Integrating Machine Learning into K–12 Classrooms: Enhancing Critical Thinking Through Interactive AI Tasks

Technical Summary: Integrating Machine Learning into K–12 Classrooms

Problem Statement
While the integration of Artificial Intelligence (AI) and Machine Learning (ML) into education is accelerating, empirical research regarding their impact on Critical Thinking (CT) development in K–12 settings remains limited. Existing literature often treats ML as a tool for predictive analytics (e.g., student profiling) rather than a pedagogical medium for cognitive development. Furthermore, studies frequently conflate Computational Thinking (CT) with Critical Thinking, or focus solely on technical skill acquisition without examining how ML tasks activate specific CT processes such as inference, evaluation, and evidence-based reasoning. There is a notable gap in understanding how middle school students engage with CT indicators during specific phases of the ML workflow (training, testing, and application) and how these engagements vary across different learner profiles.

Methodology
This study employed an explanatory case study design involving 15 voluntary fifth- and sixth-grade students (7 girls, 8 boys) from a public school in Türkiye. Participants, who possessed prior proficiency in block-based programming, engaged in an eight-week intervention consisting of five structured ML activities using the Machine Learning for Kids platform. The activities included three text-based tasks (Smart Classroom, Quiz Bot, Make Me Happy), one image-based task (Car or Cup?), and one number-based task (Pac-Man Learner).

Data collection was multi-modal:

1. Screen Recordings: Captured student interactions, workflow sequences, and verbal reasoning during the ML process.
2. Researcher-Designed Worksheets: Documented student outputs and decision-making processes.
3. Semi-Structured Interviews: Conducted post-activity to probe student reflections, cognitive strategies, and perceived challenges.
4. Clinical Interviews: Utilized specific prompts to assess goal identification, question formulation, and inference drawing.

Data Analysis

• Quantitative: A researcher-developed ML and CT Performance Rubric was used to score student performance across three ML phases (Training, Learning & Testing, Application) and eight CT dimensions (Goal Identification, Question Formulation, Information Selection, Concept Identification, Logical Reasoning, Application-Oriented Judgment, Inference, Conclusion). Scores were categorized into three levels: Insufficient (0), Needs Improvement (1), and Proficient (2).
• Cluster Analysis: K-means clustering (using RStudio) was applied to the rubric scores to identify distinct student performance profiles.
• Qualitative: Directed content analysis was performed on interview transcripts and screen recording data, mapping student responses to specific ML stages and CT components. Inter-rater reliability was established using quadratic-weighted Cohen's κ on 30% of the data.

Key Results
The study found that ML-based instructional interventions can support observable CT performance and related behaviors, particularly in goal-oriented reasoning, application-based evaluation, and inference.

• Performance Profiles: Cluster analysis revealed three distinct groups: low, medium, and high CT engagement.
  • Higher-performing students consistently identified relevant concepts during training and testing phases and engaged in analytic monitoring throughout the task cycle.
  • Lower-scoring students struggled with selecting relevant concepts and completing the full ML cycle, indicating a need for stronger scaffolding during learning phases.
• Phase Sensitivity: The study highlighted that the application-oriented tasks elicited higher-order thinking in authentic contexts more effectively than earlier phases.
• CT Indicators: Students demonstrated the ability to articulate reasoning, evaluate model outputs, and revise approaches, aligning with the APA Delphi consensus components of CT (interpretation, analysis, evaluation, inference, explanation, and self-regulation).

Key Contributions

1. Pedagogical Framework: The study provides evidence that ML activities can serve as a designed practice for making CT visible, moving beyond "content" to a structured pedagogical framework where decision criteria are explained and evidence is weighed.
2. Methodological Innovation: It introduces a multi-method approach combining screen recordings, rubric-based performance assessment, and cluster analysis to map CT dimensions to specific ML phases, addressing the lack of studies that differentiate learner profiles in ML-integrated environments.
3. Curriculum Design: The findings suggest that sequencing ML tasks is critical, with application-oriented tasks potentially driving higher-order thinking. It also identifies the necessity of scaffolding for students who struggle with concept selection and cycle completion.

Significance and Claims
The authors claim that this study bridges the gap between ML education and CT development by demonstrating that middle school students can meaningfully engage in ML tasks that activate higher-order cognitive skills. The paper posits that ML is not merely a technical competency but a powerful framework for promoting CT skills in K–12 contexts. By aligning data-driven feedback loops in inquiry-based learning with CT components, the study justifies design principles such as transparency, criterion-based assessment, and the pedagogical use of error analysis. The authors conclude that observed changes across the five activities highlight the value of structured ML task sequencing for fostering evidence-based reasoning and justification in authentic learning contexts.