Vibe Research

A few months ago, Andrej Karpathy coined the term “Vibe Coding.” The idea is simple: you write code by describing what you want to an AI, it generates the code, and you run it to see if it works. You don’t read the code. You don’t understand the syntax. You just check the output. If the app runs, you move on. If it crashes, you paste the error back and try again.

Karpathy wasn’t being critical. He was being descriptive. This is how a lot of code gets written now, and honestly, a lot of it works fine. The app runs. The feature ships. Nobody asks whether you understood what you built.

I’ve been thinking about what happens when this same pattern migrates to research (this has already begun).

The Messy Middle

I have a folder on my hard drive from my PhD days. It’s a graveyard of Word documents: draft_v2.docx, draft_FINAL.docx, draft_FINAL_revised.docx, draft_FINAL_revised_USE_THIS_ONE.docx. For years, I was mildly embarrassed by this. It felt like evidence of disorganization, of a process that should have been cleaner.

I now think that mess was the work.

Those seventeen drafts weren’t inefficiency. They were the thinking process made visible. Each revision was a confrontation with my own confusion. The document wasn’t just an artifact—it was a forcing function. I couldn’t write “this relationship is significant” until I actually understood why. The friction of putting words down, getting stuck, deleting, rewriting—that friction was doing something.

What was it doing? It was making me internalize the logic. By the time I submitted the paper, I didn’t just have a document. I had understanding. I could defend any sentence because I had fought with every sentence.

Enter the Vibes

Now imagine a different process. You have a research question. You prompt an AI to review the literature. It produces a comprehensive summary with citations. You prompt it to identify gaps. It identifies gaps. You prompt it to draft a methodology section. It drafts one. You read through it, and it looks… fine. It reads like a paper. The structure is correct. The language is appropriate. The citations exist.

You’ve produced an artifact that looks like research. But did you do research?

This is what I’m calling Vibe Research: the production of scholarly output by pattern-matching to the form of scholarship, without the underlying epistemic work.

And I want to be careful here. I’m not saying AI assistance is bad. I’m not saying these tools aren’t useful. I’m saying something more specific: there’s a difference between using AI to augment thinking and using AI to replace thinking while keeping up the appearance of having thought.

The danger is that from the outside - maybe even from the inside - the two can look identical.

The Question I Can’t Shake

Here’s what I keep coming back to: are we solving better problems now, or are we solving worse problems faster?

Because there’s a version of this that’s optimistic. AI handles the drudgery—formatting, literature search, grammar—and frees you to focus on the hard thinking. The ideas. The synthesis. The creative leap that only a human can make.

But there’s another version. AI handles so much of the process that you never develop the capacity for the hard thinking in the first place. You become skilled at prompting and evaluating outputs, not at generating insights. You become an editor of machine text, not a creator of ideas.

I don’t know which version we’re living in. Probably both, in different proportions, for different people. But I notice that I rarely hear people talk about this tradeoff. We talk about productivity. We talk about efficiency. We don’t talk about what we might be losing in the parts of the process we’ve automated away.

The Maker-Checker Problem

In software, there’s a concept called the “compiler.” You write code, the compiler checks it, and it either works or it doesn’t. The compiler is the ground truth. It doesn’t care about your intentions or your prose style. It cares about whether your logic is valid.

Research has its own version of this: peer review. You submit a paper, experts in your field read it, and they check whether your arguments hold. In theory, this is the quality control mechanism. The maker (you) creates; the checker (the reviewer) verifies.

But here’s what I’ve started noticing. When I write with AI assistance, and my reviewer reads with AI assistance, what exactly is being checked?

I submit a paper where AI helped draft sections. The reviewer uses AI to summarize the paper and identify weaknesses. The reviewer’s comments come back to me—and they read like prompts. I feed them back into my AI. It produces responses. I revise.

At what point did a human critically engage with the ideas?

I’m not saying this is happening everywhere. I’m saying the infrastructure for it to happen is now in place, and we have no way of knowing how often it does. The maker-checker system assumes that the checker is doing something fundamentally different from the maker. But if both are prompting the same underlying technology, we’re in a strange loop. An AI writes. An AI reviews. Humans manage the workflow.

The ground truth has left the building.

What Gets Lost

There’s a quote I keep thinking about: “Critical thinking is the habit that keeps you in control.”

The word habit is doing a lot of work there. Critical thinking isn’t just a skill you apply when needed. It’s a disposition. A muscle. And like any muscle, it atrophies without use.

AI is exceptionally good at producing plausible text. It writes with confidence. It doesn’t hedge awkwardly. It doesn’t leave gaps where it got confused. It smooths over the rough edges that, in human writing, signal uncertainty or unresolved thinking.

This is a feature if you’re editing a document. It’s a bug if you’re trying to learn. Because the rough edges were information. They were the markers that said “I haven’t figured this out yet.” When AI polishes those away, you lose the signal that was supposed to trigger more thinking.

I’ve noticed this in my own work. When I write something myself and it sounds wrong, I stop and think. When AI writes something and it sounds smooth, I tend to accept it. The very fluency that makes AI useful is also what makes it dangerous. It speaks to my biases in well-formed sentences. It makes the unexamined assumption sound like the obvious conclusion.

The De-AI Future

Here’s a prediction I’ll probably regret: within five years, there will be a premium on “certified human” knowledge work.

I’m already seeing early signs. There are services that “de-AI” your content—scrubbing text to remove the telltale signs of machine generation. There are journals starting to ask for disclosure of AI use. There are whispered conversations at conferences about which papers “feel” generated.

We’re developing an immune response. But it’s a weird one. We’re not checking whether the ideas are good. We’re checking whether a human had them. Those aren’t the same thing—but increasingly, the former depends on the latter.

If we can’t trust that a human struggled with the thinking, we can’t trust that thinking happened at all. The “proof of work” in academia was never just the paper. It was the process that produced the paper. And if that process is now opaque—if we can’t tell whether it was a human wrestling with ideas or an AI generating plausible text—then the paper itself becomes suspect.

I Don’t Have a Conclusion

I’m aware that I’m writing this essay with the help of the very tools I’m worried about. The irony isn’t lost on me. Maybe that’s the point.

I’m not arguing we should abandon AI in research. I’m not even arguing that “Vibe Research” is always bad. Sometimes you need to produce a document and move on. Not everything needs to be a journey of intellectual self-discovery.

But I think we should be honest about what we’re doing. If you’re using AI to skip the hard parts, you’re not just saving time. You’re making a trade. And the terms of that trade aren’t always clear until much later, when you realize you have a folder full of documents you couldn’t defend, on topics you never really understood.

The mess was the work. I’m not sure we’ve found a replacement for it.