Vector Institute’s UnBias-Plus is a free tool that automatically strips linguistic and contextual bias from data pipelines and enterprise documents.
On 30 June 2026, Toronto’s independent, not-for-profit Vector Institute launched UnBias-Plus (UnBias+), a free text-processing toolkit designed to automatically detect, explain, and rewrite biased English text across race, gender, age, and political framing.
Unlike simple keyword blockers, the tool unifies four functions into a single pipeline: segment-level classification, character-level span localisation, neutral rewriting, and structured reasoning.
-
Technical Workflows: It serves as a pre-processing filter for AI developers to sanitise unstructured data, prompts, and training sets before models ingest them.
-
Enterprise Operations: It offers localized auditing to help HR screen job descriptions for gendered or racialised language, assist newsrooms in catching editorial bias, and allow healthcare providers to strip stigmatising terminology from clinical notes.
The framework runs on fine-tuned open-weight models from the Qwen3 family (Qwen3-8B and Qwen3.5-4B). It provides structured JSON files tracking character offsets, bias types, severity estimations, and justifications alongside a safe rewritten copy.
UnBias-Plus is available via a web demo (up to 750 words) and a developer installer on PyPI. While publicized as a free tool, the repository’s initial LICENSE.md file restricts internal deployment primarily to academic entities, sponsors, and official Vector partners—meaning enterprise engineering teams should verify compliance terms before establishing production dependencies.
The pipeline strictly targets language bias; it does not check factual accuracy, track misinformation, or detect AI watermarks. Subsequent versions plan to expand into audio and video bias detection.
