Research Will Get Better. Is That a Problem?

My father has always been an optimistic person. Whatever the question, he will always say that we will figure it out. In that spirit, I want to write about the role of AI for research, inspired by recent posts by Scott Cunningham and Alexander Kustov, as well as subsequent social media debate on how AI will change publishing. The concern is that AI weakens traditional signals of research quality. If everyone can produce polished writing, it becomes harder to tell good work from mediocre work. I don’t think they are wrong about what they write, but I want to offer an optimistic take on AI and research quality.

My main point is that we need to separate publishing from research quality. I can absolutely see that AI will pose some problems for publishing. I can also see that there may be a temptation to produce slop that will be difficult for journals to deal with. But we must not forget that research will improve in many ways with AI. The optimistic case is that AI lets good ideas shine by lowering language barriers, reducing the need for technical skills, and making implicit knowledge more accessible. That has to be a good thing.

AI tools like Claude Code are tremendously helpful for generating very detailed feedback that can really improve a paper. But if you are a PhD student at a top economics department, you have always had at least some access to extraordinarily smart, experienced people who can read your work and tell you what is wrong with it. Whether that is through seminars, advisor meetings, or random coffee-shop meetings, it doesn’t matter. Whether the feedback concerns modern research methods, what editors are interested in, what a referee would say, or tacit knowledge of how to write good research papers, it does not matter. The main thing is that this kind of feedback is invaluable and extremely important in academia.

But not everyone has access to this kind of feedback. Most researchers operate in places with very limited access to great feedback, and instead have to figure things out on their own. Think about a PhD student at a good but not elite university, or a PhD student with a supervisor who does not have enough time to help out. Or a researcher at a central bank in a smaller country. Or an economist at a government ministry who is trying to write their first academic paper on the side. Most researchers worldwide receive very limited feedback on their work, except for at conferences or in referee reports. Both of those come with their own well-known problems.

Another way to see this is that everyone now has access to a highly qualified research assistant. Research assistance is already widespread in top economics papers. Recent work by Florian Caro shows that roughly 40% of NBER papers acknowledge research assistants, often several per paper. In top journals, it is not unusual to see three to five RAs listed.1 Most researchers worldwide do not have access to anything like this. AI changes that.

And it’s not just about feedback on ideas. Economics is conducted almost entirely in English, but most economists are not native English speakers. I am Swedish and have been writing in English for a long time, but I still make many grammar mistakes. I also use awkward phrasing or write weird things (exhibit A: this post?). In the past, I could use a copy editor for writing advice, but it was costly and not always that great.

AI tools are very useful for overcoming the feedback problem; I have written about using them for feedback. Morten Nyboe Tabor wrote well about AI being like having “a room full of experts while you think”. The optimistic take is that AI will be great for those with a great idea who are not able to communicate it well. That is not a small thing!

Gauti Eggertsson wrote: “We are in the business of figuring out how the world works and generating new knowledge. There is plenty we don’t understand, and no shortage of questions to answer. AI just accelerates the process.” I agree with this sentiment. Tools like Claude Code will almost certainly raise the average quality of research by helping researchers check code, explore ideas, and improve their writing.

The common worry is that this will make it harder to distinguish good papers from great ones. If AI can help everyone produce polished manuscripts, traditional signals such as writing quality or presentation become less informative, and journals may face a flood of competent-looking papers.

That concern is real and seems to be already happening. But it also helps to remember that academic publishing has always adapted to new technologies and rising standards. What counts as a publishable contribution has never been fixed. In the 1980s, you could publish a paper in a top journal that ran country-level regressions with no serious identification strategy. That would be extremely difficult today. The bar went up, people adapted, and research got better. If AI makes polished papers commonplace, the adjustment will likely be similar: the bar will move again. Alex Imas made a similar point on Twitter, noting that AI can now write the types of articles we have written in the past, but that this just means that we will write better papers. What will matter even more is the originality of the question, the credibility of the research design, and whether the results actually teach us something new.

It is also worth remembering that a large share of economically valuable research never appears in academic journals. Work produced at central banks, government ministries, and international organizations shapes policy and public debate without ever passing through peer review. The real test of research quality is not simply whether a paper gets published, but whether other people find it useful enough to read, cite, and build on.

One of my favorite recent papers studies how race and gender are depicted in children’s books. It’s a fantastic paper, published in the QJE. Genuinely interesting on an important topic, and very well executed. Could AI have created this paper? Maybe, but I’m not so sure. What is considered interesting in social science is, after all, determined by humans, and we have peculiar tastes that evolve over time. The real question is whether the underlying ideas are valuable. In the social sciences, this is ultimately determined by whether other researchers find the work useful enough to build on. That market test does not disappear with AI.

In the end, the scarce resource in research is good ideas. If AI makes it easier for more people to develop and communicate those ideas, that is more likely to improve research than to degrade it.

This post contains mostly my idea. I benefited from discussions with some colleagues here at SAFE. I also used AI tools to help sharpen some of the discussion — it gave me nice feedback on some points. That’s good!