AI-Generated Slides: Are They Good? Can Students Tell?

Juho Leinonen, Lisa Zhang, Arto HellasarXiv 2026 · arXiv

TL;DR

This paper explores how generative AI can assist educators by automatically transforming online textbook chapters into lecture slides. The authors find that developer-focused coding assistants (like Cursor and Claude Code) vastly outperform general-purpose LLMs in creating accurate, pedagogically sound slides. Furthermore, when deployed in a real university classroom, students rated the AI-assisted slides as equal in quality to instructor-made ones and were entirely unable to tell them apart.

What Problem Does This Paper Address?

Preparing high-quality lecture slides from existing course notes or textbooks is a notoriously time-consuming task for instructors. While Generative AI (GenAI) has shown promise in creating basic educational content, generating slides is uniquely challenging because it requires coordinating text, code snippets, and visual structure without overwhelming the learner.

This paper addresses two major questions:

  1. Are they good? Can current GenAI tools generate slides from technical text that are factually accurate, complete, and pedagogically sound?
  2. Can students tell? How do students perceive the quality of AI-generated slides compared to human-generated ones in a live classroom setting, and can they accurately identify the source?

Methods

The researchers conducted a two-phase evaluation within an upper-year undergraduate Web Software Development course:

  • Phase 1: Instructor Evaluation of Tools. The authors tested five accessible GenAI tools representing different workflows: NotebookLM (education-focused), M365 Copilot and Claude (general-purpose LLMs), and Cursor and Claude Code (developer-focused coding assistants). Using simple, iterative prompts, they tasked each tool with converting markdown textbook chapters into a 10–15 minute lecture segment.
  • Phase 2: In-Class Student Evaluation. The best-performing tools from Phase 1 were used to generate slides for 10 lecture segments across a 5-week period. To ensure a fair test, the instructor used the AI tools to match their personal visual styling (fonts, code formatting, transitions) and pruned redundant content. Students were informed that some content might be AI-generated but were blinded to which segments. After each segment, students used a 7-point Likert scale to rate the slide quality and guess whether the slides were human- or AI-made.

Datasets

While this paper does not introduce a machine learning dataset, it utilizes and evaluates:

  • Course Materials: Markdown source files from an online interactive textbook for a Web Software Development course covering HTML, Svelte, state management, CRUD operations, and the Fetch API.
  • Evaluation Responses: A total of 459 distinct student evaluation responses collected across multiple lecture segments to gauge quality and origin guesses.

Key Insights and Ideas

  • Coding Assistants Make the Best Slide Designers: Counterintuitively, tools built for software engineering (Cursor and Claude Code) produced far better educational slides than general-purpose or education-specific tools. Because these tools excel at handling structured text formats (like Quarto or Markdown) and code blocks, they maintained logical progression and technical accuracy.
  • The "AI Bias" Illusion: The study revealed a significant negative correlation between perceived quality and the student's guess that a slide was AI-generated. If a slide had a minor formatting glitch or felt dry, students assumed it was AI-made—even if it was actually designed by the instructor. Conversely, if an AI slide included a human touch (like a popular webcomic), students confidently guessed it was human-made.
  • The "Prune and Style" Workflow: GenAI tools are currently poor at visual layout and styling but excellent at structural drafting. The authors suggest that the most efficient workflow for educators is to let a coding assistant generate the raw content structure in a text-based slide format (like Quarto/HTML) and then manually refine the aesthetics and delete excess slides.

Key Results

  • Tool Comparison: In the qualitative analysis, Cursor and Claude Code achieved "Accurate", "Complete", and "Strong" pedagogical ratings, whereas general LLMs frequently hallucinated, missed key context, or inserted irrelevant imagery (like random pictures of birds).
Evaluation of GenAI tools
  • Perceived Quality: Quantitatively, students rated the overall quality of AI-generated slides identically to human-generated slides, with both receiving a median score of 5.0 out of 7.
  • Detection Accuracy: Students could not reliably distinguish between human and AI authorship. The median guess for both categories sat dead-center at 4.0 (neutral).
  • Statistical Correlation: A Spearman correlation analysis ($\rho = -0.28, p < 0.01$) confirmed that lower quality ratings strongly correlated with a student guessing the slides were AI-generated, regardless of the actual source.
Table 2: Overall quality and human vs. AI guesses by segment on a 7-point Likert scale. For quality, 1 = very bad, 7 = excellent, for human vs. AI, 1 = definitely human, 7 = definitely AI, and for source AI-C = Cursor, AI-CC = Claude Code. Only students who reported attending the lecture are included. Also Spearman correlation between overall quality and human vs. AI guess.

Limitations and Future Work

The authors note several limitations to their case study. First, the experiment took place within a single web development course; coding assistants might not perform as well when generating slides for non-technical disciplines (e.g., humanities or qualitative social sciences). Second, the study measured student perception of quality, not actual learning outcomes or knowledge retention.

Future research is needed to evaluate whether AI-assisted workflows impact student performance and to explore more advanced multi-agent AI frameworks that can handle visual layout design automatically.

Summary updated

Abstract

As generative AI (GenAI) tools become easily accessible, there is promise in using such tools to support instructors. To that end, this paper examines using GenAI to help generate slides from instructor authored course notes, emphasizing instructor and student perceptions. We examine an end-to-end education tool (NotebookLM), two general-purpose LLMs (Claude, M365 Copilot), and two coding assistants (Cursor, Claude Code). We first analyze whether GenAI generated slides are ``good'' via narrative assessment by educators. We choose the best slides to use (with some modification) in a real course setting, and compare the student perception of human vs. AI generated slides. We find that coding assistant tools produce slides that were most accurate, complete, and pedagogically sound. Additionally, students rate GenAI slides to be of similar quality as instructor-created slides, and cannot reliably identify which slides are AI-generated. Additionally, we find a negative correlation between a high quality rating and a high ``AI-generated'' rating, suggesting students associate poor quality with the source of the slides being AI. These findings highlight promising opportunities for integrating GenAI into instructional design workflows and call for further research on how educators can best harness such tools responsibly and effectively.
BibTeX
@misc{leinonen2026ai,
  author = {Juho Leinonen and Lisa Zhang and Arto Hellas},
  title = {{AI-Generated Slides: Are They Good? Can Students Tell?}},
  year = {2026},
  eprint = {2605.13532},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2605.13532},
}

Loading related content...

Figure