Can Students Tell AI-Generated Slides from Human-Made Ones? Their Detection Heuristic Is Backwards

Forum topic · 小凯 · 2026-05-18

Summary

A study by Leinonen, Zhang, and Hellas had five AI tools—NotebookLM, Claude, M365 Copilot, Cursor, and Claude Code—generate lecture slides from instructor-written notes. Education experts evaluated quality, and students in real classrooms rated the slides and attempted to identify which were AI-generated. Coding assistants (Cursor, Claude Code) outperformed general-purpose chatbots on accuracy and completeness, likely because slide generation is a structured-output task suited to coding tools. In the classroom, students rated AI-generated slides no differently from instructor-made ones and could only barely beat random guessing when identifying AI slides. Notably, the study found a negative correlation: students gave higher-quality slides higher odds of being human-made and assumed low-quality slides were AI-generated. This quality-based heuristic fails now that AI slides reach human-level quality. Open questions remain about subject matter, the baseline quality of instructor slides, and evaluation weighting. Source: arXiv:2605.13532.

Every university instructor faces the same reality: there is never enough time to prepare course materials. Leinonen, Zhang, and Hellas ran a study in which five different AI tools—NotebookLM, Claude, M365 Copilot, Cursor, and Claude Code—generated lecture slides from instructor-written course notes. Education experts then assessed the slides' quality, and the materials were deployed in real classrooms where students rated them and guessed which ones were AI-made.

Gap Between AI Tools

Slides generated by general-purpose large language models (Claude, Copilot) were less accurate and complete than those from the two coding assistants (Cursor, Claude Code). The reason: coding assistants are better at handling structured output—slide generation is essentially a structured-document task, and coding assistants produce tightly formatted blocks of code every day. This is a subtle observation: coding ability transferred to the slide-making scenario, while general chatbots underperformed.

Key Findings from the Classroom

Students' ratings of AI-generated slides showed no significant difference from instructor-made slides.
Students could not reliably identify which slides were AI-generated; accuracy was only slightly better than random guessing.

The Most Interesting Result: A Negative Correlation

The more highly a student rated a slide, the more likely they were to believe a human made it. Lower ratings correlated with attributions to AI. Students were not detecting an "AI style"—they were inferring the source from quality: high quality → human-made; poor quality → AI-made. This heuristic breaks down completely once AI slides reach human-level quality.

What Remains Unclear

Which course and subject the slides came from—AI slide quality may vary considerably across disciplines.
The quality baseline of the instructor-made slides—were they ordinary PowerPoint-template decks or carefully designed teaching materials?
How the evaluation criteria (accuracy, completeness, pedagogical soundness) were weighted.

References

1. Leinonen, J., Zhang, L., & Hellas, A. (2026). *AI-Generated Slides: Are They Good? Can Students Tell?* arXiv:2605.13532 [cs.AI]. 2. Mollick, E., & Mollick, L. (2023). *Using AI to Implement Effective Teaching Strategies in Classrooms*. Wharton School. 3. Hellas, A., et al. (2018). *Predicting Academic Performance: A Systematic Literature Review*. ITiCSE.