Leveraging AI as a Data Scientist

1. Introduction

AI can accelerate routine coding and documentation, but data scientists must retain control over methods and interpretation. Generated code may contain wrong assumptions, unsafe dependencies, or fabricated results.

2. Core concepts

Provide a minimal schema and a precise task; do not provide private data unless the tool is approved. Ask for explanations and tests, then inspect the code and run it on a controlled sample. A model can propose hypotheses but cannot establish that a claim is supported by evidence.

3. Topics covered in this module

• AI support for exploration, documentation, and debugging
• Cursor-like coding assistants and agentic terminal tools
• Context minimization, permissions, review, and reproducible checks
• Human accountability, source verification, and evaluation

4. Syntax and how to read it

import pandas as pd

frame = pd.DataFrame({'id': [1, 2], 'amount': [10, 20]})
assert frame['id'].is_unique
assert frame['amount'].notna().all()

These checks make assumptions visible. Replace them with domain-appropriate rules and report violations rather than silently hiding them.

5. Practical workflow

Ask for one bounded change, review its diff and dependencies, run relevant tests, and compare outputs against a known case. Keep private records and credentials out of prompts and logs.

6. Common mistakes

• Running an agent with broad filesystem or network permissions it does not need.
• Treating a generated interpretation as a finding without checking source data and methodology.

7. Best practices

• Use least privilege and review every proposed file change.
• Keep an audit trail of material transformations and human decisions.

8. Practice task

A coding agent asks to install a new package and read your home directory. What should you do?

9. One possible solution

Pause and review why each permission/dependency is needed, narrow access to the project and approved package sources, and do not grant broad access by default.

10. Summary

Use coding assistants for bounded tasks and verify generated code, analysis, and claims. Use least privilege and review every proposed file change. Keep an audit trail of material transformations and human decisions.

11. What comes next

The next module, Git for Data Scientists, builds on this foundation and applies it to a new part of the analytics workflow.