Vibe research, explained (and why I'm optimistic)

Early in 2025, Andrej Karpathy described something he called vibe coding: you tell the model what you want, it writes the code, and you barely look at what comes back. You accept the suggestions. You run the thing. If it works, it works. Karpathy's own phrasing was that you "forget that the code even exists" (Karpathy, 2025). For throwaway prototypes and weekend hacks, that turns out to be a wonderful way to build.

I kept wondering what happens when you point that same instinct at research instead of software. That question is basically what people now call vibe research, and I've spent a couple of nights reading about it.

So what is it, actually

The short version: the human sets the direction and stays accountable, and AI agents do the heavy execution. You decide the question worth asking. The agent goes off and reads the literature, writes the analysis code, wrangles the data, makes the plots, and drafts the writeup. You steer, you judge, you sign your name to the result. Feng and Liu (2026) place this between plain AI-for-science tooling and fully autonomous research, and they keep returning to one point: human accountability is the guardrail that doesn't move.

That last part is where research splits hard from coding. With a prototype you can ship something buggy and fix it next week. You can't (or shouldn't) publish a paper you never actually read. "Accept all" attitude doesn't survive contact with science, and most of the serious writing on vibe research says so plainly (Feng & Liu, 2026). The agent can do the labour. It can't take responsibility for the claim. That stays with you.

A spectrum, not a switch

The thing that finally made vibe research click for me was stopping to think of it as one of three modes rather than a single new tool.

On one end you have AI-assisted research. You're the one doing the work, and AI speeds up specific pieces. You ask it to summarise a paper, brainstorm a few angles, clean up a function, tighten a paragraph. The AI is a tool. You're still the executor. Most people reading this already work this way, often without thinking of it as anything special. Deep Research tools fit here: genuinely powerful, but they wait for your instructions and your filtering. The cleanest way I've seen the line drawn is by role, AI as a tool versus AI as a collaborator (Yang et al., 2026).

On the far end you have auto research, where the system runs the whole loop itself. Idea, code, experiment, paper, even a simulated peer review, with the human barely touching it. Sakana's AI Scientist is the clearest example anyone points to. It really does generate ML papers end to end, for something like ten or fifteen dollars each (Lu et al., 2024).

Vibe research sits in the middle, and the middle is the interesting part. The AI is closer to a collaborator than a tool, but you're still the orchestrator deciding what happens next. You give up some of the raw throughput of full automation in exchange for keeping your judgement in the loop at every turn. Bertina (2025) framed the shift well: in vibe research the human stops being the labourer and becomes the strategist and the critic.

That trade is the whole point. You're moving yourself up the stack.

Why the upside is real

Here's what gets me genuinely optimistic. The productivity numbers are striking, but the bigger thing is the reallocation of attention. Most research time goes into work that isn't the thinking. Chasing down references, reformatting data, debugging a plotting script at midnight, rewriting the same methods paragraph for the fourth time. None of that is where the insight lives. If an agent can carry that load competently, you get more of your hours back for the part that actually needs a human mind: deciding which question matters and whether the answer holds up.

There's a democratisation angle too, and I think it's underrated. A researcher who isn't a strong programmer can run a real computational analysis. Someone trained in one field can poke into an adjacent one without spending six months learning its tooling first. In biology, Owkin's K Navigator is pitched as letting scientists "talk to their data as easily as they talk to a colleague," with claims of multiplying output by as much as twenty times (Owkin, 2025). I take the specific multiplier with a grain of salt, but the direction is right. Lower the cost of execution and more people get to do the exploring.

The real risk: vibe science

The thing that worries me, and plenty of researchers besides, already has a name. Rose et al. (2025) put a sharp label on it: vibe science.

Vibe science is what you get when AI-mediated work comes out fluent, well-structured, confidently cited, and hollow. It reads like rigour. It has the shape of a careful study. Underneath, the method doesn't hold, or the citations don't say what the text claims, or the result is a plausible recombination of things the model has seen rather than anything new and true. The line that stuck with me: AI errors don't just propagate through the literature, they explode through it, because a polished wrong answer gets cited, scraped, and fed back into the next model (Rose et al., 2025).

Sitting underneath vibe science is a problem researchers have started calling verification asymmetry. Checking an AI's output can cost almost as much expertise as producing it would have (Feng & Liu, 2026). Auditing generated code means reconstructing the agent's logic. Trusting a literature summary means knowing the literature. And because the output looks finished, skipping the check is tempting in a way that a messy human draft never is. The coding world already has a name for the same trap: "trust but don't verify." Working code ships fast, nobody reads it, and the bug reaches production (Hienonen, 2026).

