Economists Got 5x Faster. But What About Quality?

Opening

Dear reader, something interesting is happening in economics right now. Dartmouth’s Professor Paul Novosad says AI has given him 5x more time to actually think about research questions. ETH Zurich’s Professor Elliott Ash says the productivity gains have him so energized that he wants to work even more.

Meanwhile, on economics social media, warnings are piling up about low-quality AI-generated content — the so-called “AI slop” — with people basically pleading, “please, just stop.” FT columnist Tim Harford tackled both sides of this head-on. I want to take his analysis a step further and argue that this isn’t just an economics problem — it’s a structural shift happening across knowledge production as a whole.

Three Things AI Is Changing: Productivity, Scope, Verification

Harford laid out three paths through which AI could reshape economics. Let’s walk through them one by one.

1. Productivity — “Economists Freed from Grunt Work”

Data cleaning, grant applications, formatting tables — these are the classic chores that eat away at an economist’s time. AI automates a substantial chunk of this. When Novosad says “5x,” he doesn’t mean the research itself is 5x faster — he means the time available for actual thinking has grown 5x.

Here’s the interesting part, though: this productivity boost hasn’t yet translated into visible results. According to Erzo Luttmer, editor of the American Economic Review, roughly 25% of submitted papers now disclose AI use — mostly for editing or coding assistance — but submission quality hasn’t noticeably changed.

Harford himself used an AI agent to analyze abstracts from NBER¹​ working papers. Since ChatGPT’s launch, average sentence length has dropped, but that’s just a continuation of a prior trend — meanwhile, word complexity actually increased. Submission volume at elite journals showed no clear break from existing trends either. Things got faster. They didn’t get better.

2. Scope — “Measuring the Previously Unmeasurable”

A more interesting shift than productivity is the expansion of research scope itself. Qualitative data used to be expensive to collect and hard to analyze systematically. Now economists are using AI to measure the effects of zoning regulations, analyze the impact of difficult job interviews, and mine massive volumes of corporate earnings calls to detect patterns in how firms respond to tariffs.

Progress in forecasting is also worth noting. This past March, the Bank for International Settlements (BIS) released BISTRO, a general-purpose foundation model for macroeconomic time-series forecasting. Built on a transformer architecture²​, this model — unlike traditional econometric models — doesn’t need to be custom-designed for a specific task. BIS reported that the model would have accurately predicted the persistence of 2021–2022 inflation — a period when most traditional models mechanically forecast “reversion to the mean” and missed badly.

3. Verification — “AI Catches Errors, Humans Make Them”

The third possibility is the subtlest, and the most important. Can AI help filter out errors in research?

There’s a tool called Refine.ink, co-founded by Northwestern’s Professor Ben Golub. It’s an AI-based paper review system that systematically detects mathematical errors, gaps in empirical strategy, and logical inconsistencies. According to Golub, it finds problems in at least a third of papers that had already passed peer review at top journals. University of Chicago’s Professor John Cochrane ran his own book on inflation through Refine and called it “the best comments I’ve received in 40 years as an academic.”

Several of economics’ top five journals are already experimenting with Refine. It seems to fit most naturally right before a conditional acceptance — as a final check that catches mistakes an author would otherwise be embarrassed by after publication. (There are plenty of similar services out there. Among the ones I’ve personally tried, I’d recommend https://jenni.ai/, which was built by a Korean founder.)