In my acceptance remarks at the Duke Med Phys program graduation ceremony, I thanked ChatGPT for arriving at the right time and making our lives as PhD students easier. I genuinely meant it.
Then I started thinking about how AI tools have changed scientific research and efficiency, especially for PhD students.
Doing research in a PhD is about reading many published works, bringing ideas from conferences, defining hypotheses, consulting ideas with advisors and peers, designing research roadmaps, doing experiments, writing codes, analyzing and making sense of results (including failed), rerunning everything again if needed, and eventually producing papers, abstracts, software, or data – with the hope of making some transitional impact.
For me personally, ChatGPT significantly reduced the time spent on redundant coding, finding bugs, and organizing scattered ideas that would otherwise take a very long time. I personally never used it to write an entire abstract or paper, but honestly, as long as the science is rigorous, meaningful, and the scientific creativity remains intact, it shouldn’t matter much.
I remember when many students were shy or hesitant about using ChatGPT, and today industries are asking applicants how they used AI tools in their workflow to improve efficiency.
That shift alone says a lot.
Throwing a research idea here – quantifying the efficiency in PhD research with vs. without the use of AI tools. Very difficult study to perform because I don’t think there are many students doing research without using some form of AI assistance anymore. But maybe we could dare to ask students to do at least one project without using any AI tool. I wouldn’t volunteer for such a study, but I’m sure there are brave souls willing to suffer for science.
Anyway, I came across this article in Nature – Why AI cannot do good science without humans. It mentions a few examples of how AI tools helped accelerate research and development, such as FutureHouse’s Robin, which accelerated parts of a drug-discovery workflow by nearly 200-fold compared to human-only workflows, and Google’s Co-Scientist, working with humans in the loop, generated hypotheses within days that could otherwise take years or decades to formulate.
At first glance, it may sound like AI replacing scientists. But in both examples, the original (training) data came from humans, humans asked the right questions, and humans checked whether the outputs made sense. What people call “prompt engineering” today is basically scientific intuition translated into words.
Could future agentic AI systems transform this completely? Of course. Are we there yet? I honestly don’t think so. But the people investing billions into AI development are not irrational and, of course, not dumb. The direction is clear. AI will fundamentally reshape research, medicine, engineering, professions, and most probably our entire way of life.
But also, replacement is different.
As mentioned in that article and what we see in research, AI is good at identifying patterns, synthesizing information, searching literature, connecting previously unrelated ideas, and accelerating hypothesis generation.
I believe in many ways, scientific discovery is often an accumulation and reinterpretation of previously done work for entirely different purposes. For example, photon-counting detectors in medical imaging.
Photon-counting detectors were used for decades in nuclear medicine applications to bin photons into different energy windows. The idea of binning photons and direct-conversion detectors already existed. What changed was the advancement of ASICs, signal processing algorithms, and computational capabilities, which eventually led to photon-counting detectors becoming viable for high-flux CT systems.
That is how science progresses – not always through new ideas, but through the convergence of previous discoveries with some improvements.
And AI is extremely good at driving this convergence process.
For a human researcher, reading hundreds of papers and forming a coherent conclusion requires enormous time and expertise across multiple fields associated with those papers. AI tools, however, can process that information at a scale humans cannot match. So, it is expected for AI to outperform humans in such tasks involving data aggregation and pattern discovery.
But there is another side to this discussion.
I believe human civilization reached the top of the pyramid partly because of the process of making mistakes and learning lessons from those mistakes. Trial and error shaped science, medicine, engineering, and society itself. Maybe that’s why there are tools that try to humanize AI responses.
Future AI systems may hallucinate less, misinterpret less, and perhaps approach perfection in many domains. But if a machine never fails, struggles, adapts, or learns through consequences the way humans historically have, can it evolve in the same way?
I honestly do not know.
There is a statement in that article – “not every grant proposal needs to include AI”. I completely agree with that statement. Including AI into every proposal simply because it is trending doesn’t automatically make the science better. But whether we like it or not, AI is becoming part of the scientific language itself. I don’t think that means human expertise is being replaced, but the scientists, researchers, and engineers who understand both domain science and AI will likely shape the next generation of discoveries.
And maybe that is the real point. AI is not replacing scientists today. It is changing what it means to be one.
So, the deeper question perhaps is not whether science needs AI but whether science done using AI would still need humanity in the end.
Share your thoughts…