Grace Huckins has been thinking about the tension between fundamental and applied science since beginning work on a PhD in neuroscience and philosophy at Stanford in 2018. 

“Practical problems are the dimension that you often have to emphasize in grant applications – how will your research help fight diseases, or advance technology, or otherwise improve people’s everyday lives?” said Huckins, who uses they/them pronouns. “But scientists are also really curious people. Even if science had no obvious practical benefits, many of us would still be driven to understand how and why the world works the way that it does.” 

In a recent essay that received the international Nine Dots Prize, Huckins, a lecturer in Stanford’s first-year undergraduate requirement program Civic, Liberal, and Global Education (COLLEGE), argued that unprecedented amounts of data and the increasingly powerful AI tools for analyzing it are making practical advances possible without a corresponding increase in understanding.

“Never before has it made sense to ask whether science is about developing new technologies and interventions or about understanding the universe – for centuries, those two goals have been one and the same,” Huckins wrote in their winning entry. “Now that big data and AI have dissociated science’s two objectives, we have the responsibility to decide which matters most.”

The Nine Dots Prize, named for a lateral thinking puzzle that demonstrates how unseen assumptions limit problem solving, recognizes innovative thinking on contemporary issues and comes with a $100,000 award to support the development of a book to be published by Cambridge University Press.

Huckins shared their thoughts on AI-driven science and humanity’s desire for discovery. 

When did you first begin to think about this tension between how AI is applied to science vs more traditional scientific methods? 

I’ve actually been thinking about this tension since the start of my PhD. I worked in Russ Poldrack’s lab here at Stanford, where I focused on the use of AI or machine learning in neuroimaging research. It’s really common these days to use machine learning to try to predict some attribute of an individual – like personality, intelligence, or psychiatric diagnosis – based on a brain scan. And some of that research does have real practical benefits: If you could perfectly predict the ideal psychiatric drug for a given patient, that would change lives. But I didn’t see the benefit of, say, predicting whether or not someone has depression, since it’s much cheaper and easier to identify depression with a diagnostic interview. And those depression prediction studies don’t actually tell us much about how depression works in the brain, partly because some of those machine learning approaches can be really tough to interpret. It seemed to me like there was a trend in the field of valorizing prediction for its own sake, and that didn’t seem like the best approach.

What opportunities do you see for AI and data to enhance our understanding of the world? 

There are tons of ways, and that’s something that I’m really excited to explore in the book. AlphaFold, the protein-folding AI from Google DeepMind, is a great example. AlphaFold can take in the sequence of amino acids that make up a protein and accurately predict the three-dimensional structure of that protein. That brings real, practical benefits. I was reading a paper the other day about using AlphaFold to design small proteins that can help improve cancer treatment. 

AlphaFold is a huge, ridiculously complicated system, and no one understands how it works. It makes great predictions, but the source of those predictions is a mystery. But those predictions themselves can be grist for the scientific mill. Previously, if you wanted to collect data to test a hypothesis about protein folding, you’d need to do a lot of laborious structural biology work to determine how different sequences of amino acids change protein structure. But now you don’t have to observe those proteins in the real world; instead, you can use AlphaFold as a source for protein structure data. In that sense, tools like AlphaFold could potentially accelerate scientific understanding. 

But today, that understanding is secondary. The predictions and practical benefits come first, and then scientists can use those predictions to gain understanding if they so desire. So a big question that I have is whether understanding might be deprioritized in science, because it’s no longer a key step on the way to many practical benefits. 

How has this inquiry informed discussions in your COLLEGE courses? 

While this book is specifically about AI and science, it’s also more broadly about how AI is forcing us to reassess and redefine so many human endeavors, from writing to visual art to relationships. I’m really curious about what it means to be human in the era of AI, and I think my students are as well – especially because they are living through this transformation at the same time as they are trying to determine what they want for themselves and for their futures. That means they have much richer insight into this question than I do, and I have learned so much from our in-class discussions about AI and how they see it factoring into their lives. For the most part, my students aren’t just AI boosters or AI doomers. They see potential applications for AI that I never would have imagined, and they have the creativity and skill to try to put those applications into practice; at the same time, they also worry about AI changing the way they engage with their education and with the world. That lesson – that AI brings potential and risk in equal measure – is something I’m working hard to reflect in my book.