotheses that a human researcher might not have been able to formulate in parallel. Active learning entities can even decide on their own which experiments have the highest probability of producing information, limiting the number of low-value experimental runs.
All this automation dismisses the dirty work – such as literature reviews, data cleaning, or negative results that you still had to run to recheck that they were negative. All that remains is the creative and interpretive work. Formulating the right hypotheses. Considering unexpected outcomes. Interpreting the results.
This is not a lesser role. This is a different one.
We have been at these crossroads before, in a sense: every major tool change in science had the same displacement anxiety, eventually prompting acceleration. But this time it’s different. It is not about solving the same problems quicker. It is about being able to handle new problems. And that’s the interesting part.
