This summer I built a pipeline that automates the identification, screening, and extraction of summary statistics from peer-reviewed publications. Several components of it came from a software engineering perspective that was never part of my training: I have a medical degree and a background in biostatistics, not computer science. I built it anyway, and the results looked good. None of that would have been possible without the AI tools we have today.
I was interning in computational biology and human genetics at a pharmaceutical company, where AI is gradually taking over methods that used to define the work: protein–protein interaction prediction, small molecule construction, peptide folding, patient enrollment, literature review. Watching my own pipeline run, my first reaction was not excitement. It was something closer to grief. The reason is simple. The coding skills I spent nearly a decade acquiring can now be reproduced in minutes from a few sentences of instruction. The tech giants tell us that improvements in these models will eventually make learning to code obsolete. But I wonder if that is true, or is it a bubble inflated to raise more money for their own cause?
My second reaction came about an hour later, when I looked closely and found the output was not perfect. Obvious errors had survived the model’s own quality control: mismatched DOIs, missing results, wrong statistics. The grief disappeared entirely once I started reviewing the code by hand. What I found was that the AI had chosen simple but poorly considered solutions to the requests I gave it. Established methods were not applied correctly, because the model had instead invented its own approach based on a web search — a decision buried in the execution and never surfaced to me. The pipeline only worked because I gave detailed instructions after reading almost every line in the output.
The scientific coding layer of our work has always been enormous. As lead statistician on more than thirty cardiology studies, I spent years assembling datasets, verifying variables, and writing data dictionaries so that others could use the data without misunderstanding it. That work required my hours more than my judgment. If that layer shrinks, what remains is the part I was actually trained for: deciding what question is being asked, whether the data can answer it, and what would make the answer wrong. In principle, AI moves us up the stack toward reasoning, and that is a genuine gain in efficiency. It also lowers a real barrier to brainstorming for a new method, a new study design, an analysis you would not have attempted because the implementation cost was a real burden. Plenty of excellent public health projects never made it to practice simply because the person holding the question cannot write the pipeline and cannot find support, methodological or applied, from someone who can.
Without proper supervision, these pipelines fail fluently. Common pitfalls such as miscalculated numbers, mislabeled samples, arbitrarily chosen parameters are prevalent, and forcing through the details of a scientific project without checking them produces repercussions in every possible way. I cannot count how many times an AI has readily admitted its error and then offered nothing useful for getting out of it. Reviewing what AI generates has quietly become the single most important skill in using it for academic or professional work. In public health research, this means students still need hands-on experience in project design, literature review, survey design, statistical reasoning, etc. This is not asking them to produce every result by hand, but to recognize when an output doesn’t add up and to know where to look when it doesn’t. Getting back to the part that concerns me most during my internship, I could not meaningfully review the software engineering portion of my own pipeline. I turned to my colleagues and manager for advice, but they’re computational biologists with little background in Python development. And all we did was spot-check the output, not the code producing it. This is the problem: we cannot guarantee the quality of work done by AI in a domain where we ourselves were not sufficiently trained. Increasingly, that is exactly how we are being encouraged to use it. We may produce a generation that can operate everything through AI and cannot identify when something is off in their own work. As the disputes over data center expansion have made clear, we are building our scientific workflows on infrastructure that a handful of companies can degrade, reprice, or withdraw without their customers’ consent, which makes the ecosystem more fragile than we thought.
Perhaps the most striking news of 2026 is that AI is reshaping the job market so thoroughly that new graduates are finding it far harder to land a position than the cohorts just ahead of them. I cannot argue with the claim that AI is taking over many people’s jobs. But when the tide of a technological revolution comes in, no one survives by standing still. AI will be part of our working lives, and we need to embrace it, assess its benefits, and stay alert to its pitfalls. I believe we can achieve more with AI than without it, so my conclusion is a simple one: trust its transformational power but beware the delusions it invites. Fundamentally, we still have to be able to review what AI does for us.
