AI Writing in PhD Dissertations

Sources#

• How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations

Summary#

Gross, Shvadron & Zhang (Duke Fuqua / Toronto Rotman; NBER Working Paper 35859, October 2026) give the corpus its first behavioural trace measured on the doctoral population itself, at national scale. Earlier evidence on that population was attitudinal (the Nature Graduate Survey's comfort items) or came from published papers, whose authors are mostly not students. Their argument for treating writing as a training endpoint is that writing is where ideas must be stated precisely, so it is where gaps in logic and method get found. Delegating it may raise short-run output and cost the student the skill.

Four findings, in the paper's order:

1. Adoption is new and steep. AI-generated abstracts are effectively absent before 2023, then reach 1.7% (2023), 5.5% (2024), 18.8% (2025) and 29.4% (2026, filings through May).
2. Use is stratified. It is heavier among students educated in non-English-speaking countries, at less research-intensive institutions, and in lower-ranked programs. It also clusters within programs and under the same advisor.
3. It predicts leaving academic research. Among US-origin R1 graduates, those with AI-written dissertations are less likely to be employed in academia in 2026 and more likely to be in industry. Within academia they are more likely to hold non-tenure-track posts.
4. The interpretation is open. The authors call the link correlational and give the reverse-causal reading (students not seeking research careers write lower-effort dissertations) equal standing. They still conclude that "we thus far view the evidence as suggesting an adverse impact of AI use on research capability."

Evidence note. empirical, kept. The data are administrative-grade: ProQuest plus scraped university repositories, benchmarked against the NSF Survey of Earned Doctorates, and linked to LinkedIn through Revelio. The weak links are three proxies. AI use is a commercial detector's verdict on the abstract. Origin is the country of a graduate's high school or bachelor's degree. Career is a self-maintained LinkedIn profile. NBER working papers are not peer-reviewed. COI: Pangram supplied API credits and discounted pricing. Disclosure: the authors state that Claude Fable 5.1 drafted the initial version of the discussion of the empirical findings and a subset of the data appendix, with all AI text "substantially manually revised". So the paper's interpretive section is itself partly offloaded writing.

The instrument#

• Sample. U.S. STEM PhD graduates, January 2019 to May 2026: 226,911 dissertations from ProQuest Dissertations & Theses Global (PQDT), extended with scraped repositories for universities missing from PQDT. Coverage is 90–100% of SED STEM graduates in every year through 2024. It is roughly complete in life, physical and mathematical sciences and about 85% in engineering. There are 17 NCSES major fields.
• Detector. Pangram v3.3.2 on every abstract (226,591 scorable). A stratified random full-text sample covers 5 dissertations per field per year for 2023–2026 (340 scored; 339 used).
• Origin. The country of a graduate's high school or bachelor's degree, taken from linked LinkedIn education histories. It is observed for about 60% of the sample. The proxy favors precision over recall, so foreign-born students with US bachelor's degrees count as US-origin, which should attenuate every origin gap.
• Careers. 159,297 of 226,911 (70.2%) link to a Revelio profile. The outcome is any position held in the first half of 2026. Employers are classified into eleven sectors by an LLM ensemble: Claude Sonnet 5 and GPT-5.6 Terra classify, and Gemini 3.5 Flash adjudicates. The ensemble agrees with a blind hand label on 49 of 50 employers. Academic job titles go through the same ensemble and agree with 44 of 50 hand labels, with all six misses in administrative roles that do not enter any outcome. Name matching in the record linkage also uses batched LLM review. Measurement in this paper runs on LLMs at three separate points.

The detector holds up on the two checks that matter most#

False positives. Pangram flags six abstracts in about 126,000 filed 2019–2022, before LLMs were public, which implies a false-positive rate under 0.01%. The concern from Liang et al. (2023, cited by the paper, not in this corpus) was that detectors flag non-native human prose as machine-written. The paper tests it directly: among pre-2023 graduates, abstracts by students from non-English-speaking countries are flagged at a 0% rate, identical to English-speaking peers, on about 18,000 graduates per year. The ESL gap below appears only after 2022 and grows with adoption. That is the right negative control for this detector on this population, though it says nothing about the earlier detectors Liang et al. tested.

