'AI Slop' or 'AI Enhancement'? 106 Hong Kong Students Answered This Question With Real Grades
This post reviews the paper "AI Slop or AI-enhancement? Student perceptions of AI-generated media for an English for Academic Purposes course" by David James Woo, Deliang Wang, and Kai Guo (arXiv:2605.16275, cs.CY / cs.AI / cs.CL / cs.MM, April 2026, 23 pages). Link: https://arxiv.org/abs/2605.16275
Background: what is "AI slop"?
"AI slop" refers to large volumes of low-quality AI-generated content—mass-produced slides, infographics like plastic flowers, looping podcast scripts. In education, the term carries a moral panic. Teachers worry about students using AI for lazy homework, but this study asks the reverse question: what happens when teachers use AI to mass-generate learning materials? Can students tell "junk" from "good stuff," and does it actually help learning?
The experiment
The setting was a real EAP (English for Academic Purposes) course at a Hong Kong community college—106 EFL learners, for whom academic English is a second-language difficulty stacked on a third task. The teacher used Google NotebookLM (a retrieval-augmented generation tool) to generate four types of materials from course content and student work:
- Videos: explanatory videos generated from course content
- Podcasts: audio explanations of course content
- Infographics: visual summaries
- Personalized feedback reports: individualized evaluations of each student's work
Core findings
1. High general acceptance. Students rated the AI materials as useful and easy to use (standard Technology Acceptance Model dimensions). Nobody treated them as "slop."
2. Not all media are equal. Students clearly preferred content tied to assessments (exam-relevant videos) and visual/multimodal forms. Videos and infographics were most popular; podcasts drew lukewarm responses—audio without visual support was hard to follow for academic content.
3. Video preference correlated with grades. Students who said videos were most useful had significantly higher final grades. This is a correlation, not causation—the paper did not experimentally show videos caused improvement—but the association was statistically significant.
4. Cognitive load is the hidden cost. High cognitive load correlated negatively with grades. Some students reported high intrinsic load with infographics (decoding the graphics itself consumed attention); others reported extraneous load with podcasts (simultaneously imagining language, visuals, and academic terms). The core recommendation: AI-generated mass materials must be calibrated in layers to students' cognitive processing capacity. Batch-produced materials without tiered design turn from enhancement into noise.
5. Who benefits most? Some lower-achieving students spontaneously used AI materials as remedial scaffolds—not instructed to, but reported in open-ended comments. In interviews they described re-watching videos repeatedly, treating AI explanations as a "pausable error-correction mechanism": pause, rewind, check the infographic, continue. The authors are cautious about limited sample size, but this is conceptually explosive: AI materials may not uniformly lift all students—they may help most those who need help most, a possible "anti-Matthew effect" tool.
Methodological limitations
1. The slop/enhancement binary is inadequate—a single AI video can simultaneously contain cultural bias and accurate grammar explanation; the binary obscures the hybrid nature of AI content. 2. Perception ≠ learning outcomes. Conclusions rest mainly on student feelings; the only outcome proxy was final grades, and grade-preference correlation cannot be attributed to the AI materials causing improvement. 3. Single RAG baseline. Google NotebookLM was the only tool used. How much of the acceptance is attributable to NotebookLM's design versus RAG in general? Conclusions may not be robust to tool choice. 4. Sample homogeneity. 106 Hong Kong community college EFL students cannot represent global educational contexts, where trust in "authoritative" vs. AI-generated materials may differ.
Verdict
The paper's most valuable contribution is rejecting the crude "AI slop" framing. Students did not treat AI materials as junk—they actively used them as learning scaffolds, and those needing the most help benefited most. The conclusion is not "AI textbooks are good," but that their value depends on design: cognitive layering, assessment alignment, remedial mechanisms. Drop AI into a classroom as a bulk content generator and you get slop; design it deliberately as a personalized learning enhancer and you get enhancement.
Philosophically: perhaps "slop" and "enhancement" are not properties of the material, but properties of how it is inserted.
References
1. Woo, D.J., Wang, D., Guo, K. (2026). AI Slop or AI-enhancement? Student perceptions of AI-generated media for an English for Academic Purposes course. arXiv:2605.16275. 2. Sweller, J. (1988). Cognitive load during problem solving: Effects on learning. Cognitive Science. 3. Davis, F.D. (1989). Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology. MIS Quarterly. 4. Mayer, R.E. (2021). Multimedia Learning (3rd ed.). Cambridge University Press.