By Tori Ellison (University of Illinois Urbana-Champaign)
Information
Since 2023 and the widespread adoption of LLM tools by college students (Flaherty, 2025), many STEM educators have observed a shift in the nature of student assignment submissions (Sah et al., 2025). In my data science and machine learning courses, I have observed these shifts both in writing style and in technical and conceptual accuracy. At the level of tone and presentation, student work may feel “off” in ways that are difficult to articulate but misaligned with disciplinary norms. At the level of analysis, instructors may see incorrect, illogical, or misaligned code, interpretations, and conceptual statements. Unlike pre-AI errors, which tended to follow predictable patterns tied to student misunderstandings, these newer errors are often idiosyncratic and harder to diagnose. As a result, it becomes more difficult to determine what a student actually understands and how to assign credit. Over the past three years, I have identified seven recurring types of “AI slop” in student work. These include stylistic indicators (e.g., excessive em dashes), provable unchecked AI output (e.g., leaving prompts in submissions), vague or speculative interpretations (“blindfolded tour guide”), unnecessary or incorrect added information (“TMI”), out-of-scope claims, repeated AI-consistent logic patterns, and unexplained or off-menu code errors. This session presents a course policy that evaluates student work based on its alignment with professional norms rather than attempting to prove AI use. Each type of AI slop is paired with examples and corresponding penalties that reflect the level of professional risk. For example, stylistic indicators may reduce confidence in the work, while submitting unedited AI output may eliminate that confidence entirely. Students are introduced to this typology at the beginning of the course and are expected to revise their work accordingly, regardless of whether AI was used. This approach shifts the focus from detecting misconduct to evaluating the quality, responsibility, and professionalism of submitted work. This work is based on upper-level undergraduate and master’s-level data science courses (25–65 students) at a large public R1 university.