Some time ago I wrote a blog post where I said 2 things:
- The state of the art in detecting AI-generated-text at that time was not good enough to be reliable.
- Even if reliable, it would not be possible to assume that genAI detection was a sign of research-misconduct because there are lots of legitimate use-cases for generative AI in writing up research.
My view on (1) was changed recently. I tested a genAI detection tool from a company called Pangram, which claims to be ‘genAI detection that actually works’. Pangram estimate their false-positive rate at 1 in 10,000.
I ran the following test:
- ~1,000 random abstracts were retrieved from articles published in 2010 (pre-genAI control group).
- ~1,000 random abstracts were retrieved from articles published in 2025 (post-genAI group).
If the false-positive rate on Pangram’s genAI detection is 1 in 10,000, we should expect that the detections on the pre-genAI group will be zero, or close.
Assuming that’s what we find, given that we don’t know the false-negative rate of Pangram’s software, we can only assume that what we find in the post-genAI group is indeed genAI.
Here are the results:
There are a number of interesting things here.
- That 1 positive detection in 2010 data? That’s a very short piece of text that I should probably have filtered out. The false-positive rate is indeed very low and so far consistent with the 1 in 10,000 claim.
- We don’t know the false-negative rate, so the detection rate shown in the upper left square is best thought of as an estimated lower-bound.
Bluntly, I found these results difficult to believe at first. But they do fit with what I predicted in (2) above. In 2025, GenAI usage is widespread and this preliminary estimate suggests that it is in use in some way in at least a third of the research literature.
That said: there are some limitations to this estimate which I’ll detail below. In my opinion, most of that usage is not indicative of misconduct. Some of it is, but I think that while the false-positive rate on the task of genAI detection is clearly very low, the false-positive rate on the task of misconduct detection will be much higher.
So, that begs the question, is genAI detection useful for detecting research misconduct? I think, despite not being a smoking gun, it’s useful context. If I was editing a research paper, I would want to know if it had been AI-generated in whole or part. So we will make Pangram genAI detection available to our subscribers. We look forward to hearing how you get on with it.
Some nerdy details
- We don’t have exactly 1,000 records in each group. The way that I retrieved the datasets was using the random article function in OpenAlex. The data was retrieved in batches and then filtered to remove duplicates, articles with missing abstracts, or very short pieces of text. I simply told the retrieval process to keep acquiring and filtering batches until the total was >1,000. Then there was some checking of the data and some more filtering.
- I had thought that the best check would be to compare 2021 (the last full year before chatGPT was launched) with 2025YTD. However, on running 2021 data I found a number of positive predictions from Pangram. It occurred to me that various language models, including GPT-2, and GPT-3 were in use at that time for various tasks. There were also language-model-based translation services around at that time. So it’s possible that these were not false-positives. Given the ambiguity, I halted the check on 2021 and switched to 2010.
- I didn’t filter the data for article type. There are several article types here. Most are research papers, but there are also book-chapters, preprints, reviews etc. It might be interesting to look at those separately, but I’ll leave that for another day.
There are a few sources of bias:
- Pangram detects a number of different genAI tools (including GPT-5). There are some data tables showing how accuracy varies depending on which tool is being detected.
- There will be differences between 2010 and 2025 that could cause false-positives. E.g. imagine if false-positives were triggered by papers about deep-learning, or papers about covid-19, there will be more of those in 2025 than in 2010. This kind of (potential) temporal bias in the data is probably the thing I’d be most concerned about because — if present — it would mean that the measurement of FP rate in 2010 is not going to be accurate for 2025.
- Abstracts may be disproportionately AI-generated. I can imagine someone writing a paper and then asking a genAI tool to create the abstract given full text. It isn’t currently clear if there’s a relationship between genAI in full text vs genAI in an abstract, but we will check as we develop the functionality in Oversight.
- In recent years, a number of publishers have removed abstracts from their Crossref data. I think I can guess why, but for whatever reason, this could bias the dataset in some way.
- I broke the data down by publisher and did find some stark differences in the data. Given the small sample size I wouldn’t say it’s conclusive, but some publishers do appear to be putting out more genAI abstracts than others.
- I did take a look at the national breakdown. Affiliations were missing on a large proportion of the articles. Where present, rates of genAI usage were higher in non-native-English-speaking countries. I interpret that as being consistent with genAI usage for language editing and improvement.
Bottom line: if someone wants to do a survey of genAI usage in the published literature, I think there’s a great opportunity to do that with Pangram’s tools.
As a method to detect misconduct, it’s unlikely that genAI detection will be a smoking gun. But I think it will be very useful.