There's a reproducibility study I keep coming back to here, REPRO-BENCH, that makes the asymmetry concrete. It asked AI agents to do something that sounds like exactly the kind of grunt work we'd love to delegate: take a paper plus its code, and judge whether the results actually reproduce. The agents were not good at it then. The best structured system landed around 37 percent accuracy (Hu et al., 2025). So the very task you'd most want an agent to handle, the checking, is currently the one it handles worst.

What keeps it real

None of this makes me want to retreat, though. It makes me want to build the guardrails, because they're not mysterious.

Disclosure is the first one. Say what the AI did, at which step, and how you checked it (Zhang, 2026). Several venues already ask for this, and the better vibe research tools bake it in rather than bolting it on. Notez Nerd, for instance, is built around traceable research chains: every claim links back to a source, every tool call is logged, the whole path is auditable after the fact (Notez Nerd, 2026). When the trail exists, an agent's output becomes something you can critique instead of a black box you have to trust.

The second guardrail is keeping your own skills sharp. The researcher who never does the work loses the ability to tell whether the work was done right (Zhang, 2026). That's the real danger to graduate training, and the fix isn't to ban the tools. It's to learn the craft first and delegate second, so you're delegating from competence rather than from ignorance.

The third is just making verification cheap enough that nobody skips it. Citation checks, code review, statistical sanity tests, reproducibility logs. The less the checking costs, the more it actually happens.

Notice that all three already exist in some form. That's why I land on optimism. The failure mode has a name, the guardrails against it are real, and the open question is whether we bother to use them. Vibe science is vibe research with the rigour dropped. Keep the rigour, and I think the next few years of science get a lot more interesting.

Sources

Bertina, A. (2025, May 24). Vibe-research instead of deep-research: AI agents and the future of scientific discovery. Medium. https://medium.com/@abertina/vibe-research-instead-of-deep-research-ai-agents-and-the-future-of-scientific-discovery-4d561248f3e2

Feng, Y., & Liu, Y. (2026). A visionary look at vibe researching. arXiv. https://arxiv.org/abs/2604.00945

Hienonen, M. (2026, April 7). Mitä on fiiliskoodaus ja miten se eroaa AI-avusteisesta ohjelmoinnista [What is vibe coding and how it differs from AI-assisted programming]. Lapin AMK. https://lapinamk.fi/blogiartikkeli/mita-on-fiiliskoodaus-ja-miten-se-eroaa-ai-avusteisesta-ohjelmoinnista/

Hu, C., Zhang, L., Lim, Y., Wadhwani, A., Peters, A., & Kang, D. (2025). REPRO-Bench: Can agentic AI systems assess the reproducibility of social science research? arXiv. https://arxiv.org/abs/2507.18901

Karpathy, A. (2025, February 2). There's a new kind of coding I call "vibe coding" [Post]. X. https://x.com/karpathy/status/1886192184808149383

Lu, C., Lu, C., Lange, R. T., Foerster, J., Clune, J., & Ha, D. (2024). The AI Scientist: Towards fully automated open-ended scientific discovery. arXiv. https://arxiv.org/abs/2408.06292

Notez Nerd. (2026, February 6). Vibe research 2026: The new paradigm of human-AI collaborative research. https://www.noteznerd.com/blog/vibe-research-2026

Owkin. (2025, June 16). Vibing with your research: The future of vibe research for biology. https://www.owkin.com/blogs-case-studies/vibing-with-your-research-the-future-of-vibe-research-for-biology

Rose, I., Celi, L. A., Corti, C., Ellen, J., Fiske, A., Kebede, M., Kwak, G. H., Lunde, T. M., Ordóñez, S. C., Pucher, G., Sadrolashrafi, M., & Sauer, C. M. (2025, August 24). The rise of "vibe science": How AI threatens scientific rigor. TechRxiv. https://doi.org/10.36227/techrxiv.175606229.90717623/v1

Yang, N., Wang, X., & Hyrynsalmi, S. (2026). Positioning vibe researching in software engineering: Dual perspectives from junior and senior researchers. https://doi.org/10.5281/zenodo.18454000

Zhang, Y. (2026). Vibe researching as wolf coming: Can AI agents with skills replace or augment social scientists? arXiv. https://arxiv.org/abs/2602.22401