A new Vector tool screens AI training data for bias

A caregiver writes “noncompliant patient refused medication” in a chart. That one word, noncompliant, will likely follow the patient into every appointment after, colouring how the next clinician reads them before they’ve said anything.

The Vector Institute’s new tool wants to catch these buried assumptions. 

On June 30, the Toronto-based national AI institute released UnBias-Plus. Its researchers explain that the free, open-source AI tool can detect biased language across race, gender, age, and political framing, explain why a phrase was flagged, and rewrite it. 

The platform has two tiers. A browser-based version is available for public use, supporting content of up to 750 words. A developer installer can be integrated into company applications, systems and workflows. 

In the above healthcare example, the tool would suggest “patient declined medication,” favouring factual language over judgment.

Newsroom and HR use cases are easy to picture. Inclusive reporting, job descriptions, promotion cases, and internal communications all require accuracy and the removal of unintentional barriers. For another example, insurance teams can use the tool to review benefits letters, eligibility notices, and coverage communications.

Technology leaders need to look at the data layer. 

UnBias-Plus is built to screen the annotations, prompts, and source text that feed model training and fine-tuning, before bias is learned and reproduced at scale, across everything the model touches.

Bias in a single document affects one reader. Bias in training data gets encoded once and reproduced en masse. 

Vector points to research measuring bias in AI training data at rates from 3.4 to 38.5%, and notes even safety-tuned models still show implicit racial and gender bias.

“The people most harmed by biased language are often the last to know it’s there,” says Shaina Raza, applied machine learning scientist for responsible AI at Vector.

The timing fits where boards are looking. 

Vector positions UnBias-Plus within Canada’s national AI strategy, which names algorithmic bias as a pillar of responsible AI. In 2026, enterprises are being asked to prove they can govern, defend, and measure their AI, with real scrutiny and consequences. Bias in the data a model learned from is one of those proof points.

A tool that audits bias at the data layer puts the question in front of the people accountable for what their models learned.

Final Shots

• Bias in training data is both a governance and ethics question. It gets encoded once and reproduced everywhere the model runs.
• The value for technology leaders sits at the data layer, screening what feeds a model.
• UnBias-Plus handles English text only, and doesn’t verify facts or flag AI-generated content. Vector says audio, video, and misinformation detection are planned.