broodinggoat/qwen-privacy-codegen-pilot — explained in plain English
Analysis updated 2026-05-18
Compare how different privacy prompt styles affect AI-generated code.
Study whether W3C privacy principles translate into concrete implementation choices.
Review example generated FastAPI apps handling sensitive health data under different prompts.
Use the documented evaluation layers as a template for testing other privacy prompts.
| broodinggoat/qwen-privacy-codegen-pilot | 0xustaz/streamgate | a-bissell/unleash-lite | |
|---|---|---|---|
| Stars | 1 | 1 | 1 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | hard | hard |
| Complexity | 3/5 | 4/5 | 4/5 |
| Audience | researcher | developer | researcher |
Figures from each repo's GitHub metadata at analysis time.
Requires access to run Qwen for code generation across the four documented prompt conditions.
qwen-privacy-codegen-pilot is a small research project that tests whether telling an AI code generation model, Qwen, to follow privacy principles actually changes the code it writes. The model was asked to build the same FastAPI application, one that handles sensitive family health data, four separate times, each time with a different level of privacy instruction in the prompt. The four conditions ranged from giving the model no privacy guidance at all, to a short general reminder about privacy and security, to guidance based on official W3C privacy principles, to that same W3C guidance combined with concrete operational details like specific field rules, API expectations, examples, and required tests. The project then compared the generated applications to see how privacy-related choices, such as how consent is checked, what data gets logged, and what gets shared externally, differed between the four versions. The early result reported is that the most detailed condition, combining W3C principles with concrete operational guidance, produced the strongest privacy-related implementation. That version included more privacy tests, clearer rules about which data fields could be stored or shared, and more complete handling of user consent. The author is careful to note this is an interpretation of a small pilot, not a proven conclusion, and that the sample size is limited. The repository is explicit about what it is not: it is not a claim that Qwen or any of the generated apps are actually safe to use with real privacy data, not a general coding ability benchmark, and not an official W3C compliance check. It is meant purely as an early pilot exploring whether more specific privacy instructions lead to more privacy-aware code, with further testing and independent verification planned.
A research pilot testing whether giving an AI code generator detailed privacy instructions produces more privacy-aware application code.
Mainly Python. The stack also includes Python, FastAPI, Qwen.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly researcher.
This repo across BitVibe Labs
double-check against the repo, no cap.