Despite the scandal, Dartmouth President Sian Leah Beilock wants more AI. (He wrote an article for The Atlantic in September titled “Why I Want More AI at Dartmouth.”) Crucially, however, school leaders must “hold themselves to the highest standards,” he said. This seems to mean one main thing: disclosure.
In announcing Schnell’s resignation, Beilock wrote that “each of us brings a different perspective on how and when to incorporate AI into our work and our lives. Furthermore, keeping Dartmouth at the forefront of AI inevitably means leading in a fluid and rapidly changing environment. That is why, as leaders, it is so critical that we model the transparency and human judgment we expect from our students and scholars.”
Who wrote this?
What does “disclosure” mean? In a September 2026 article on AI detection tools, titled “Who wrote this?”, Schnell himself goes to great lengths to argue that “written by humans” or “written by AI” is too simple a binary. In his opinion, everything will become a mix of human judgment and AI work, and there is nothing wrong with that. The key for him is that the human author takes full responsibility for the text, regardless of how it was composed and edited.
In the article, Schnell lays out his method for producing it. It is worth quoting at some length:
In September 2026, I asked OpenAI’s GPT-5.6 Pro if an AI-written essay could ever be reliably identified, and I suggested connecting the problem to the Turing imitation game and to the “reverse” Turing test posed by AI text detection. GPT-5.6 Pro connected the question to binary hypothesis testing and statistical decision theory, proposed initial structures and titles, identified candidate literature, and generated portions of first drafts. I chose the problem, the journal, the audience, and the final direction of the argument. I decided which examples and results should remain central, including the decision to retain Liang’s study while confronting it with important counter-evidence. I reviewed the cited literature, rejected or corrected statements that I judged to be inaccurate, corrected mathematical and conceptual errors, and reorganized the argument.
In revising the manuscript, I rewrote most sections and added substantial explanations. From time to time I asked GPT-5.6 Sol to help me with language editing of sections I had rewritten. I created the figures and made final decisions about mathematics, interpretation, sources, organization, and writing. I take full responsibility for the resulting manuscript.
Before submitting the manuscript for publication, I also asked GPT-5.6 Pro and Anthropic Claude Fable 5.1 to critically review it and identify any issues the reviewers might raise. I evaluated the problems they identified and made any corrections I deemed necessary.
Despite all this help, Schnell resists the idea of AI being a “co-author.” Instead, he writes: “Calling [the paper] “co-authored” would introduce a different confusion. Mathematical tools have made substantial contributions for a long time without becoming authors… The point is not that software has not contributed anything. “It’s just that contribution and authorship are different roles.”
