AI Slop or AI-enhancement? Student perceptions of AI-generated media for an English for Academic Purposes course
Paper · 2026 · David James Woo, Deliang Wang and Kai Guo
Summary
The paper asks whether multimedia that a teacher produces with retrieval-augmented generation (RAG) tools actually helps learners, or whether it only fills the course site with low-value "AI slop". In a compulsory English for Academic Purposes course at a Hong Kong community college, the first author mainly used Google NotebookLM. He turned worksheets, other sources and student submissions into videos, podcasts, infographics and individualised feedback reports for 106 EFL students in three classes.
The evaluation used an explanatory sequential mixed-methods design: a survey, then semi-structured interviews, then correlations between survey responses and course scores. The results were read through the Technology Acceptance Model, Cognitive Load Theory and Mayer's multimedia learning theory.
Learners rated the materials as useful and easy to use. They preferred visual content tied to assessments, especially video and infographics. A preference for video went with better grades, while higher reported cognitive load went with lower grades. Some weaker students adopted the materials on their own to catch up. The study matters because it shows an inexpensive path to personalised feedback at scale, as long as complexity is kept under control.
Key points
- Setting: the fall 2025-26 semester, with 106 second-year associate degree students taught by the first author. Only materials made between September and 5 December 2025 were counted. Custom prompts were used for some outputs, for example to strip student names.
- Course structure: 24 lessons of 90 minutes (36 contact hours) and five intended learning outcomes (ILO1–ILO5).
- Grading: CA1 is an adapted IELTS academic writing task 1 (30%). CA2 is a group research report of 3,000–4,000 words (30%). CA3 is an oral presentation of that report (30%). Participation counts for 10%. Scores are normalised per class, so grading is norm-referenced.
- Results: perceived usefulness (PU), perceived ease of use (PEU) and germane load scored high, and mean cognitive load was moderate. Video preference correlated positively with performance, and cognitive load correlated negatively with final grades.
- Limitations noted by the authors:
- small sample and low survey response;
- a cross-sectional survey taken before individualised feedback for CA2/CA3 existed;
- uneven distribution of media across curriculum areas;
- no data on how students actually used the materials;
- no comparison with teacher-made equivalents.
Interesting ideas
- I1. Each generated item was grounded in specific course sources and mapped to the learning outcome it supports, which makes it possible to audit what every artifact is for.
- I2. The study measures two things separately: whether a resource is valued (usefulness, goal relevance) and whether it is cognitively manageable. It then checks each against outcomes instead of trusting a single satisfaction score.
- I3. The authors warn that norm-referenced, per-class-normalised grades can hide or distort relationships with outcomes, so the way the metric is built shapes the conclusions.
- I4. The authors recommend shorter outputs and fewer extraneous details, with language calibrated to the reader's level, following the coherence principle.
- I5. High reported difficulty after the first assessment is suggested as an early warning to identify at-risk learners for direct help.
For Mova
How these ideas could serve Mova, our small open language model — its files, stages and tests:
- I1 (answer): School-course test items and the explanations Mova gives could carry an explicit tag naming the skill or curriculum target they exercise, so gaps in coverage become visible.
- I2 (learn): Evaluation reports could pair a correctness metric with a difficulty measure, such as inference-chain length per question, to see where failures concentrate.
- I3 (learn): Benchmark scoring should use absolute, criterion-style checks rather than relative rankings among model versions, so that real improvement is not masked.
- I4 (back): Generation from the meaning graph could prune peripheral nodes and choose vocabulary according to a target-reader profile, keeping output explainable while lowering load.
- I5 (answer): Mova's student readers could flag early quiz questions with low confidence or long reasoning chains, routing those text types for targeted repair before full evaluation.
Institutions
China159 works · 16,154 citations🎓University of Hong KongPok Fu Lam · China118 works · 13,496 citationsFields, topics and concepts
📚 Shelves 3The field of science this work belongs to (OpenAlex fields).Computer Science · 18,614 worksArtificial Intelligence · 17,965 worksInformation Systems · 1,730 works
🧭 Topics 1Research subjects and languages the work is about.English · 2,379 works
Cite
Cite as David James Woo, Deliang Wang and Kai Guo. 2026. AI Slop or AI-enhancement? Student perceptions of AI-generated media for an English for Academic Purposes course. arXiv:2605.16275.
