🤖 AI

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

This study evaluates various generative AI tools for creating educational slides, finding that coding assistants produce the most pedagogically sound content and that students perceive AI-generated slides as comparable in quality to human-created ones, often associating perceived lower quality with the AI source.

Original authors: Juho Leinonen, Lisa Zhang, Arto Hellas

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Juho Leinonen, Lisa Zhang, Arto Hellas

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a teacher who has written a fantastic, thick textbook for your class. Now, you need to turn those pages into colorful slides for your lectures. It's a lot of work, like trying to build a house brick-by-brick when you'd rather just assemble a pre-fabricated kit.

This paper is like a report card on using AI tools to do that heavy lifting. The researchers asked two simple questions:

Are the AI-made slides actually good?
Can the students tell the difference between AI-made slides and human-made ones?

Here is the story of what they found, explained simply.

The "Toolbox" Test

The researchers didn't just pick one AI. They tested five different types of tools, like trying out five different brands of power drills to see which one builds the best shelf:

The "Generalist" Tools: These are like Swiss Army knives (e.g., NotebookLM, M365 Copilot, Claude). They are good at many things but sometimes miss the mark on specific details.
The "Coding Assistants": These are like specialized, high-end power tools designed for programmers (e.g., Cursor, Claude Code). Even though making slides isn't strictly "coding," these tools turned out to be the best at the job.

The Verdict on Quality:
The "Generalist" tools were a bit like a clumsy apprentice. They made slides that looked okay on the surface but were missing key facts, had confusing diagrams, or included irrelevant pictures (like a slide about HTML code featuring a picture of a bird).

The "Coding Assistants," however, were like master craftsmen. They produced slides that were accurate, complete, and followed a logical teaching flow. They didn't just copy-paste; they understood how to structure the information for a classroom.

The "Blind Taste Test"

Once the researchers picked the best AI slides (from the coding assistants), they did something tricky. They took those AI slides, fixed the fonts and colors to match the teacher's usual style, and mixed them into the real lectures alongside slides the teacher made by hand.

The students were told, "Hey, some of these slides might be made by AI," but they weren't told which ones. After every lecture segment, the students had to:

Rate how good the slides were.
Guess: "Was this made by a human or a robot?"

The Results:

The Taste Test: The students rated the AI slides and the human slides as equally good. They couldn't tell the difference. The average score for both was a solid "5 out of 7."
The Guessing Game: The students were terrible at guessing. They guessed "AI" about as often as they guessed "Human," which is basically just random guessing.
The "Bias" Twist: Here is the interesting part. The researchers found a funny pattern: When students thought a slide was made by AI, they tended to rate it lower. Even if the slide was actually great, if they suspected a robot made it, they gave it a lower score. It's like a food critic saying, "This soup tastes great, but since I think a machine cooked it, I'll give it a lower rating just in case."

The Big Takeaways

Coding Tools Win: If you want AI to make educational materials, tools designed for programmers (coding assistants) work better than general chatbots, even for non-coding tasks like making slides.
Humans are Still Needed: The AI isn't perfect yet. It needs a human to check the work, fix the formatting, and make sure the style matches. Think of the AI as a very fast, very knowledgeable intern who needs a manager to review their work before it goes to the client.
Students Don't Care (or Can't Tell): As long as the human teacher fixes the little errors and makes it look professional, students can't tell the difference and rate the quality the same as if the teacher made it themselves.

In a nutshell: AI can be a powerful helper for teachers to save time, but it works best when the teacher acts as the "editor-in-chief" rather than just hitting a "generate" button and walking away. And don't worry, the students won't know the difference if you do it right.

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

Problem Statement

As Generative AI (GenAI) and Large Language Models (LLMs) become increasingly accessible, computing instructors face both opportunities and uncertainties regarding their integration into pedagogical workflows. While GenAI tools show promise in increasing productivity for programming and writing tasks, they frequently generate incorrect or hallucinated outputs. A specific gap exists in evaluating GenAI's ability to transform instructor-authored course notes (specifically online textbook chapters) into high-quality lecture slides. This transformation is challenging because it requires coordinating text, code, and diagrams while maintaining pedagogical soundness. Furthermore, there is limited understanding of how students perceive AI-generated educational materials compared to human-created ones, and whether students can reliably distinguish between the two sources.

Methodology

The study employed a mixed-methods case study approach within an upper-year undergraduate Web Software Development course at Aalto University. The research was divided into two primary phases:

Phase 1: Instructor Assessment of Tool Quality

Researchers evaluated five distinct GenAI tools across a spectrum of accessibility and workflow integration:

End-to-end education tools: NotebookLM.
General-purpose LLMs: M365 Copilot and Claude (Sonnet 4.5).
Coding assistants: Cursor and Claude Code.