Abstract as a surrogate for full text (Table A.1, individual level, n=339). An AI-flagged abstract predicts AI in the full text at +0.556 (base rate 0.153). It predicts the AI share of the full text at +0.111 (base 0.023), and more than 20% AI-written full text at +0.204 (base 0.035). The reverse regressions (Table A.2) carry the same correlation. Reweighting the abstract series by Pr(AI in full text | AI abstract) gives about 8% in 2024 and 31% in 2026, so the headline is a lower bound, though not a loose one. One wording slip: the appendix prose describes a 0.565 coefficient as "56% more likely", but the coefficient is a percentage-point difference.

Diffusion, and the lag behind published science#

The share of dissertations with an AI-flagged abstract, by year (Figure 1A, read from the image and matching the prose):

Dissertations lag published research by one to two years. Liang et al. (2025) and Kobak et al. (2025) estimate that 13–22% of 2024 abstracts in computer science and biomedicine were LLM-modified, against about 5% of 2024 dissertations. Two explanations are mixed in that gap and the paper names only the first. (1) Dissertations take years to write, so 2024 filings were drafted mostly before strong models existed. (2) The estimators measure different constructs. Liang's distributional estimator counts LLM-modified text, which includes polishing a human draft. Pangram classifies text as AI-generated. A construct gap alone would make the dissertation number lower, and the lag would then be partly an artifact.

The authors also extrapolate: "at its current growth rate, a majority of US STEM dissertations will have AI-generated content by 2028." That is a prediction.

Who writes with AI#

By field (2026 cohort, Figures 1B–C). Math/CS 36%, engineering 35%, life sciences 26%, physical sciences 22%. Among major fields, civil engineering 51% and computer science 48% are highest. Mathematics and statistics 17% and chemical engineering 18% are lowest. The 2025 cohort has the same rank order (engineering 24%, math/CS 24%, life 16%, physical 14%). The authors read this as field differences in pace, not in eventual reach. Chemistry and physics abstracts, they note, carry more notation, compounds and quantities. This ordering matches The Solo-Authorship Rebound's trend-break ordering on a different instrument and a different population. Engineering is at the top of both, and chemistry and physics are at the bottom of both.

By language and origin (Figure 2A, 2026). 41% of dissertations by students educated in non-English-speaking countries have AI-flagged abstracts, against 23% for English-speaking origin. By country: Iran 51%, South Korea 49%, China 44%, India 43%, US 23%. The ESL-to-English gap widened each year: 2.8% vs 0.9% (2023), 8.3% vs 3.7% (2024), 28% vs 14% (2025). Within program (university-field fixed effects) the gap is 1.5, 3.7 and 11.7pp in those three years. In 2025–26 it is largest for Iran (23pp), South Korea (15pp), China (13pp) and India (11pp). Canada and the UK do not differ from the US. In the raw Figure 2A bars, Canada (~37%) and the UK (~36%) sit well above the US (~23%), a vault reading of the image. The fixed-effects null therefore means those students are sorted into high-use programs, not that their raw rate equals the US rate.

By institution and rank. In 2026: R1 28%, R2 31%, other doctoral institutions 37%. The gradient is concentrated among US-origin students, while foreign-origin students use AI heavily everywhere (Figure 2B). Within R1s for 2025–26: top-10 programs 15%, programs ranked 40–50 22%, programs ranked 90–100 26%. That is about 0.10pp per rank position, falling to 0.07 once origin is held fixed, so about a third of the gradient is international composition. By university (R1, at least 50 graduates since 2025; Table B.1, reconciled against pdftotext): MIT 9%, Caltech 10%, Princeton 13%, against 28–36% for the top ten. The prose says "30%+" for that top ten, but three of its ten rows (Idaho, Lehigh, Arizona State) are at 28–29%, and the lowest university in the table is UMass Amherst at 4%, not one the prose names. Several program cells, including Berkeley mathematics, Northwestern bioengineering and Ohio State chemistry, have no AI-flagged abstracts at all.

