Fostering AI Literacy through Opening the Black Box of Foundational Text Generation Models

Daria Smyslova, Franziska Bickel, Qiankun Li, Shiyan Jiang, Rebecca Ellis, Jie ChaoProceedings 2025 · DOI

TL;DR

This pilot study explored using simple Markov Chain models to teach ninth-graders about foundational text generation, aiming to boost AI literacy. While students showed improved confidence and understanding of basic concepts, they still struggled to connect this knowledge to advanced models like ChatGPT, highlighting a crucial gap in current AI education.

What Problem Does This Paper Address?

As AI becomes increasingly integrated into our lives, fostering AI literacy is paramount. However, many students find it challenging to grasp how AI systems, particularly complex ones like large language models (LLMs), actually work. This paper addresses the gap in understanding how to effectively teach AI concepts to K-12 students, specifically focusing on text generation. The authors propose that by demystifying foundational models, like Markov Chains, students can build a more concrete understanding of AI principles, which can then serve as a bridge to comprehending more advanced, "black box" systems.

Methods

The study employed a design-based research approach, focusing on a pilot phase of a larger project. The core intervention involved using a Markov Chain text generation demo in English Language Arts classrooms. Markov Chains were chosen for their foundational role in probabilistic modeling and their relative simplicity, allowing students to visualize concepts like state-based predictions and transition probabilities.

The intervention consisted of five educational activities:

  1. AI is Everywhere: Exploring societal impacts of language-based AI.
  2. Who Wrote It: Comparing student and AI-generated writing.
  3. Writing by Prediction: Understanding language prediction mechanisms.
  4. Markov Chains: Hands-on exploration of text generation using the demo.
  5. What is Art: Examining AI art's implications for creativity.

Students interacted with an online demo that visualized Markov Chain models, allowing them to manually or automatically generate models, adjust parameters, and observe the step-by-step generation process. Data was collected through pre- and post-intervention surveys assessing self-reported competence and knowledge of Natural Language Generation (NLG). Qualitative analysis of open-ended responses provided deeper insights into students' understanding.

Datasets

Not discussed in the paper. The study used a Markov Chain text generation demo as the primary tool, and student-generated text for comparison in one activity.

Key Insights and Ideas

  • Demystifying the "Black Box": Using simpler models like Markov Chains can be an effective gateway to understanding more complex AI, making abstract concepts more tangible.
  • Bridging the Gap is Crucial: While foundational knowledge is important, there's a significant challenge in helping students connect these basic principles to sophisticated, real-world AI systems like ChatGPT.
  • Self-Assessment Matters: Students' confidence and ability to articulate concepts improved significantly after the intervention, suggesting that even basic AI literacy interventions can have a positive impact on self-perceived understanding.
  • Qualitative Insights Reveal Nuance: Beyond quantitative gains, qualitative data showed students developing a more sophisticated vocabulary to describe AI limitations, moving from "random" to discussing "memory constraints" and "lack of creativity."

Key Results

  • Self-Assessment: Students showed significant improvements in their ability to describe AI text generation limitations (moderate effect size, d=0.61) and explain how AI text generation works (small-to-moderate effect size, d=0.46).
  • Knowledge Assessment: Scores on a Natural Language Generation knowledge assessment increased significantly from pre-test (M=2.14) to post-test (M=2.88), with a moderate effect size (d=0.61).
  • Qualitative Findings: Post-intervention, students moved beyond simplistic explanations of AI text generation (e.g., "random," "chance") to discussing concepts like Markov chains, predictive mechanisms, and specific limitations such as memory constraints, lack of creativity, and inability to handle personal topics or emotions.

Limitations and Future Work

Limitations: * This was a pilot study, and the findings are exploratory, requiring further investigation. * The study focused on ninth-graders, and the effectiveness of this approach may vary across different age groups.

Future Work: * The authors emphasize the need to develop strategies that explicitly link foundational models (like Markov Chains) to advanced AI systems (like ChatGPT). * Future iterations will explore how to teach students about enhancements in advanced systems, such as incorporating longer memory, leveraging large datasets, and managing complex contexts. * The goal is to build comprehensive AI literacy in K-12 education by creating scalable and practical approaches that bridge theoretical understanding with practical application.

Summary updated

Abstract

This pilot study investigates the impact of an intervention aimed at demystifying foundational text generation models to enhance AI literacy among ninth-grade English Language Arts students.Pre-and post-surveys revealed significant gains in self-assessed performance, while qualitative analysis showed improved understanding of concepts like predictive modeling.However, students struggled to connect foundational knowledge to advanced AI systems like ChatGPT.The findings highlight the importance of teaching foundational models and developing strategies to bridge this gap, supporting comprehensive AI literacy in K-12 education.
BibTeX
@inproceedings{smyslova2025fostering,
  author = {Daria Smyslova and Franziska Bickel and Qiankun Li and Shiyan Jiang and Rebecca Ellis and Jie Chao},
  title = {{Fostering AI Literacy through Opening the Black Box of Foundational Text Generation Models}},
  booktitle = {Proceedings},
  year = {2025},
  doi = {10.22318/icls2025.326214},
}
Figure