Process:

Input: Four textbook chapters covering basic HTML/Svelte, state sharing, CRUD operations, and the Fetch API were used as source material.
Prompting: Researchers used iterative, straightforward prompts without advanced prompt engineering, asking the tools to generate slides for 10–15 minute lecture segments including peer-instruction exercises.
Evaluation Criteria: Two researchers (both computing educators) and the course instructor conducted a qualitative narrative assessment based on three factors:
1. Factual Accuracy: Verification against source materials and identification of hallucinations.
2. Completeness: Assessment of whether key concepts and examples were omitted.
3. Pedagogical Soundness: Evaluation of cognitive load management, dual coding (visual/text balance), and scaffolding of concepts.

Phase 2: Student Perception Study

Based on Phase 1, the researchers selected the highest-performing tools (Cursor and Claude Code) for a classroom experiment.

Experimental Design: Over five weeks, the course utilized a mix of human-generated slides and AI-generated slides (modified to match the instructor's visual style and formatting).
Blinding: Students were informed that lectures might contain GenAI content but were not told which specific segments were AI-generated.
Data Collection: After each lecture segment, students completed a survey rating the slide quality on a 7-point Likert scale (1 = Very bad, 7 = Excellent) and guessing the source (1 = Definitely Human, 7 = Definitely AI).
Analysis: The study analyzed mean ratings, medians, and correlations between quality ratings and source guesses using Mann-Whitney U tests and Spearman correlation coefficients.

Key Contributions

Comparative Tool Evaluation: The paper provides a practical, narrative assessment of five different GenAI tool categories for educational slide generation, moving beyond theoretical system development to real-world instructor experience.
Identification of Optimal Tool Types: The study identifies that coding assistants (specifically Cursor and Claude Code) outperform general-purpose LLMs and education-focused tools for this specific task, producing slides that are more accurate, complete, and pedagogically sound.
Student Perception Data: The study offers empirical data on student perceptions, demonstrating that students cannot reliably distinguish between high-quality AI-generated slides and human-created slides.
Workflow Insights: The paper documents the "instructor-in-the-loop" workflow, highlighting that while AI drafts require iteration (particularly for styling and formatting), they can be effectively integrated into existing teaching practices.

Results

Instructor Assessment (Phase 1)

Coding Assistants (Cursor, Claude Code): Produced the most compelling results. They were rated as Accurate, Complete, and Strong in pedagogical soundness. They successfully generated code examples, maintained logical progression, and leveraged parallelism in content presentation. Minor issues included occasional rendering errors and content that was sometimes "too complete," requiring instructor curation.
General-Purpose LLMs (Claude, M365 Copilot): Produced slides that were generally accurate but often Incomplete and Problematic pedagogically. Issues included missing code examples, superficial content, irrelevant large images, and poor text-image balance.
Education Tools (NotebookLM): While visually appealing, the output was deemed Very Incomplete and Very Problematic pedagogically. The content was often superficial, lacked necessary code examples, and the "Video Overview" format was too long-winded for lecture segments.

Student Perception (Phase 2)

Quality Ratings: There was no statistically significant difference in the quality ratings between AI-generated slides (Median = 5.0) and human-generated slides (Median = 5.0). The Mann-Whitney U test confirmed that students perceived the quality of both sources as equivalent ( $p = 0.84$ ).
Source Identification: Students could not reliably identify the source of the slides. The median guess for both AI and human slides was 4.0 (neutral), and the Mann-Whitney U test showed no significant difference in guessing accuracy ( $p = 0.1$ ).
Correlation: A significant negative correlation ( $\rho = -0.28, p < 0.01$ ) was found between quality ratings and the guess that slides were AI-generated. This suggests that when students perceived slides as lower quality, they were more likely to attribute them to AI, indicating a bias associating poor quality with AI generation.

Significance and Claims

The paper claims that integrating GenAI into instructional design workflows is promising, particularly when using coding assistants rather than general-purpose chatbots or dedicated education tools. The authors emphasize that:

Time Savings: Using these tools can save significant instructor time, provided the instructor remains in the loop to iterate on style and verify content.
Student Experience: When AI-generated materials are of high quality and stylistically consistent with human materials, they do not negatively impact student perception or experience.
Pedagogical Approach: Instructors should approach GenAI usage similarly to how programmers use coding assistants: through iterative prompting, refinement, and leveraging existing documentation, rather than expecting a perfect first draft.

The authors modestly frame this work as a "snapshot" of late 2025 capabilities and a case study rather than a definitive, generalizable research conclusion. They acknowledge limitations, including the single-institution context, the specific nature of the course (web programming), and the lack of measurement regarding actual learning outcomes. They call for further research on how educators can responsibly and effectively harness these evolving tools.