Clustering (Table B.3, reconciled). A 10pp higher AI share among the previous year's graduates of the same program predicts a ~3pp higher own probability (0.305; R1-only 0.277). For the same advisor it is ~2pp (0.183; R1-only 0.172). The authors decline to treat this as exogenous variation: peer norms, shared access and advisor practice could all affect outcomes directly.

The career result#

Design. 2023–2025 R1 graduates and their 2026 employment, regressed on AI-abstract × foreign-lower-degree with university-field and year fixed effects, clustered by university-field. The results hold with program-cohort fixed effects. Sectors are not mutually exclusive.

Tenure-track and postdoc shares are not significantly different. The foreign-origin column is a vault reading of the Figure 3 point estimates. The text says only that the differences "attenuate substantially", which understates the industry row: the foreign-origin industry gap is smaller but its interval clears zero.

Selection checks (Table A.3, reconciled). Graduates with AI-flagged abstracts are 3.0pp more likely to be linked to a LinkedIn profile (base 69.6%). Given a profile, they are no more likely to report a 2026 position (+0.7pp, n.s.). Given a position, they are only marginally more likely to keep it after stale-PhD-position cleaning (+1.0pp, p<0.1). Trimming the AI group by the excess link rate (Lee 2009 bounds) puts the industry gap at +6 to +11pp and the academic gap at −10 to −6pp. Dissertation length, the authors' proxy for low effort, does not correlate with AI use, and the employment gaps persist when it is controlled. That regression is not tabulated.

Why the authors call it conservative. Abstract-only detection is measurement error that biases toward zero. And the sample is R1-only, the group where AI use is lowest and research careers most common.

Reading the career result#

The causal reading is that writing with AI leaves a training deficit. The reverse reading is that students already headed for industry put less effort into the dissertation. The length test is weak against the reverse reading. A student who writes with AI because they no longer plan a research career need not write a shorter dissertation, because AI makes length cheap. So length cannot separate the two readings. What would separate them is career intent measured before writing. The SED asks every graduate about definite postgraduation plans, but intent formed during the PhD is still unobserved. The authors' own leading indicators, independent publication and competitive grants, need more years to accumulate.

The heterogeneity is the paper's most informative result. If AI writing simply marked low effort, the gap should look the same in both origin groups. It is concentrated in US-origin graduates. For students whose English writing cost is highest, AI writing is common and carries less career signal, which fits the paper's both-at-once discussion: AI closes a language gap and removes the thinking that writing forces. Visa-linked career constraints on foreign-origin graduates are an alternative the paper does not discuss. They would compress sector choice regardless of AI use.

Where this is fragile#

1. Detector-defined treatment. "AI writing" means Pangram's verdict on one abstract. Students who polish their own prose with AI and students who generate whole chapters fall into the same bin unless the detector separates them, and the validation sample is 339 dissertations.
2. The 2026 cohort is partial (filings through May) and may be selected on filing timing.
3. Origin is a proxy observed for about 60%, and it attenuates by design.
4. Careers come from self-maintained LinkedIn profiles. Profiles update late, which is why 2026 graduates are excluded and every included graduate has had at least half a year to update.
5. The outcome window is very early. Academic employment two years after a PhD is mostly postdoc and staff positions, and the tenure-track contrast is null.
6. Citation slip. The appendix references give Shvadron et al. (2025a) as "NBER working paper no. 33944", the same number the main references give Agrawal, McHale & Oettl, AI in Science. One of them is wrong. Nothing on this page depends on either.

