Do we need to reclaim the hypothesis?
Scientific knowledge grows through cycles of conjecture and criticism. But in his recent piece, Hypothesis Retrofitting, Dr. Frezza warns us that big data and AI are disrupting this cycle, pulling science firmly toward data collection and away from hypothesis generation. The end result is scientific papers brimming with data yet devoid of explanation.
Frezza is right to suggest that we risk conflating data collection for discovery. But as I read his essay, I found myself questioning some of the points he raised: that omics encourages HARKing, that we lean too heavily on omics for hypothesis generation, and that AI will accelerate the decline of scientific creativity. These are important concerns, but they are more nuanced than he describes.
Here, I'll examine these claims in turn. I hope that by identifying the real issues at hand, we will be better equipped to resolve them.
Exploratory omics needs post-hoc hypotheses
Frezza warns that omics is turning us into hypothesis retrofitters at scale.
In confirmatory science, that's a problem. It's called Hypothesizing After Results are Known, or HARKing. A helpful analogy is the Texas Sharpshooter: a shooter fires bullets into a barn, then conveniently draws a target around the holes, claiming his brilliance. Post-hoc claims of this kind are much less impressive than genuinely pre-specified predictions (i.e., I'm going to hit the bullseye). They're less impressive because the experiment wasn't designed to test them, inviting false positives and corroding the credibility of the conclusions.
But most omics isn’t confirmatory science. It’s exploratory.
Omics experiments generate vast datasets, revealing metabolites, transcripts, proteins and entire pathways that change with an intervention. And we really care about the surprises. The observations that conflict with, or aren’t accounted for by, our best available knowledge—because they expose a gap in our understanding. And the only way to close that gap is to construct a new hypothesis that accounts for what we didn't expect to find.
Without a new explanation, all you can say is that a particular set of molecules went up or down. But the interesting question is why—which is exactly where the knowledge deficit lies.
In exploratory science then, HARKing isn't an epistemic sin. It's the mechanism by which surprise is turned into understanding—earned through independent, confirmatory tests. The real sin comes when people present exploratory conjectures as confirmatory claims.
Are we relying too heavily on omics for ideas?
Perhaps Frezza's real concern was that we risk relying on omics too heavily for new hypotheses?
Although we tend to call omics assays hypothesis-generating, that label doesn't quite capture their essence. Rather, they're problem-generating. These tools don't point us toward an explanation so much as they expand the set of observations that need explaining. And this can be incredibly helpful, because problems are the substrates of hypotheses.
Seen this way, Frezza's concern must shift: the issue isn't over-reliance on omics for new hypotheses, but for problems.
Sure, a carefully chosen omic assay can reveal interesting problems that targeted approaches would miss, precisely because we tend to focus on familiar regions of the molecular landscape.
But outsourcing the discovery of scientific problems to omics could have severe consequences. Identifying important problems requires wide reading, conceptual synthesis, and integration of disparate evidence. As pressure for new research questions rises, it becomes tempting to defer that labour to the tools. Just spray-and-pray omics. The result is research organised around whichever signal happens to emerge, rather than around questions grounded in deep understanding and clinical relevance.
What's at risk then isn't the capacity to hypothesise. Instead, it's the habits that curate scientific problems and make hypotheses worth having: reading widely, making tenuous connections and holding half-baked ideas in tension.
AI won't replace your hypotheses
The last worry is that large language models (LLMs) will render our creativity obsolete. But this concern mischaracterizes what LLMs do and how they do it.
These models are trained on incredibly large bodies of text and designed to produce coherent, contextually relevant responses through pattern matching—which is why they can generate plausible sounding explanations. But we shouldn't mistake this for understanding. They carry no model of biological reality and no capacity to discern whether an explanation is credible or merely coherent-sounding. More fundamentally, they can't generate novel ideas, only combinations of those that were latent in their training data. We can think of them as sophisticated indexes of existing human knowledge—useful for retrieval and recombination, but not for the kind of explanatory leap that constitutes a truly original hypothesis.
This is why Frezza's concern doesn't quite land. At its most valuable, scientific creativity involves constructing new explanations for phenomena that current frameworks fail to account for. That capacity isn't something LLMs possess.
That said, the candidate explanations LLMs produce could still be useful in practice. Through recombining ideas across fields, they may spit out hypotheses that you hadn't considered before. Most of the outputs will be wrong, trivial, and need triage. But expanding the set of potential explanations—unconstrained by your own priors—might enhance your personal creative process.
The scientist's role in an AI-assisted research environment is therefore not diminished but clarified. It is to identify real-world problems to solve, produce novel hypotheses, and judge which of them deserve to be tested. That judgment requires exactly what LLMs lack: a genuine understanding of the world. Which means it remains, at least for now, irreducibly ours.