A professor from Berkeley wrote an op-ed discussing students who take shortcuts, but a detector marked her article.

      On August 15, Zvezdelina Stankova published an op-ed asserting that Berkeley admits students who struggle with middle school math. Shortly thereafter, the focus shifted to how the op-ed itself was created. Stankova, a teaching professor of mathematics at UC Berkeley, presented her argument in the San Francisco Standard under the title “I teach calculus at Berkeley. Some of my students can’t do middle school math.”

      The article gained traction quickly, attracting attention from Fox News and Townhall, and became a focal point in the ongoing controversy regarding the University of California’s test-blind admissions policy. Subsequently, Pangram, the tool used by Substack to identify machine-written content, flagged the piece, indicating approximately 33% of it was generated or assisted by AI. Chris Hoofnagle, a professor at Berkeley Law, shared a post about this finding, which garnered nearly three million views on X, shifting the narrative away from calculus.

      Stankova has not refuted the use of technology. In an interview with the Daily Californian, she acknowledged utilizing AI for editing the piece and claimed the article involved “several hundred person-hours of intensive human work and deliberation, of which about 80 hours are my own.” This defense carries more weight than it initially seems. Under most campus policies, editing assistance is fully allowed, and the challenge lies in distinguishing between a tool that revises sentences and one that produces them—an issue universities have struggled to clarify over the past three years.

      However, not everyone at Berkeley was convinced. Hannes Bajohr, a German instructor at the university who writes on machine authorship, told the paper that the situation “seems like deception” or at least somewhat dishonest, particularly in the absence of a disclosure at publication.

      The op-ed itself presents claims that warrant careful examination, and they are significant. Stankova indicates that prior to 2020, when testing was mandatory, 71% of her Calculus I students were ready or nearly ready for the course, while by 2023, that number had fallen to 26%, with the most frequent diagnostic score being zero. She also points to disparities in admissions among Bay Area schools and mentions that five Nobel laureates, including Jennifer Doudna, have signed open letters advocating for the reinstatement of standardized tests. These statistics are drawn from her classroom diagnostics and campaign materials, and the university has yet to issue a rebuttal to either.

      The discomfort in this situation arises not just from the controversy but also from the topic at hand. A discussion about academic rigor, partly formulated with a tool that students are penalized for using, evokes an irony that feels almost fictional. Moreover, it occurs during a semester in which Berkeley’s own computer science faculty have reported increased failure rates coinciding with greater AI usage in coursework, indicating a disconnect within the campus discussions about the same technology.

      The issue of detection remains unresolved. While Pangram is considered one of the more reliable tools in a sparse field, independent researchers have suggested that its false-positive rate may be underestimated, and a 33% result is a probability measure rather than an admission. Detection tools have been asked to evaluate authorship in unusual contexts, such as a project analyzing the Pope’s writings for validation. However, none can yet effectively differentiate between a mildly edited human draft and a heavily prompted machine-generated piece, which is the critical point.

      Faculty at other institutions have taken matters into their own hands concerning enforcement. For instance, a professor at Brown University reverted assessments to supervised, in-person examinations, effectively creating a controlled measure of the extent to which AI was influencing his students’ work.

      Regulatory measures have not progressed. Under the EU’s labeling system, an AI-written article may go unlabeled, while a proofread email may carry a designation, and American universities lack any comparable framework to reference.

      Neither Stankova nor the San Francisco Standard has commented on whether a disclosure will be appended to the article. The argument regarding UC admissions has not been retracted and, for now, stands as the more significant of the two contentious issues.