@misc{woo2026ai,
title = {{AI Slop or AI-enhancement? Student perceptions of AI-generated media for an English for Academic Purposes course}},
author = {David James Woo, Deliang Wang and Kai Guo},
year = {2026},
doi = {10.48550/arxiv.2605.16275},
eprint = {2605.16275},
archivePrefix = {arXiv},
url = {https://doi.org/10.48550/arxiv.2605.16275}
}References
- Artificial intelligence (AI) capabilities have advanced beyond text generation to include automated generation of multimedia content (e.g. audio; images; and video; Wang et al., 2025) and transformation of content through retrieval-augmented generation (RAG), which combines information retrieval with generative models to enhance content accuracy and relevance (Li et al., 2025). Tools such as Google’s NotebookLM combine AI multimedia and RAG capabilities and allow educators to transform content into diverse multimedia at little or no cost (Wang, 2024). The integration of these tools into school and university pedagogical practices could constitute a curricular innovation if the proliferation of multimedia enhances student learning. On the other hand, it may merely lead to AI slop, high-volume, low-quality, AI-generated materials that clutter learning management systems (LMSs) (Jones et al., 2025). find
- Distinguishing enhancement from slop requires attending to at least three theoretical dimensions. First, the Technology Acceptance Model (TAM) predicts that students will engage with supplemental materials only to the extent that they perceive them as useful and easy to use (Granić & Marangunić, 2019). In high-stakes academic pathways, these perceptions may be shaped by whether materials align with assessed Abstract tasks. Second, Cognitive Load Theory (CLT) warns that the volume and complexity AI tools can produce risk imposing unnecessary cognitive demand, exhausting the processing capacity EFL learners need for deeper understanding, even as the germane load that builds schemas and integrates knowledge should be preserved (Sweller, 2011). However, whether such engagement translates into higher grades may depend on how performance is assessed and scored. find
- Third, Mayer’s (1997) Cognitive Theory of Multimedia Learning suggests that formats like video and infographics may lower comprehension barriers by distributing processing across visual and auditory channels, but the coherence principle cautions that extraneous elements can distract rather than support learning. These dimensions interact: a cognitively manageable, visually engaging resource may still go unused if students see it as irrelevant to their grades. On the other hand, a demanding text report may be valued if it delivers individualized, grade-relevant feedback. find
- Student perceptions and performance can signal whether AI-generated multimedia functions as a pedagogical scaffold or AI slop. In the context of English for Academic Purposes (EAP) instruction for English as a Foreign Language (EFL) learners, empirical evidence is needed not only to evaluate effectiveness, but to theorize what AI integration enhances (Kohnke, 2024). Thus, this innovative practice paper reports Hong Kong community college EFL students’ preferences and perceptions of AI-generated multimedia materials developed and implemented by their EAP course instructor (the first author) to supplement instruction. Reporting preferences and perceptions along with learning outcomes, and drawing on TAM, CLT and multimedia learning principles, we argue that the AI-generated multimedia enhanced the course ecosystem in terms of perceived relevance and cognitive accessibility. Importantly, the innovative practice demonstrates how free, RAG-based AI tools enable scalable personalization of feedback that would be less feasible with traditional methods. find
- Materials and Methods Context The practice was implemented during the fall 2025-26 semester in a Hong Kong Community College that delivers two-year associate degree programs. Associate degree students may receive offers for undergraduate programs in Hong Kong universities based on their academic performance in the College. Students at the College include those who were not offered undergraduate program places based on their university entrance exam results from secondary school; and students who were offered undergraduate places, but sought to improve their academic performance and receive more preferred undergraduate admission offers. In that way, the range of learner diversity could be wide. find
Updated
This page in Markdown → https://md.research.elamur.ai/free/en/papers/woo-2026-ai-slop-ai-enhancement-student/