Using ChatGPT for Data Science Analyses

1. Problem Statement

The paper addresses the integration of Large Language Models (LLMs), specifically OpenAI's Data Analysis (DA) plugin for ChatGPT, into the Data Science workflow. While LLMs offer the potential to automate statistical tasks and lower the barrier to entry for non-programmers, there is a lack of empirical evidence regarding their reliability, accuracy, and limitations in real-world statistical modeling. The authors aim to assess whether the DA plugin can serve as a viable "co-pilot" for data exploration, visualization, and modeling, or if its tendency for hallucinations, biases, and technical errors renders it unsafe for unsupervised use.

2. Methodology

The authors conducted a series of empirical experiments using the ChatGPT DA plugin as a quantitative co-pilot. The methodology involved:

• Datasets:
  • Laptop Price Dataset (Kaggle): Used for Exploratory Data Analysis (EDA), data loading, preprocessing, and visualization.
  • Duke Forest Dataset (openintro R package): Used for supervised and unsupervised learning tasks (regression and clustering).
• Prompting Strategy: The study utilized a "zero-shot" prompting approach, keeping prompts simple and following the plugin's own suggestions to simulate a novice user's interaction. They also tested the effects of repeated prompting.
• Evaluation Framework:
  • The authors scored the DA's outputs using a star rating system (1 to 5 stars).
  • They critically evaluated the generated Python code, statistical interpretations, visualizations, and model diagnostics.
  • They compared the DA's outputs against standard statistical best practices (e.g., checking for linearity, normality, causality vs. correlation, and appropriate metric selection).

3. Key Contributions

The paper provides a comprehensive "hands-on" review of the DA plugin's capabilities across the Data Science lifecycle:

• Data Loading & Preprocessing: Demonstrated that the plugin handles CSV, XLSX, and TXT files seamlessly but struggles with R/SAS formats (.RDATA, .SAV), requiring manual conversion.
• Visualization Capabilities: Evaluated the plugin's ability to generate histograms, boxplots, mosaic plots, and correlation heatmaps.
• Supervised Learning: Tested the plugin's ability to select, implement, and interpret Linear Regression, Decision Trees, Random Forests, Gradient Boosting, Support Vector Regression (SVR), and Neural Networks.
• Unsupervised Learning: Assessed the implementation of K-Means clustering and the Elbow method for determining cluster counts.
• Critical Gap Analysis: Highlighted specific failures in statistical reasoning, such as misinterpreting correlation coefficients, ignoring model assumptions (e.g., negative price predictions in linear regression), and failing to detect non-linearities in model diagnostics.

4. Results

The experiments yielded mixed results, highlighting a dichotomy between the plugin's interface capabilities and its statistical reasoning:

• Strengths:

  • Workflow Automation: The plugin excels at translating natural language queries into executable Python code, handling data loading, and generating summary statistics.
  • Visualization: It successfully generates standard plots (bar charts, scatter plots) and can iterate on user feedback to refine visualizations.
  • Educational Value: For novices, it provides a reasonable roadmap for analysis steps (e.g., splitting data, selecting features).

• Weaknesses & Errors:

  • Hallucinations in Interpretation: The plugin frequently misinterpreted its own generated data. For example, in the laptop dataset, it incorrectly claimed prices were on a log scale when they were not. In the correlation heatmap, it reported correlation values (e.g., 0.74) that did not match the actual plot values (0.68).
  • Statistical Missteps:
    • Model Selection: It suggested using $R^2$ for non-linear models (like SVR), which is statistically inappropriate and led to negative $R^2$ values that were not properly contextualized.
    • Assumption Violations: In linear regression, it failed to warn about the possibility of negative price predictions and did not initially check for residual normality or homoscedasticity without specific prompting.
    • Technical Limitations: The plugin failed to run Bayesian regression due to computational complexity and could not execute Neural Networks because the TensorFlow library was unavailable in the sandbox environment.
  • Visual Flaws: Generated mosaic plots often suffered from overlapping labels and poor color choices, requiring manual intervention.

5. Significance and Conclusion

The paper concludes that while the ChatGPT DA plugin is a powerful assistive tool (a "co-pilot"), it is not yet a substitute for a professional data analyst.

• Human-in-the-Loop is Essential: The authors emphasize that no AI-powered statistical software should operate without human critique and oversight. The plugin's outputs can be misleading, particularly for novice learners who may lack the domain expertise to detect statistical errors.
• Paradigm Shift: The interface between LLMs and statistical models represents a significant shift in statistical software, potentially democratizing data science but introducing new risks regarding accuracy and reproducibility.
• Future Directions: The authors suggest that future developments should focus on multi-agent frameworks and hybrid deployments (combining LLMs with established AutoML systems) rather than relying on a single LLM for end-to-end analysis. They call for rigorous benchmarking of LLMs against standard statistical test problems to better understand their performance boundaries.

In summary, the paper serves as a cautionary yet optimistic guide, urging practitioners to leverage the efficiency of LLMs for workflow automation while maintaining strict human oversight over statistical validity and interpretation.