Vector Institute launches open-source writing bias detector

Toronto's Vector Institute has launched a new application to detect bias in writing.

UnBias-Plus is a free, open-source AI tool that users can integrate into their organisation's workflows using publicly released code on GitHub or, for more personal uses on the browser-based site, which supports content of up to 750 words.

Shaina Raza, Applied Machine Learning Scientist, Responsible AI at the Vector Institute, was a principal architect on the project. The inspiration for the project: she said that although the Canadian Government has released its AI strategy, trust in these technologies remains low across the country. 

"With Canada's AI strategy, they want people to trust AI, but the problem is that if AI is itself biassed on certain demographics or certain group of people, people will be hesitant to use AI, and we don't want that situation," said Raza. "Once people trust AI, they will use it more. It will definitely also positively impact our economy. There will be more production and more software, so we definitely want to bridge this gap."

Last year, Canada ranked among the lowest of 47 countries surveyed by KPMG and the University of Melbourne in terms of regular use of and trust in AI. Only 34 per cent of Canadian respondents were "willing to trust AI."

May's AI Monitor Report from Ipsos, which assessed global AI sentiment, found that among 32 countries surveyed, Canada consistently scored in the bottom tier for trust. In response to the prompt "I trust artificial intelligence to not discriminate or show bias towards any group of people," 36 per cent of Canadians agreed.

Raza said UnBias-AI's goal is to mitigate bias from both algorithmic and stakeholder-management perspectives: the next generation, who rely heavily on these applications, could benefit from education and training, while industries that use agents and LLMs need to adapt.

While the program flags bias and suggests ways to address it, it does not correct the user's work. Raza said this allows users to reflect on their own writing and mitigate bias in a deliberate way.

She confirmed that the model was trained on extensive English-language media works from 50 to 60 news organisations around the world, across various genres. From how a person has been described to the connotation of a verb used in a headline.

"One example: 'the immigrants flooded the border,' is normally a sentence we see in the news headlines. But compare that to 'the immigrants arrived at the border.' There is a tiny word like flooded, which caused this bias. So we give this kind of suggestions for journalism, for healthcare, and for students who are writing the essays."

The institute is working with media organisations to explore future implementation in their workflows. In addition to journalism, the institute cited use cases in human resources, insurance and health care communications.

While Vector launched the first version of UnBias in 2023, just a few months following the launch of ChatGPT, this week's launch marks a catch-up to the highly generative and agentic nature of AI in 2026.

"The original UnBias was mostly on encoder-only models and one decoder-only model, because the science of LLMs was not very advanced that time, and there was no specific advanced agents. That was small scale, that was built more on rule-based approaches and some generative AI approaches," said Raza. "In UnBias-Plus, we totally approached the generative way, autoregressive approach."

As the latest version continues its pilot, Raza is already thinking ahead to stay ahead of the curve. The next analysis: deepfakes.