This is a short and simple post. I continue to be interested in whether and how we can see inklings of the impact of LLMs on science now, and how to increase that impact over time. My previous post on the topic found that most impact is pretty isolated to math and ML, that a lot of it is that research about LLMs has exploded, and that math papers with theorems don’t seem to have gotten more correct according to LLM reviewers.
However, that post didn’t use any AI detection, so one could have believed that the AI-written papers were getting better on average while others degraded. Thanks to a recent post by Paul Goldsmith-Pinkham, I was able to take a quick pass at something here.
Paul’s post, How do we know if writing is AI?, used Pangram’s AI detection tool to evaluate whether the use of AI in economics (on arXiv and the NBER working paper series) has expanded over time. Indeed, he finds that it has, with his key figure copied below:
So thanks to Paul, we now have a sample of papers tagged as having used AI to some degree. Now we can ask the question (thanks to Paul sharing some of the data with me!): if economics researchers are using AI, are they less likely to post papers that contain things frontier AI models would flag as errors? Personally, I have AI review my drafts now as a detailed copy-editor (and of course there are paid products out there to give even more in-depth reviews); I guess I have some intuition that if that review step is not being done well and carefully then I worry about the underlying AI-augmented science.
Error detection
Quick note on the very AI-coded flow here. I took a random sample of about 200 economics papers per year from 2023 to 2025 and had two GPT-5.6 Terra agents review the papers in extreme detail for errors. They were looking both for internal consistency (e.g., does the description of the model match the equation?) and for accuracy of claims, especially in proofs. The variation across agents, even with a detailed rubric, in what they determined to be an error and how severe they deemed it was large, so we can look at error flagging either as the intersection or the union of the two agents’ findings.
AI-flagged error rates in papers with AI use
Now here are the top-line results (below). In general, no matter how we combine the two agents’ findings ("union" meaning we only require one reviewer to flag the error and "rediscovered" requiring both), a larger share of papers with AI use have at least one AI-flagged error.
Next, I had Terra categorize errors into a few simple groups and flag which papers had content in which those errors could occur. For example, how many papers have theorems and how many of those papers have an error in a theorem? I did the same for numerical reporting and "claims/empirical design" (which relates broadly to either misstating the empirical work done, making important estimation mistakes, or overclaiming result strength).
Finally, we can see if these comparisons hold conditional on some simple paper attributes. There are lots of issues with this comparison from a causal perspective (AI usage probably affects length and content directly), but it is useful to see that the entirety of the difference doesn’t seem to be explained by these simple paths.
These results feel consistent with my previous post and in contrast to recent narratives. There might soon be real acceleration in science, but I am really not convinced that we are seeing it at scale.
What I continue to find over-theorized and under-empiricized is that not every part of science is bottlenecked by the availability of intelligence. We need more research on the mechanics of scientific production and more discussion of the precise ways AI might accelerate those mechanics. Production functions are a weak spot of the economics literature — which has spent a lot of time arguing over statistical minutiae but doesn’t really understand production at the firm or task level — but understanding them is going to be critical to making acceleration happen.