The emergence of agentic AI1 has radically changed how we can conduct quantitative social science.2 It seems clear that we should use agentic AI in our work—critics can no longer dispute the benefits.
But how should we think about how AI models can help? How can we make sure that we do not underuse or overuse the models? I cannot answer these questions confidently. However, I can describe how I am thinking about these questions.
What We Do (or Perhaps “What We Did”)
In social science, we write papers to help other social scientists understand the world. For quantitative social science, I think of projects as having three parts:
A dataset.
A paper.
A process that takes the researcher from the dataset to the paper.
I’ll focus on this process in this essay.
At our best, the process consists of well-documented computer code that implements our principled decisions (that we carefully defend in the paper).
The process is long. A quick look across several replication archives in political science suggests that a typical project uses one to six scripts and 500 to 2,500 lines of code. Some use many more scripts and many more lines of code.
The process is important. We might even argue that the process is the primary research product. Buckheit and Donoho (1995, 5) write:
An article about computational science in a scientific publication is not the scholarship itself, it is merely advertising of the scholarship. The actual scholarship is the complete software development environment and the complete set of instructions which generated the figures.
So what should this process look like—what features should it have? I suggest three:
Principled: the researcher made reasonable decisions in analyzing their data.
Implemented: the researcher actually implemented those decisions.
Documented: the researcher made it easy to check that they actually implemented those decisions.
We could imagine, for example, that a researcher comes up with a clever research design but fails to code an important variable correctly. Or perhaps they did everything correctly, but they pointed-and-clicked their way to their final estimates and left no way to verify they did anything at all.
I’ve long advocated for what Healy (2020) calls “plain-text social science” because tools like CSV datasets, R scripts, and LaTeX files make it easier to document your process (and thus easier to implement). As Healy (2020, 8) notes:
[S]oftware applications are not all created equal, and some make it easier than others to do the Right Thing.
This leaves us with two uncontroversial claims:
We should make smart choices, implement those choices, and document those choices.
We should use tools that make it easy to do those things.
Enter AI
AI models enter almost seamlessly into this framework, especially when used agentically.
It is indisputable that AI models are tools that can help implement our choices and document those choices. They can write clear and correct R scripts. They can refactor and debug human-written R scripts. They can create careful records. They can document revisions.3
This is clear: at a minimum, we should use AI models extensively for error-checking (Aronow 2026).
Mistakes We Can Make
We can make two types of mistakes:
We can use AI too little.
We can use AI too much (or for the wrong things).
And for PhD students and junior faculty, it is important that they thread this needle.
Too Little
The risks of too little seem clear: we fail to discover and correct errors that AI models would have detected.4
Real examples of errors.
Example 1: A paper describes a prior distribution as “N(0, 2).” From context, “2” refers to a variance. The Stan model uses
normal(0, 2), and Stan treats2as a standard deviation.Example 2: A paper describes a complex algorithm implemented in C++. A subtle but important difference exists between the algorithm described in the paper and the algorithm implemented by the C++ code.
While these errors are nearly impossible for a human to detect, an AI model will almost always discover them.
Too Much
The risks of too much seem less clear. But here is my thinking.
First, we might turn over tasks to AI models in which the human should remain deeply involved. I suspect that current models are not great at things like justifying the substantive credibility of assumptions used for causal inference. I suspect they are not great at finding open and important questions.5
Second, we might use AI models to quickly complete tasks that ought not be completed at all. Thus, rather than making less work, they make more work. It is clear that we can use agents to do lots of tasks very quickly and with many fewer errors. What matters now is that we point those agents towards worthwhile tasks.6
An Attitude for the Future
Some researchers seem optimistic that agents will allow us to work faster, collect more data, or push theories further. Other researchers seem worried about being overwhelmed by slop. My feelings are different.
I’m excited about this: AI agents will allow us to work more carefully.7
And I’m worried about this: AI agents will allow us to pursue all sorts of computational projects that are not worth doing at all.
So here’s an attitude for the future: let’s continue our work. Let’s continue to do work that’s worth doing. Let’s continue to write papers worth reading. And let’s continue to use tools that help us do that work well (even if it slows us down a little).
Such as Anthropic’s Claude Code interface and Opus 5.5 model, as of October 2026.
Perhaps I should write “computational” social science for this essay. That’s what I mean—social science that combines quantitative datasets, data wrangling, and statistical modeling.
They can also help make principled decisions, but work in this dimension is more complicated, I think.
Others seem to think that AI allows us to work faster, incorporate more data, or push further theoretically. Perhaps there is lots of opportunity there, as well.
I don’t worry about researchers overusing models for these things, though. I’m actually optimistic that researchers care about the quality of the papers they put out under their name. I suspect we have (or will quickly develop) good ways to filter out “slop.”
Take the example of data collection. Prior to the emergence of agents, extensive data collection was difficult, expensive, and time-consuming. The high costs made us think carefully about what we wanted and why before we went ahead. The difficulty encouraged and rewarded discernment and judgment. With AI agents, extensive data collection is easy and cheap—you can do it while you sleep. The low costs make it easy to “just go ahead” before we’ve put any thought into the matter.
When I’ve used them best, I feel like agents have slowed me down, not sped me up, without a huge jump in theoretical or empirical ambition.

