---
id: "lib-2026-10-05-how-heavily-are-phd-students-offloading-their"
type: "library"
title: "How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations (w35859)"
lang: "EN"
org: "NBER (Duke)"
source_type: "Paper"
date: "2026-10-05"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-05-how-heavily-are-phd-students-offloading-their/"
original_url: "https://www.nber.org/papers/w35859"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "NBER working paper by Daniel Gross (Duke) and co-authors measuring AI writing in a near-census of about 226,000 US STEM dissertations and linking graduates to employment records."
summary_basis: "full-text"
original_minutes: "45"
tags: ["future of work", "enablement", "ai writing", "skill formation"]
---

# How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations (w35859)

Tier A (Strong evidence): NBER working paper by Daniel Gross (Duke) and co-authors measuring AI writing in a near-census of about 226,000 US STEM dissertations and linking graduates to employment records.

Summary based on the full text.

## In short

NBER Working Paper 35859 (October 2026) by Daniel P. Gross and Hansen Zhang (Duke Fuqua) and Dror Shvadron (Toronto Rotman). It measures AI-generated writing in nearly all US STEM PhD dissertations from 2019 to May 2026 and links graduates to their later jobs. AI writing is absent before 2023 and present in 29% of 2026 dissertations. Graduates whose dissertations contain AI writing are less likely to work in academic research, which the authors read as a possible sign that AI is replacing work through which expertise is built.

## Key numbers

- 29.4%: 2026 US STEM dissertations with AI-generated abstracts (filed through May 2026)
- 6 of ~126,000: abstracts flagged in 2019-2022, before LLMs, a false positive rate under 0.01%
- 41% vs 23%: AI-written abstracts among students from non-English vs English-speaking countries, 2026
- -7.4 p.p.: lower probability of academic employment for US-origin graduates with AI-written dissertations

## What it reports

- Share of dissertations with AI-generated abstracts: 1.7% (2023), 5.5% (2024), 18.8% (2025), 29.4% (2026). Projected to full text: about 8% in 2024 and 31% in 2026.
- 2026 by field: mathematical and computer sciences 36%, engineering 35%, life sciences 26%, physical sciences 22%; from 17% in mathematics and statistics to 48% in computer science and 51% in civil engineering.
- Students from China 44%, India 43%, South Korea 49%, Iran 51%, against 23% for those with US lower degrees.
- R1 universities 28%, R2 31%, other doctoral institutions 37%. Programs ranked top ten: 15%; ranked 40-50: 22%; ranked 90-100: 26%.
- MIT 9%, Caltech 10%, Princeton 13%, against 30%+ at the ten highest-rate universities.
- A 10 p.p. higher AI rate among a program's previous cohort is associated with a 3 p.p. higher probability for a student.
- US-origin graduates with AI-written dissertations are 7.4 p.p. less likely to work in academia (base 44%) and 8.0 p.p. more likely to work in industry (base 50%); within academia, 5.4 p.p. more likely to be non-tenure-track (base 12%).

## How they got it

About 226,000 graduates come from ProQuest plus university repositories, checked against the NSF Survey of Earned Doctorates. Abstracts were scored with the Pangram 3.3.2 AI detector, with a stratified full-text subsample used to validate abstracts as a proxy. Employment comes from LinkedIn profiles via Revelio Labs; regressions compare students in the same university and field.

## Limits

- The employment link is correlational; students not aiming for academia may simply invest less in the dissertation, though length does not differ.
- Detection covers abstracts, which the authors treat as a lower bound; detector errors after 2022 cannot be checked against a pre-LLM baseline.
- LinkedIn coverage of origin country is about 60%; Pangram provided discounted API credits.

## Why it matters for companies

Industry hires most STEM PhDs, and this paper says a growing share of them wrote parts of their training work with AI. Hiring managers cannot assume a credential carries the same depth of practice it did in 2022. The same mechanism applies to junior staff in any company: if writing and analysis are delegated early, the skills built by doing them may not form. Development programmes need tasks where people still do the hard part themselves and are assessed on it.

Original: https://www.nber.org/papers/w35859