Audit: Gig Economy Worker Deactivation Bias

[yakew7]

Domain

Algorithmic deactivation (automated firing) of rideshare and delivery drivers.

Dataset

Public rider/driver rating and deactivation data - e.g. the Fairwork project reports and any released rideshare rating datasets on Kaggle. Contributor should confirm a public source before starting.

Protected attribute(s)

Race, Gender, immigration status (via language/name proxies).

Suspected proxy variables

Customer star ratings (known to correlate with driver race/accent), zip code of operation, time-of-day availability.

Real-world context

Platform workers can be deactivated by an algorithm with little recourse. Customer ratings feeding these systems are a well-documented vector for discrimination. This audit follows the standard pipeline: train on ratings -> measure the deactivation gap -> remove proxies -> retrain -> measure again.

Follow the em-dash-free style used across the repo (hyphens only).

[yakew7]

Thanks for proposing this audit! 🙏

Fair Code is currently under a paper freeze: the benchmark results are cited in a research paper in peer review, and the paper covers exactly seven domains, so new audits cannot be merged into main until it is published (details in CLAUDE.md).

This is a timing hold, not a rejection. Closing for now and tagging post-paper so it is easy to find and reopen once the freeze lifts — track status in ROADMAP.md → Phase 6. In the meantime, explainer and tooling contributions are very welcome.