I see a large number of initial research proposals, sent to me as part of a cold-emailing strategy by prospective PhD students. The vast majority of those proposals are now obviously written by generative AI. For the most part, that makes the proposal largely worthless in terms of judging the quality of potential PhD students. If anything, it is a negative signal about their ability to conduct primary research. If they can't write a short initial proposal without significant help from generative AI, how will they cope with writing research papers and a thesis?
No doubt I am being too harsh in my approach. However, when you receive three or four cold emails per day from prospective PhD students, as I do about this time of year, you have to be a bit ruthless. I know some colleagues that delete those emails unread. I at least want to give these prospective students a chance to capture my attention.
My own experience is based on a sample of one. What does the evidence tell us about the growing use of generative AI by PhD students more generally? We receive an answer to that question in the context of STEM PhD students in this new NBER Working Paper (sorry, I don't see an ungated version online) by Daniel Gross (Duke University), Dror Shvadron (University of Toronto), and Hansen Zhang (Duke University). They collected data on US-trained PhD students from 2019 to May 2026, ran their dissertation abstracts through the AI-detection tool Pangram, and then collected data from each student's LinkedIn profile to determine their post-graduation outcomes. The growth in generative AI writing in dissertation abstracts is striking, as shown in Figure 1 from the paper:
Prior to 2023, almost no dissertation abstracts are flagged as having AI content. By the first five months of 2026, 29.4 percent of abstracts contain AI writing (at least, according to Pangram). Looking across fields, AI adoption has been most rapid in civil engineering and computer science (both around 50 percent in 2026), and lowest in mathematics and statistics (17 percent in 2026).
Gross et al. then dig into the characteristics of the PhD students using generative AI in their dissertation abstracts, finding that:
Non-native English-speaking students use AI far more intensively than native speakers... Among 2026 graduates, 41% of dissertations by students educated in non-English-speaking countries have AI-generated abstracts, against 23% for students from English-speaking countries...
AI use also increases at institutions with lower research intensity and programs of lower rank. In the 2026 cohort, 28% of dissertations at R1 universities have AI writing, compared with 31% at R2 universities and 37% at other doctoral institutions, and this gradient is particularly strong among US-origin students (foreign-origin students write with AI at high rates at all institutions...).
Gross et al. then look at where students go after receiving their PhD, finding that:
Among US-origin graduates, those with AI-written dissertations are 7.4 p.p. less likely to be employed in academia in 2026 (17% over the 44% base rate) and 8.0 p.p. more likely to be employed in industry (16% over a 50% base), with no difference in government employment. These differences attenuate substantially for likely foreign-origin graduates. We also find differences within academia: conditional on academic employment, graduates with AI-written dissertations are 5.4 p.p. more likely to hold non-tenure-track positions (on a base of 12%), with no significant difference in the likelihood of tenure-track or postdoctoral positions.
Gross et al. explicitly caution against taking a causal interpretation of the effect of generative AI use on student outcomes. I didn't interpret the results that way, instead seeing them as perhaps indicating that students with dissertation abstracts written by generative AI may be lower-quality students, or those with less of an academic focus for their future employment. Alternatively, some students may simply be using AI to overcome language barriers, or to polish their writing, rather than to substitute for their own research skills. However, the results don't definitively establish either of those interpretations either.
What the results do show is more banal - PhD students are increasingly using generative AI. Of course, it is important to keep in mind that these results tell us about the prevalence of AI-generated writing in dissertation abstracts, rather than the dissertation more generally. Since generative AI is quite good at summarising, and that is essentially the purpose of an abstract, I would expect AI-generated writing to be more prevalent in abstracts than in the main text of the dissertations.
Finally, while these results apply to STEM subjects, it seems likely that this is indicative of a more general trend across all fields. If this trend continues, it will not be long before AI-assisted dissertation writing becomes the norm rather than the exception. What we should be wondering now is whether that trend will hamper PhD students in their development as researchers, or not.
[HT: Marginal Revolution]
Read more:
No comments:
Post a Comment