Published:
Artificial intelligence (AI) is transforming software engineering and mathematics first, because answers in these fields are readily checkable: one can determine whether a solution is correct by testing it. The same transformation will reach other fields. I expect people and AI together to make more progress than either can alone. An unguided AI makes many mistakes, but a properly guided one is a powerful tool for gaining knowledge.
This raises the question of what science is for. Science is a tool for advancing civilization and improving people’s lives. It also lets us marvel at the world, work out how things function, and satisfy our curiosity. I have used AI in my own research, and it has reduced tasks that once took weeks or months to 30 minutes or an afternoon. What constitutes responsible use of AI in science remains open, and a stigma around its use likely persists among many researchers.
Where do we go from here, and how do we reach common ground on using these tools well? Universal agreement is unlikely, because the three purposes of science above pull in different directions. A strictly utilitarian view of the scientific endeavor pushes toward unfettered use of AI to make as much progress as possible. Science as an artistic expression in service of one’s own curiosity remains strictly human (at least until the AIs are conscious…). I have used AI in my own research and have seen a clear effect on what I can do scientifically. One of my goals as a scientist is to increase the sum of human knowledge, and with AI I am better able to do that.
As an example, a scientific product I made in one afternoon is available at https://michael-s-phillips.github.io/mars-crater-graph/explorer/, along with the accompanying paper. I built it with Claude Opus 5.5, using the design canvas, Claude Code, and Claude Cowork. During a meeting, I had an idea for a directed graph showing how impact ejecta are distributed across the surface of Mars. I prompted Opus 5.5 with just a few lines: “the relative stratigraphy of craters on Mars: an idea to build a directed graph of all craters on mars to relate them by their relative stratigraphy. have a go at making this a reality”
That was it.
I sent it as an after thought during the meeting because I thought the idea was interesting. After an afternoon of iteration, we had a working prototype: an interactive website showing the directed relationships between craters, and a scientific analysis of what those relationships might reveal about the Martian surface and how erosion could have operated as a function of time on Mars. The product is close to the scope of a master’s thesis. The data wrangling, processing, and derivation of equations needed to build the directed graph reliably would have taken me at least several weeks. A student new to planetary science, data science, coding, and visualization would have needed much longer. The web application adds further work: it runs quickly and reliably, includes useful visuals, documents its methods, and lists all of the crater IDs. Building it by hand would have taken months. Not to mention the scientific analysis that followed, which included landscape evolution modeling, spatial statistical analyses, and an exploration of geometric saturation of the surface against crater production functions for Mars. Oh, and of course the writing. The paper actually reads quite nicely and if I didn’t tell you an LLM generated, you might not have guessed.
I still don’t know what to think about a tool that lets us answer questions about the world we live in, and about other worlds such as Mars, this quickly.
Perhaps luckily, I finished my graduate training before these tools were released into the world. Part of my fear is that students of all ages, from elementary school through professional training and graduate studies, will bypass the struggle that builds expertise. In my experience, learning came from the struggle. Working through math homework builds an understanding of the mechanics and embeds the concepts as intuition. Grasping a subject takes hard work. If that struggle fades, I do not know what follows, and I do not think any of us do. Similar concerns were raised about nearly every technology that has since transformed our lives. Is AI that much different? I worry that it might be. My other concern is that this tool will create silos of individuals who can do whatever they can conceive of in any scientific discipline. I can imagine a future in which I need far fewer colleagues. My AI assistant can now do work for which I would have previously needed colleagues. A year ago, heck even a few months ago, it could not. In many cases it now does that work with more rigor and higher accuracy, and much faster. I am of two minds about this tool. I see its utility, and I marvel at how quickly it helps me answer questions. I also worry for the future of scientific collaborations.
So, I wrote this blog post to explore these thoughts. I do not expect many people to read it. I posted the paper on my website because I do not know what to make of it or what else to do with it. I think it’s good, but it is not really my work. The initial idea was mine, but what was generated was not. I gave a very simple prompt and received extraordinary output—at least I think it’s extraordinary. I cannot publish this with my name as first author, and I do not expect a journal to accept a paper with Claude Opus 5.5 as first author. The paper will live here on my website at least for now. I encourage you to read it and struggle with these ideas as I have.