MEDIA RELEASE — November 19, 2025 — R0136
“I wan chop. Come chop.”
In Pidgin English, spoken by 140 million people in West Africa, these sentences express a friendly invitation to have a meal. But as Brock University Assistant Professor of Computer Science Blessing Ogbuokiri explains, artificial intelligence (AI) is likely to put a negative spin on the words because of inherent biases in its models against African English dialects from Nigeria and around the world.
“When AI misinterprets the real meaning of Pidgin English and other underrepresented languages, it puts a toxic meaning or interpretation on the language, resulting in what we call amplification of bias,” he says.
This negative bias can pose risks to Pidgin English speakers’ health and well-being, Ogbuokiri says. Individuals may be locked out of medical chatbots or other vital online services, have accounts censored or shut down, and encounter other forms of discrimination in service-delivery systems.
Ogbuokiri and York University Assistant Professor Laleh Seyyed-Kalantari are co-directing a new project that seeks to remove biases against Pidgin English by reforming the dialectic preferences of large language models (LLM).
They will conduct this work through the new Mitigating Dialect Bias Solution Network, created by the Canadian AI Safety Institute Research Program at the Canadian Institute for Advanced Research, which is funded by the Government of Canada. Ogbuokiri and Seyyed-Kalantari’s project is also receiving additional funding from the International Development Research Centre.
“This highly innovative project will boost AI accuracy not only for those living in West Africa but also among immigrants and Indigenous communities in Canada who use non-standard English varieties in their communications,” says St. Catharines Member of Parliament Chris Bittle. “It’s exciting to know that Dr. Ogbuokiri and his lab at Brock have such specialized knowledge.”
In his work, Ogbuokiri describes Pidgin English as a type of dialect that uses abridged versions of words, different ways of pronouncing words and a few local terms.
Because LLMs are trained on standard English, they don’t recognize Pidgin English and tend to interpret some words incorrectly, he says.
Ogbuokiri explains that a sentence written in African American English dialect may read as “these ppl irking my nerves,” for example, compared to “these people are getting on my nerves” in Standard English dialect.
LLMs would assign a negative sentiment to the African American sentence and a positive sentiment to the Standard English sentence, he says, even though the content is the same.
Ogbuokiri and Seyyed-Kalantari’s team aims to create benchmarks for Pidgin English, audit LLMs to identify biases and develop a tool that will reduce these biases in these models.
A benchmark is a standardized dataset or evaluation framework that enables AI models to be trained and tested, much like how humans learn and adapt when they become familiar with a language, says Ogbuokiri.
“If someone is chatting with you using that kind of language, you might not understand what the person is saying at first,” he says. “But if you stayed with this person for some time, you get to know the person’s dialect and you’ll no longer struggle to understand the language.”
The research team of Canadian and African academics plans to work with industry officials, linguists, policy-makers, end users and other community members in Nigeria to create and use the new training model.
“We believe this project will ensure that speakers of underrepresented dialects have a true voice in the future of AI, promoting safer, more accountable and more equitable technologies for all, including Canadians who rely on AI systems that understand and respect their diverse linguistic identities,” Ogbuokiri says.
For more in formation or for assistance arranging interviews:
*Sarah Ackles, Communications Specialist, Brock University [email protected] or 289-241-5483
– 30 –
Recent Comments