The exact text we saved

This is the version exactly as it was saved, character for character. The SHA-256 hash below is a 64-character code worked out from that text. Run the same text through SHA-256 and you get the same code; change a single letter and the code comes out completely different. Each version also stores the previous version’s hash, which chains them together so no version can be swapped out quietly.

TRUVACE RECORD VERSION
record: TRV-2026-2486
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-11-12T07:41:17.267130Z
status: published
lens: p_space
sector: science
headline: How Heavily Are PhD Students Offloading Their Scientific Writing to AI? Evidence from US STEM PhD Dissertations
dek: Large language models are diffusing rapidly into science, raising concerns that delegating research writing to AI may undermine the development of expertise that scientific training is supposed to build. We measure AI-generated writing in the dissertations of a near-census of US STEM PhD graduates from 2019 to 2026 and link graduates to their subsequent employment. AI writing is absent before 2023, present in 29% of dissertations filed in 2026, and rapidly growing. Use is more common among students from non-Engl…
gain_title: (none)
problem_title: A working paper covering about 226,000 US STEM PhD graduates found AI-written abstracts in 29.4% of 2026 dissertations, and US-origin graduates with them were 7.4 percentage points less likely to hold academic jobs in 2026.
trace_subject: (none)
gain_reading: (none)
gain_evidence: (none)
problem_reading: A working paper covering about 226,000 US STEM PhD graduates found AI-written abstracts in 29.4% of 2026 dissertations, and US-origin graduates with them were 7.4 percentage points less likely to hold academic jobs in 2026.
problem_evidence: AI writing is absent before 2023, present in 29% of dissertations filed in 2026, and rapidly growing. | Among US-origin graduates, those with AI-written dissertations are 7.4 p.p. less likely to be employed in academia in 2026
quick_read: A working paper from researchers at Duke and the University of Toronto scanned dissertation abstracts from a near-census of about 226,000 US STEM PhD graduates from 2019 to 2026. AI-generated writing was effectively absent before 2023 and appeared in 1.7% of 2023 dissertations, 5.5% in 2024, 18.8% in 2025 and 29.4% in 2026 (filed through May).

Use differed by background and institution. Among 2026 graduates, 41% of dissertations by students educated in non-English-speaking countries had AI-generated abstracts, against 23% for students from English-speaking countries. At R1 universities the share was 28%, versus 37% at other doctoral institutions.

Graduates whose dissertations contained AI writing were less likely to hold academic jobs in 2026 and more likely to work in industry. The authors stress the link is correlational and that it is too early to say whether AI use weakens research training.
limitation: The authors say the employment analysis is correlational, not causal, and could reflect students who do not want research careers writing lower-effort dissertations; the paper also says the evidence cannot resolve whether AI use degrades research preparation.
tag: Evidence-backed problem
key_points: Detector Pangram flagged only six of roughly 126,000 abstracts filed from 2019 to 2022, implying a false positive rate under 0.01%. | In 2026, AI-generated abstracts were most common in mathematical and computer sciences (36%) and engineering (35%), versus life sciences (26%) and physical sciences (22%). | By field, rates ran from 17% in mathematics and statistics to 48% in computer science and 51% in civil engineering. | Among 2025-26 graduates, 15% of top-ten-ranked programs had AI-generated abstracts, versus 22% at programs ranked 40-50 and 26% at programs ranked 90-100. | Rates were 9% at MIT, 10% at Caltech and 13% at Princeton, compared with 30% or more at the ten highest-rate universities. | Among US-origin graduates of R1 universities, AI-written dissertations went with being 8.0 points more likely to work in industry and 5.4 points more likely to hold non-tenure-track academic posts. | At the current growth rate, the authors project a majority of US STEM dissertations will contain AI-generated content by 2028.
rundown: The sample came from ProQuest Dissertations & Theses Global plus university repositories, matching closely the National Science Foundation's Survey of Earned Doctorates: near-complete coverage in life, physical and mathematical sciences and 85% in engineering. The authors applied the Pangram 3.3.2 detector to all abstracts and to 340 randomly sampled full dissertations to check that abstracts track full-text AI use.

Gaps between language groups widened each year. Non-English-speaking students' rate was 2.8% in 2023 against 0.9% for English-speaking students, 8.3% against 3.7% in 2024, and 28% against 14% in 2025. Within the same doctoral program the gap was 11.7 percentage points in 2025-26, largest for students from Iran (23 points), South Korea (15), China (13) and India (11).

Employment was measured through LinkedIn profiles in 2026 for 2023-2025 graduates of R1 universities, compared within the same university and field. Among US-origin graduates, AI-written dissertations went with a 7.4-point lower chance of academic employment and an 8.0-point higher chance of industry employment, with no difference in government jobs. The differences were much smaller for likely foreign-origin graduates.

The authors say AI use also clusters within programs: a 10-point higher rate among the previous year's graduates of a program went with a 3-point higher chance a student's own writing was AI-generated. They note prior experiments finding generative AI helps disadvantaged writers most, so both benefits and costs may be real.
sources:
- peer_reviewed | NBER Working Paper | https://doi.org/10.3386/w35859 | 2026-10-01
prev: 0000000000000000000000000000000000000000000000000000000000000000

SHA-256 hash

afa866183038ce0f82d8221bb4d6869d5c8da226bc8f9269180afb5e978c5ee7

Previous version’s hash

0000000000000000000000000000000000000000000000000000000000000000