Dartmouth is at a crossroads. Recent reports in Semafor and The Dartmouth have indicated that a top official at our institution, Provost Santiago Schnell, may have produced a number of AI-generated publications, including a Washington Post guest column titled “Universities are fighting AI cheating, but there’s a deeper problem.”
Pangram 4.0, the latest version of the artificial intelligence detector, indicated that Provost Schnell’s Washington Post column was 100% “AI-written.” Provost Schnell has acknowledged that he uses AI for “making writing processes more efficient” but stated that he “conduct[s] the research, evaluate[s] the evidence and sources, ...and decide[s] what appears in the final text” and emphasized that AI does not “substitute for my research, intellectual judgment or authorship.”
In our judgment, definitive conclusions on this issue cannot hinge on proprietary software tools. Thus, we have conducted our own independent assessment of Provost Schnell’s Washington Post column, focusing on its substantive content rather than its stylistic features. We find that Provost Schnell’s column contains substantive errors and omissions that are characteristic of AI-generated text and that would be practically inconceivable in a column written by a distinguished human expert in the field of scientific measurement.
Lack of causality
A systematic limitation of AI tools is the inability to clarify causal relationships. That failure is readily apparent in a central paragraph of Provost Schnell’s Washington Post column: “As a university provost, I initially viewed generative AI chiefly as an academic integrity problem, requiring clear rules about permitted use and ways to identify violations. My scientific work concerns measurement, which led me to a more basic question: What capability does the submitted work actually reveal?” Those two sentences incorporate a shift from academic integrity to capability revelation without providing any coherent explanation of what caused that shift. The “basic question” identified in the latter sentence is practically a non sequitur rather than an implication of what precedes it.
Mischaracterization of scientific evidence
In his column, Provost Schnell discusses prior scientific studies at a superficial level that mischaracterizes key elements. Specifically, he refers to a 2025 study authored by André Barcaui as a “randomized study of 120 students.” The phrase “n=120” appears in the abstract of Barcaui’s article, but the body of the article clarifies that “the primary outcome measure was a knowledge retention test administered approximately 45 days after the learning intervention … Of the 120 participants who began the study, 85 completed the retention test.” Furthermore, Barcaui writes that, “Students who used ChatGPT scored significantly lower on the retention test (57.5% correct) compared to those who studied traditionally (68.5% correct), t(83) = -3.19.” The degrees of freedom — 83 — equals the sample size of 85 observations minus two estimated parameters.
All of Barcaui’s key results are based on those 85 students — not the initial recruited sample of 120 students. Such a distinction is commonly overlooked by AI tools, which tend to focus on an article’s abstract while neglecting subsequent text. In contrast, an expert in measurement familiar with the content of Barcaui’s study would surely refer to the actual number of observations used to produce the key results.
Provost Schnell’s column also makes no qualifications regarding Barcaui’s methodology. In contrast, Barcaui himself noted in the text of his article that his study did not utilize a representative sample: “Participants were undergraduate business administration students recruited through convenience sampling from a large Brazilian university.” An expert in measurement would be fully aware of the limitations of convenience sampling and remiss in omitting such information in characterizing a published study. In contrast, AI tools frequently fail to provide qualifications unless specifically prompted to do so.
In addition to the study by Barcaui, Provost Schnell’s column references a study coauthored by Greg Kestin, Kelly Miller, Anna Klales, Timothy Milbourne and Gregorio Ponti. Both studies were published last year. Provost Schnell omits any reference to subsequent studies, including a working paper posted in December 2025 and a peer-reviewed journal article published in April 2026. Such omissions of more recent and directly relevant work would be extraordinary for a human expert but are commonplace for AI tools.
Dartmouth-specific information
Provost Schnell’s Washington Post column makes a single specific reference to the use of AI at Dartmouth: “Some first-year seminars ask students to compare their own summaries of an assigned essay with an AI-generated version, identify inaccuracies and defend their judgments aloud,” he writes. This sentence practically reproduces a sentence from a published Dartmouth press release, which stated that first-year seminar students “compare an AI summary of an assigned essay with their own and assess the accuracy of the summary, flagging AI hallucinations and defending their choices aloud.” That press release was neither cited nor referenced in Provost Schnell’s column — an omission that is symptomatic of AI-generated text but would be extraordinary if the column had been produced by a human writer who understands the importance of citing paraphrased sources and who seeks to give credit to colleagues here at Dartmouth.
Ethical integrity
The Washington Post requires every guest writer to confirm that a submission was “not created or manipulated with artificial intelligence or editing software.” However, Provost Schnell acknowledged in a statement responding to The Dartmouth's investigation that in drafting his Washington Post column he “used ChatGPT to help refine a few arguments and check for grammatical errors.” Thus, it seems incontrovertible that Provost Schnell gave false attestations to the Washington Post — an ethical breach that is distinct from the extent of his AI usage in producing his column.
Provost Schnell wrote in his Washington Post column that AI can be used to “produce proficient prose without performing much of the intellectual work the prose appears to represent.” Our analysis highlights the deep irony that a column about AI was largely — if not wholly — generated using AI without acknowledgement.
Our analysis clearly shows how the use of AI affects the substantive content of Provost Schnell’s column — not just the stylistic and lexical patterns that can be identified using automated detection tools. However, we have only used this approach to analyze a single column, whereas Pangram 4.0 has been applied to a larger set of documents and indicates that a high proportion of Provost Schnell’s recent publications were AI-generated.
Provost Schnell wrote in his column that he “initially viewed generative AI chiefly as an academic-integrity problem requiring clear rules about permitted use and ways to identify violations.” Indeed, the formulation and enforcement of AI-related rules at Dartmouth may become practically unworkable unless the issues surrounding Provost Schnell’s recent published work are resolved transparently, decisively and promptly.
Michael Herron is the Remsen 1943 Professor of Quantitative Social Science, and Andrew Levin is a professor of economics. Guest columns represent the views of their author(s), which are not necessarily those of The Dartmouth.
This report is solely the responsibility of its authors. AI tools were used in conducting background research but not in the analysis, writing or editing of this report.
