GPT-4 zero-shot PII redaction prompt and pipeline for educational forum posts¶

Element · All elements
Evidence · 5 claims (3 for, 2 mixed) · 1 study (1 design), q2 · 0 of 1 report an effect size · 5 claims rest on one study

Description¶

A de-identification pipeline in which each forum post is sent individually to the OpenAI gpt-4-0613 model via API with a prompt instructing it to "remov[e] any personally identifiable information (PII)" including names, company names, places of origin, current living locations, addresses, and social media links, replacing removed PII with '[REDACTED]' while keeping the rest of the text unchanged word for word. The code for the GPT-based de-identification process was released in a GitHub repository for replication.

Design Implications¶

Context¶

Requirements¶

• API access to GPT-4 with data sent under OpenAI's privacy policy, which retains data for a maximum of 30 days

Constraints¶

• An early prompt draft explicitly instructing GPT not to redact public figures caused lower overall performance for both precision and recall; without instructions to preserve formatting, GPT corrected grammar and changed words

Target Learners¶

• Educational data mining and learning analytics researchers de-identifying student text data

Target Learning Goals¶

• Automated redaction of PII from student-generated discussion forum posts

Claims¶

• GPT-4 detected 45 PII words that human coders failed to redact across all nine courses [+W]
• GPT-4 achieves high recall (average 0.958) but low precision (average 0.526) when de-identifying MOOC forum posts [+W]
• GPT-4 over-redacts famous names, locations, and mythological creatures that are not PII in educational forum discussions [~W]
• Supervised machine learning de-identification still outperforms GPT-4, though GPT-4 exceeds class-list/regular-expression and transformer approaches in recall [+W]
• GPT-4 de-identification performance varies by course, with lower agreement in courses with longer posts and more qualitative discussion [~W]

Related Elements¶

Examples¶

• Use LLM-based de-identification as an additional verification layer to catch human redaction mistakes

Key Sources¶

• S. Singhal, A. F. Zambrano, M. Pankiewicz, X. Liu, C. Porter, and R. S. Baker. (2024). De-identifying student personally identifying information with gpt-4. Proceedings of the 17th International Conference on Educational Data Mining, pages 559-565, Atlanta, Georgia, USA. https://doi.org/10.5281/zenodo.12729884