Impact of Generative AI on Students' Learning Outcomes: An In-depth Analysis Based on 50,000 Data Entries

This study conducts data analysis on 50,000 student records to explore the relationship between generative AI usage patterns and academic performance, skill retention rate, AI dependency, test anxiety, and burnout risk. Key findings include: moderate AI usage (5-15 hours per week) is associated with better academic outcomes, but excessive dependency (over 15 hours) may lead to decreased skill retention and increased burnout risk; usage scenarios (e.g., debugging vs. directly obtaining answers) and prompt engineering skills significantly affect learning outcomes; the study provides empirical evidence for educational institutions to formulate AI usage policies.

With the rapid popularization of generative AI tools like ChatGPT and Claude in education, the question of "whether AI helps or weakens students' independent thinking" has become a key issue. By analyzing data from 50,000 students, this study aims to reveal the real relationship between AI usage patterns and academic outcomes, providing data support for educational institutions to find a rational middle ground between "complete ban" and "unrestricted use".

Dataset: Based on 50,000 student records, covering 16 core variables (academic indicators: GPA before and after the semester, skill retention score; AI usage behavior: weekly duration, traditional learning duration, scenario classification, prompt engineering skill rating, etc.; psychological state: perceived AI dependency, test anxiety level, etc.; background information: major category, grade, school AI policy, etc.).

Analysis Methods: Data quality inspection and cleaning, feature engineering, exploratory data analysis, user segmentation analysis, correlation analysis, and visual presentation.

1. Non-linear Relationship: The group using AI for 5-15 hours per week shows the strongest GPA improvement and skill retention rate; using AI for more than 15 hours leads to decreased skill retention and increased burnout risk (inverted U-shape);
2. Usage Scenario Differences: High-efficiency scenarios (debugging, feedback acquisition, concept explanation, etc.) are associated with strong academic outcomes, while low-efficiency scenarios (directly obtaining answers) show the opposite;
3. Prompt Engineering Skills: Individuals with advanced skills have higher skill retention rates and healthier dependency levels;
4. Learning Balance: Learning patterns with an excessively high proportion of AI have the risk of cognitive degradation;
5. Contextual Factors: Variables like school policies and majors affect AI usage outcomes, requiring differentiated support.

The study visually presents key findings through charts:

• GPA change and AI usage duration show an inverted U-shape (the 5-15 hours group is optimal);
• Skill retention rate is best in the moderate usage group;
• High burnout risk rises sharply with usage duration;
• The "debugging/troubleshooting" scenario is associated with the highest GPA improvement;
• Prompt skills have a strong correlation with skill retention rate.

Seven recommendations for educational institutions:

1. Teach responsible AI usage (instead of banning or unrestricted use);
2. Encourage active learning usage scenarios;
3. Be alert to the "directly obtaining answers" model;
4. Cultivate prompt engineering as an academic skill;
5. Pay attention to students who use AI for more than 15 hours per week;
6. Promote balance between AI-assisted and traditional learning;
7. Provide group-based differentiated support (e.g., STEM majors, graduate students, etc.).

Core conclusions: Moderate and purposeful AI usage is associated with stronger academic outcomes, while excessive usage leads to issues like dependency, anxiety, and burnout; those using active problem-solving scenarios and advanced prompt skills perform best.

Implications: Education should cultivate students to be wise users of AI. Core strategies include guidance rather than prohibition, maintaining balance, cultivating prompt literacy, targeted support for high-risk groups, and providing an empirical basis for AI education policies.

Limitations: Observational data cannot prove causal relationships; self-reported variables have subjective biases; lack of longitudinal behavior tracking; presence of potential confounding factors (course difficulty, teacher policies, etc.).

Future Directions: Deepen causal understanding through randomized controlled trials, long-term longitudinal tracking, and fine-grained behavior data collection.