Researchers Develop Fully Local AI System for Protecting Student Privacy in Educational Transcripts

A new machine learning approach addresses a critical challenge in educational research: protecting student privacy while preserving the curricular content needed for academic study. Researchers have developed a fully local artificial intelligence cascade that can distinguish between personally identifiable information and legitimate educational terms that happen to share names—such as whether "Riemann" refers to a student or the mathematical concept.

The core problem stems from tension between data governance and accuracy. Commercial large language models can handle semantic ambiguity but require sending sensitive student data to third parties, raising privacy concerns. Conversely, traditional local named entity recognition systems keep data on-device but frequently over-redact legitimate curricular terms. The proposed solution reframes de-identification as a constrained privacy triage task rather than open-ended entity recognition.

The system uses a two-stage cascade architecture. A recall-first union proposer combines two lightweight encoders with deterministic rules to generate candidate spans for potential redaction, intentionally over-generating to ensure no sensitive information is missed. A context-aware reviewer then examines each candidate in the surrounding dialogue and speaker role context, making a binary Redact/Keep decision. Evaluation on mathematics tutoring transcripts from two large educational platforms showed the strongest local configuration achieved a macro F1 score of 0.958, substantially outperforming a same-family LLM-only baseline (0.767) and a commercial API (0.706), while running entirely on a single laptop computer.

The research demonstrates particular strength in handling ambiguous cases. On a targeted test set of curricular-personal name ambiguity, the local system degraded by only 0.03 F1 compared to 0.19 to 0.25 for smaller alternative reviewers. The findings suggest that careful problem formulation and architectural design can outweigh the advantage of using larger models for specialized de-identification tasks in educational contexts.