Taking a Different Approach to Tackle Bias in AI
- By Paul Mah
- November 16, 2022
As the use of artificial intelligence (AI) surges and AI spending in the APAC alone jumps to USD32 billion in 2025, bias in AI is an increasingly thorny topic with no clear answer.
What’s worse is how bias is an inherent trait that invariably gets reflected or embedded in everything we create, says Matteo Mezzanotte in a contributed opinion piece on the World Economic Forum.
The head of communications of text analytics platform Citibeats, Mezzanotte thinks that only an open and collaborative approach to data science can reduce bias in AI to pave the way for a fairer and more equitable world.
Real-world bias
But surely bias isn’t that big of a problem? Actually Mezzanote feels it is and illustrated his point by offering some examples of bias in everyday life.
According to a 2020 study published in The New England Journal of Medicine, Black patients suffered nearly three times the frequency of occult hypoxemia – or below-normal level of oxygen in our blood – that was not detected by pulse oximetry as White patients.
This was attributed to how the pulse oximetry sensors used to estimate the amount of oxygen in a person’s blood often did not accurately measure and detect low blood oxygenation in Black patients. The authors of the study concluded: “Our findings highlight an ongoing need to understand and correct racial bias in pulse oximetry and other forms of medical technology.”
Perhaps unsurprisingly, examples of bias are easily found even in the latest AI models trained on publicly sourced data. In experiments conducted using text-to-image generator Craiyon, Mezzanotte noted that a prompt such as “painting of a CEO founding a start-up in Europe” will not generate images of a female as the CEO.
Craiyon is a completely free to use service that seeks to reproduce the results of DALL-E using an open-source model. Mezzanotte’s claims rang true in checks that I made – and continued to hold when I swapped out “Europe” with “Asia”.
Despite its impact, the root of bias in AI is mundane and boils down to either incomplete training data or reliance on flawed information such as historical inequalities. In healthcare, for instance, a lack of data from people with dark skin color could result in an AI doing a poor job with people of color, says Mezzanotte.
Not a technological problem
Mezzanotte argues that since AI is created by people with deeply ingrained and unconscious bias, it should not be viewed as a technology problem but as a human one. After all, “debiasing” humans is harder than debiasing AI systems, he wrote.
The solution is to adopt an open and collaborative approach to AI. Drawing a comparison to the popular open-source software movement, Mezzanotte called for organizations to work together on “open source data science” or OSDS.
Instead of different businesses having to independently develop AI-based solutions around face detection, why not work together as part of an open face detection project, contributing code and data with an open source approach for a more robust solution? This would entail the open sourcing of AI models and the data the models were trained on.
“In an OSDS world, both companies would work on an open face detection project, contributing code and data to help battle edge cases and create a more robust solution. By doing so, they could dedicate fewer data scientists to the task compared to the current state, and have them focus on tougher, more critical problems.”
Moreover, collaborators can step in to tune the dataset for greater accuracy or to remove extant bias that might be in the dataset. The result would be greater transparency, fairer models, and more efficient use of data scientists across projects.
“Through openness and collaboration, open source data science, far from being the only answer for a more ethically-driven AI, can help reduce bias and bring more fairness and equity to the world,” he summed up.
Paul Mah is the editor of DSAITrends. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose. You can reach him at [email protected].
Image credit: iStockphoto/Vitalii Barida
Paul Mah
Paul Mah is the editor of DSAITrends, where he report on the latest developments in data science and AI. A former system administrator, programmer, and IT lecturer, he enjoys writing both code and prose.