Disclaimer: In this post, I neither argue for nor against the use of Claude Code and equivalent LLM-based tools for empirical research. I neither argue here that producing papers at a higher rate is good or bad (this may be a subject for another post). I certainly object to the idea that the rate at which one publishes papers, with or without LLM assistance, is a good indicator for impact, quality, or relevance.
Multiple blog posts and many more social media posts announced something like a Claude code revolution. The increasing capabilities of LLMs, right now notably Claude Code, allow researchers to produce papers at a much higher rate than before, which is expected to lead to a significant increase in paper output that may overwhelm journals (see link to an X post making this claim in this blog post). There are open questions as to the quality and novelty of papers largely produced with LLM assistance. Even if LLMs cannot create “A papers”, however defined, right now, there seems to be agreement that LLM assistance allows one to produce papers that make a contribution to a certain field and have good chances of getting published in a solid journal.
So far, what I have seen seems to implicitly focus on quantitative research in the social sciences, broadly understood, and other disciplines. I think it is important to broaden the perspective because the LLM revolution is likely to have an uneven impact across research methods. For the purpose of this post, I distinguish between quantitative and qualitative methods and QCA.
For quantitative research, it seems safe to assume that adoption of LLM assistance for coding will be nearly uniform. It may also be used to assist the writing process or even write entire drafts of papers (unless journals or publishers impose bans on the latter use, which I don’t think will happen or, if it does, that bans can be sustained). Whether this is for the better or worse of research progress is a separate matter. The consequence for research output is that most quantitative researchers will be able to submit more papers than before. Regarding output alone, this means “the tide lifts all (quantitative) boats.” Some quantitative researchers may be better at using LLMs for research and increase their output more than others compared to the pre-LLM period. I do not see this as different from some researchers having been better able to code in Python or R, so, on average, LLM assistance will accelerate the output of all researchers who adopt it. This holds unless the productivity gains—productivity meant here as writing code or text that does the job in a certain amount of time—lead to the production of more ambitious papers that eat the efficiency gains. This may happen in some cases but is not a uniform trend I expect to materialize.
For qualitative research and QCA, I believe that the revolution is not coming for everyone. In qualitative research, the range is from unequivocal opposition to LLMs to their use for coding (I guess every major qualitative data analysis software now has AI functionalities). I haven’t tried, but I guess that Claude Code could produce a full-fledged qualitative analysis that involves LLM-based coding, so the question is less whether this can be done but whether it is done. I am grossly simplifying matters here: My understanding is that certain philosophical commitments that emphasize reflexivity and human understanding, for example, deny a role for LLMs in research, except maybe for minor purposes such as checking references for accuracy. For researchers who take this position, not much will change with the availability of Claude Code. In contrast, qualitative researchers who are open to full LLM assistance should be able to produce more papers and even books. These qualitative researchers may not be able to keep pace with the acceleration of output of quantitative research, but they are falling less behind than LLM deniers. Within qualitative research, this would create a new divide between scholars who refuse LLMs and produce less output than researchers who rely on LLMs.
For the field of QCA, I think the potential divide is the same as for qualitative research. Technically, the process from calibration to the discussion of results is standardized and follows a template that an LLM should be able to follow because it has been trained with QCA studies. The question for QCA is whether one believes in a strong role of case knowledge and reflexivity in making design choices such as calibration thresholds. If yes, this speaks against LLMs for the data analysis and, possibly, the writing process. If not, then an LLM could largely take over and should be able to run a full study. According to the textbook (like Schneider/Wagemann, Dusa, Oana/Schneider/Thomann, or Mello), QCA should be case-based. To the extent that case knowledge is not codified, for example in country memos in cross-national research, this inhibits the use of LLMs, and the potential for acceleration is limited. My reading of parts of QCA studies is that the role of case knowledge is uneven. Sometimes design choices are made based on data-dependent criteria, such as choosing the mean value of a variable for calibrating the crossover point, or without any justification at all. If such a route is taken, an LLM could play a greater role, and QCA research could be produced at a faster rate. The future will tell: I imagine the field of QCA will be split between adopters and non-adopters of LLMs, similarly to other domains of qualitative research.
There are two reasons why the acceleration (or not) of paper production through LLM adoption is likely to matter beyond self-reflection of methods used in the social sciences. First, researchers who can submit more papers to journals have higher chances of getting published and will publish more eventually. This would not be important if research “productivity” was irrelevant for hiring, promotion, funding, and awards. So far, the quantity of output still seems to matter for many decisions (at some institutions more than at others), implying that the non-adoption of LLMs may have adverse consequences for a career.
Second, the more uneven the adoption of LLM assistance by researchers who use different methods, the more uneven the “playing field” becomes. Let’s take the extreme scenario of all quantitative researchers fully endorsing Claude Code and no qualitative researcher doing the same. The number of quantitative submissions would go up sharply, while the qualitative submissions remain on the same level. Assuming LLM assistance has no negative effects on the research quality of quantitative research, on average, qualitative submissions would find it much harder to get published because they are outnumbered even more than they are at present (at many journals this is the case, I imagine). Besides other potential consequences of the “Claude Code revolution” for the social sciences and journal publishing, I think there is a high probability of disrupting the field because of the uneven adoption of LLM assistance across methods.
