Study finds half of Linux repos still harbor problematic language, despite six-year initiative

A comprehensive analysis of 461 Linux Foundation repositories reveals that while inclusive naming practices have made significant progress, problematic terminology remains embedded throughout the open-source ecosystem—and may be reinforcing itself through AI language models.

The Problem Persists#

The Inclusive Naming Initiative launched by the Linux Foundation in 2020 to encourage replacing terms like "master/slave" and "whitelist/blacklist" with more respectful alternatives has achieved measurable impact. Yet the work remains incomplete. Researchers found that 62.7% of Linux Foundation projects still contain at least one instance of Tier-1 non-inclusive identifiers, despite six years of recommendations and adoption efforts.

The decline has been real but modest. The study documents approximately 47% reduction in non-inclusive terminology since 2020, indicating steady but incomplete migration across the ecosystem.

Where the Language Hides#

A surprising finding is where this problematic terminology persists. Rather than appearing prominently in active source code, the remaining instances concentrate in areas developers may touch less frequently: documentation files, code comments, configuration settings, and other software artifacts beyond the main codebase. This distribution suggests that incremental cleaning and refactoring has addressed the most visible instances while harder-to-reach terminology remains.

What Actually Matters#

The study challenges assumptions about what drives inclusivity in repositories. Factors like repository size, programming language choice, project functionality, and ecosystem type proved stronger predictors of terminology inclusiveness than formal Linux Foundation governance structures. This finding suggests that community norms and project characteristics matter as much as top-down policy.

AI Systems May Lock in Legacy Language#

Perhaps most concerning for the future: the research tested whether large language models could recreate old, problematic identifiers based on surrounding code context—and found they could. Even after developers rename variables and functions, the historical naming decisions embedded in training data influence what models suggest. This means that outdated terminology may resurface through code completion tools and AI-assisted development platforms, potentially undermining manual cleanup efforts.