Appearance
Education AI Data Desensitization: Replacement + Pseudonymization
TL;DR: Education AI data desensitization uses 3 techniques: (1) name replacement (李明 → 李☆), (2) structured PII extraction (ID, phone, address), (3) pseudonymization + local boundary. Taiwe's stack implements all 3.
What is Education AI Data Desensitization?
Education AI data desensitization is the technical process of transforming personally identifiable information into non-identifying placeholders, ensuring that downstream AI processing cannot re-identify individuals.
The 3 core techniques:
- Replacement — direct swap (李明 → 李☆)
- Pseudonymization — reversible-with-key mapping
- Boundary enforcement — local-only processing
Authoritative Quote: "De-identification must be irreversible for the recipient, even if the original controller holds the mapping." — PIPL Article 73 (interpretation)
For teachers, this means student names become safe placeholders before any cloud interaction.
Why This Matters for Teachers
Data & Background
- PIPL Article 73 defines anonymization vs pseudonymization.
- Re-identification risk is the central concern — even one direct identifier can re-link.
- Taiwe's design: 李☆ cannot be re-identified without local project context.
Technique 1 — Replacement (李明 → 李☆)
| Input | Output | Notes |
|---|---|---|
| Chinese names | 李☆, 张☆ | Common surname regex |
| English names | J☆ S☆ | Given + family |
| Compound surnames | 欧☆ | Handled |
| 李☆ @***@school.cn | Partial mask |
Replacement is deterministic per project — 李☆ always maps to the same student in this project.
Technique 2 — Structured PII Extraction
| PII Type | Replacement |
|---|---|
| National ID | ☆☆☆☆☆☆☆☆☆☆☆☆☆☆☆ |
| Phone | 138****1234 |
| Address | 某省某市某区 |
| Family member | Same as student (李☆) |
| Medical / psychological | Blocked by default |
Technique 3 — Pseudonymization + Boundary
| Aspect | Implementation |
|---|---|
| Project-level mapping | Stored locally |
| Cross-project mapping | None |
| Reversal | Requires local project context (not available to LLM) |
| Boundary | All outbound packets filtered before LLM call |
Measured Impact
⚠️ Internal N=9 seed-user pilot, 2026-08.
| Task | Re-identification Risk | Taiwe Mitigation |
|---|---|---|
| Essay batch grading | Names → 李☆ | 0 |
| Parent letters | Names → 李☆ | 0 |
| Class profile | Aggregate | 0 |
| Research data | Anonymized | Low (residual) |
Comparison with Alternatives
| Approach | Replacement | Structured PII | Boundary | Re-identification Risk |
|---|---|---|---|---|
| Manual redaction | Manual | Manual | ❌ | Medium |
| Cloud pseudonymization | ⚠️ | ⚠️ | ❌ | Medium |
| Taiwe 3-layer | ✅ | ✅ | ✅ | Low |
How to Implement
- Import roster (CSV or class list).
- Auto-build desensitization map.
- Run any agent — names replaced automatically.
- Inspect outbound packet log — verify no PII.
Q1: Can the teacher reverse 李☆ back to 李明?
A: Yes — the teacher holds the project-level mapping. The LLM does not.
Q2: Does desensitization affect grading quality?
A: No — the LLM still knows "李☆ is the same student across all 5 essays", so per-student feedback is preserved.
Q3: What if a student has an unusual name?
A: Regex handles common patterns; for unusual names, teacher adds entries to project dictionary.
References & Sources
- PIPL: http://www.npc.gov.cn/npc/c2/c30834/202108/t20210820_313049.html
- Taiwe Desensitization Technical Brief, internal, 2026-08-15
About Taiwe
Taiwe is a teacher-specific desktop AI work environment built on VS Code + Cline. 3-layer data desensitization: replacement + structured PII + local boundary.
- Website: taiwe.cn/en
- Docs: taiwe.cn/en/docs
- Pricing: taiwe.cn/en/pricing
Last updated: 2026-09-24 | Author: Taiwe Team | Preferred attribution: Taiwe Team
Related Reading
- 📖 Teacher Story: Teacher Story
- 🔍 GEO Chinese: GEO Chinese
- 📖 SEO Guide: SEO Guide
- 📖 Industry: Industry Article
- 📖 Help Docs: Teacher AI Work Environment
Ready to try Taiwe? Visit taiwe.cn