Connections#

• AI Adoption in Scientific Work — the attitudinal and telemetry instruments on the same profession. This page supplies a matched-population behaviour measure for the comfort-vs-behaviour contradiction recorded there. In 2025, 31.0% of PhD students said they were comfortable with AI writing a research article. Of 2025 US STEM dissertations, 18.8% have AI-flagged abstracts, a behaviour rate inside the comfortable minority. Its language-covariate result (native English speakers +1.047 toward refusal) reappears here as a behavioural gap that survives program fixed effects
• The Solo-Authorship Rebound — the same field ordering from an unrelated instrument: engineering highest and chemistry/physics lowest in both the solo-authorship trend break and dissertation AI writing. That page also finds LLM-modified writing pervasive rather than concentrated, with the steepest regional gradient for non-native-English authors, the same language gradient seen here
• The Tragedy of the Cognitive Commons — the first population-scale outcome linked to offloading inside a training pipeline. It does not measure depletion: what it observes is that graduates who offloaded leave the research track. That is consistent with the commons argument, and equally with self-selection out of it
• Experimental Learning Impact of Generative AI — the randomized evidence that automation-mode use leaves no durable skill once AI is removed. This is the field analogue at high stakes, with the use mode unobserved: an AI-flagged abstract cannot say whether the student used the tool as a tutor or as a ghostwriter
• AI-Exposed College Majors at Labor-Market Entry — the same credential question one degree level up. That page's signal-erosion mechanism (AI makes grades and work samples stop signalling skill) is this paper's closing question, "whether the next generation of scientists is trained or merely credentialed". Here the doctoral defense is the verification institution the authors expect to grow in importance
• Codified vs Tacit Knowledge Exposure — the authors frame research ability as tacit and acquired through "slow, arduous, hands-on work". Writing a dissertation is the codified step through which that tacit understanding is supposed to form, which makes offloading it a direct test of whether the codified/tacit split protects trainees
• Anthropic — the drafting model the authors disclose (Claude Fable 5.1), and one of the three classifiers in their employer and job-title ensemble (Claude Sonnet 5)

Open Questions#

• Does AI writing in the dissertation predict the research-capability markers the authors name as decisive, such as independent first-author publications and competitive grants, once the 2023–2025 cohorts have had three to five years to produce them? A gap that stays in academic-employment share but closes on output, conditional on staying, would favor the self-selection reading. #oq/wait (trigger: 2028–2030 publication and grant records for the 2023–2025 cohorts)
• The reverse-causal reading predicts that the academic-exit gap is already present in career intent recorded before the dissertation was written. Does linking the SED's postgraduation-plans item, or a pre-defense career survey, to Pangram scores shrink the US-origin −7.4pp academic gap toward zero?
• Is the one-to-two-year lag behind published research a timing fact or a construct artifact? Score the same 2024–2026 dissertation abstracts with Liang et al.'s distributional (LLM-modified) estimator and compare it with Pangram's (AI-generated) rate. If the gap closes, the "lag" is polishing that Pangram does not count.

Sources#

• How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations — Daniel P. Gross, Dror Shvadron & Hansen Zhang (Duke Fuqua; Toronto Rotman), How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations, NBER Working Paper 35859, October 2026 (revised October 2026), 30pp including the online appendix, empirical. Not peer-reviewed. Pangram supplied API credits (COI). The authors disclose that Claude Fable 5.1 drafted the initial discussion of findings and part of the data appendix. Cited for the abstract and §§Methods–Discussion, Figures 1–3, and Appendix A.1.4, A.2.1–A.2.3 and B.1–B.2. Parse note: PDF-derived (docling 2.126.0, MLX layout and table stages, 30pp, 7 tables, 8 pictures, confidence excellent). Docling's only warning, table-shift on 13 label cells, is a false positive. All regression rows cited here (Tables A.1, A.2, A.3 and B.3) were reconciled cell by cell against pdftotext -layout on and match. Table B.1 was also reconciled against pdftotext; its top-ten range (28–36%) is looser than the prose's "30%+". Figure-derived values: Figure 1's year and field shares match the prose. The Canada/UK bars (Figure 2A) and the foreign-origin point estimates (Figure 3) were read from the images by the vault and are not stated in the text. Figure 3's notes are split by the parse into the middle of the §Careers paragraph ("…does not correlate with AI use, and Notes: …"), and footnote 1 interrupts the ESL paragraph. No numbers are affected.