A Detector Flagged 29% of New STEM Dissertation Abstracts. What Does That Mean?

A striking percentage can make an uncertain measure sound definitive. The new NBER working paper on STEM dissertations reports that an AI detector flagged 29.4% of sampled abstracts in 2026. The paper raises real questions about how generative tools are changing academic writing, but its headline number is about what a detector classified in a short text. It is not a census of AI authorship, student intent, or misconduct.

What the researchers counted

The authors analyze a near-census dataset of U.S. STEM PhD dissertations from 2019 through 2026, focusing on abstracts. Their working paper reports that detector-flagged text was largely absent before the generative AI era, then rose to 29% in 2026. Nature’s report on the study describes roughly 226,000 abstracts and a 29.4% rate for the latest period. ^{[1] [2]}

Nature reports that the authors used Pangram version 3.3.2 and manually reviewed a sample of 340 dissertations. That check provides a useful layer beyond an automated label, but it does not turn each flagged abstract into confirmed authorship. The rate is still specific to a detector, threshold, text type and review procedure; a different tool or writing sample could yield a different estimate. ^[2] The unit matters: an abstract is a short, highly conventional summary, not the full dissertation. A flagged abstract does not tell us how much of the dissertation involved AI assistance, whether a student used grammar or translation support, whether the text was revised, or whether a use complied with a local policy. The detector’s label is a measurement produced by a particular tool and version.

The trend is informative, but the label is not ground truth

The study’s time pattern can be useful evidence that writing practices changed after widely available generative models appeared. Yet using a detector to infer authorship introduces uncertainty. Detector systems can confuse predictable, formal or second-language writing with machine-generated text; they can also miss generated text that has been edited. A widely cited evaluation of earlier detectors found substantial false positives for essays by non-native English writers. That study tested different tools and should not be treated as a direct error-rate estimate for the detector used in the dissertation paper. ^[3]

This is why the safe reading is narrower than “three in ten STEM PhDs used AI.” The result is “the chosen detector classified roughly three in ten sampled abstracts as AI-like under the study’s method.” Whether that signal tracks AI assistance, writing style, language background, field conventions, or some combination requires validation against known provenance and human review.

What the paper says about career outcomes

The working paper also reports associations between detector flags and later career destinations, including lower academic research placement and more non-tenure-track or industry roles. Those are correlations in observational data. They do not show that AI use caused a particular career outcome. Differences in institution, field, language background, student resources, labor-market timing, and other factors could affect both writing patterns and employment.

The authors disclose that Claude Fable 5.1 helped initially draft parts of the discussion and a subset of the data appendix, which they then manually revised. That disclosure is relevant context for readers and also illustrates why transparent descriptions of AI assistance are more useful than trying to infer it from prose after the fact. ^[1]

How universities should use this finding

Universities can treat the paper as a prompt to clarify acceptable assistance, improve disclosure norms, and study how writing support affects students. They should not treat a detector score as stand-alone evidence in a disciplinary process. Any consequential assessment needs a transparent procedure, human review, a chance for the student to explain their process, and corroborating evidence beyond a classifier output.

Researchers can strengthen future estimates by validating detector predictions on text with known provenance, reporting tool versions and thresholds, testing subgroup error patterns, and separating abstract-level classification from document-level assistance. Students, in turn, benefit from policies that distinguish editing and accessibility support from undisclosed generation of assessed work.

The paper’s central contribution is a large-scale signal about a fast-changing writing environment. Its practical value depends on preserving the boundary between a signal and a conclusion. A detector can help identify questions worth investigating. It cannot, by itself, answer who wrote a dissertation or whether the work meets academic standards. See our coverage of how AI is changing academic assessment and the risks of learning with AI for the broader campus debate.