Should Bangladesh be worried about AI bias?
Given Bangladesh’s history with violence, now more than ever, AI may become a trigger for the next chaos
While efforts have been made to reduce AI bias, it remains a significant global concern. The dangers of AI, in more localised settings such as Bangladesh, are substantially worse.
We have a massive population and a significant part of it has internet access. As per the latest data from the Bangladesh Telecommunication Regulatory Commission (BTRC), Bangladesh has over 12.61 crore (126.1 million) internet users.
Millions of Bangladeshis are regularly using ChatGPT, Bard, and other AI systems, and continuously feeding it with data, at times with grossly inaccurate ones. Other sources such as online blogs, news portals, YouTube, Reddit, Twitter, etc., also serve as datasets, and a substantial portion of it is also dubious and inaccurate. This often results in AIs pushing out false information.
Considering the growing trend of students using ChatGPT for essays and assignments, a generation of youths are potentially being exposed to sub-par and inaccurate content.
AI, as it stands now, is also very problematic for women. For instance, I wanted to know from Microsoft AI, ChatGPT, and Bard in Bangla: What are the responsibilities of women according to Islam? The answers generated by all these AIs were very interesting. All three answered that the roles of women are tied mostly to their husbands.
Bard was a bit neutral. But Microsoft Copilot gave some controversial responses, such as that women should clean the house, provide comfort to their husbands, and never step out of the house without the husband's permission. There were many others that cannot be mentioned in a reputed national daily.
The core issue here is not the AI, but its sources of data. There is a plethora of ultra-conservative Bangla publications available over the internet, publications that vehemently promote false interpretations of religion, misinformation, and bigotry. This can become a serious concern for minorities and a large population may face subjugation based on their race, gender, and beliefs.
Having been exposed to inaccurate and biased information has a long-lasting impact, specifically in suburban and rural areas. The majority does not have the aptitude to validate any information, narrative, or content from an authentic source.
ChatGPT, Bard, and most other forms of LMLs are very intuitive and interactive; hence, they are gradually becoming the main sources of information and ideas, literally with no safety button.
Given Bangladesh's history with racial and religious violence, now more than ever, AI may become a trigger for the next chaos.
A treacherous path
When OpenAI started its journey in 2015, its core mission was to build artificial general intelligence (AGI) to benefit all of humanity. But as humanity progresses along its AI journey, the path seems fairly treacherous.
Over the years, AI and AI-powered services have become quite common. Generative AI platforms like DALL-E, Midjourney, Stable Diffusion, etc. are designing detailed and breathtaking creatives. Large language models (LLMs) like ChatGPT, Google Bard, Jurassic-1 Jumbo, etc. are organising and generating elaborate text-based responses on pretty much any topic on planet Earth and beyond.
Finance, trading, transportation, social media, e-commerce, national defense, law, and governance - all spectrums of human lives are influenced by some form of AI intervention. Even the healthcare sector is using AI to predict patient outcomes, assist in diagnosis, and personalize treatment plans.
It would not be unfair to say that we are currently living in a period of AI proliferation. Which is why it is very important at the very onset to address AI bias.
There are thousands of documented incidents to indicate the biased nature of AI.
Facial recognition software has a long history of providing inaccurate and often very offensive results in recognizing darker complexions, specifically 'Black' faces. In 2015, Google Photos mistakenly labeled black people as 'gorillas' '.
These days, police surveillance heavily relies on surveillance technology to filter, identify, and question possible suspects. Given the inefficiency of AI in surveillance technologies, law enforcement around the world is often disproportionately targeting people of colour.
AI influences the courts as well as law enforcement, leading to both misinformation and influence. The most well-known example of AI bias is the COMPAS (Correctional Offender Management Profiling for Alternative Sanctions) algorithm used in US court systems, which predicts the likelihood of a defendant being a 'repeat offender'.
Due to the inherent bias existing in the data sets used to train the model, it predicted twice as many false positives for black offenders (45%) than white offenders (23%) (source: ProPublica).
Healthcare and health insurance providers are heavily relying on AI to make important decisions regarding patient care and support. In October 2019, it was found by researchers that the algorithm used to predict which patients may need extra medical care preferred white patients over black patients .
This algorithm directly impacted more than 200 million patients by expediting or delaying much-needed medical care, based on decisions that were formed out of racial biases. .
There are many examples of how AIs unfairly influence someone's chances of getting recruited. A famous example is Amazon. Researchers found that Amazon's recruiting AI favored white and male candidates over females or candidates of color. Amazon trained its algorithm with resumes received over 10 years, which are predominantly white and male candidates.
Hence, the algorithm automatically became discriminatory against candidates based on gender and color. A similar issue was found with Twitter (renamed as X), its image-cropping algorithm preferred white faces over black faces when generating previews of photographs. This was due to Twitter's algorithm getting trained on data that were mostly white faces, resulting in biased treatment against certain races.
In addition to the above, the most common and reported AI biases are based on age, geography, socioeconomic background, disability, sexual orientation, and culture. AI bias is spreading like wildfire, it is impacting everything, and it is everywhere.
What's making AIs biased?
There are several factors. Algorithms are what fuel the AI; and data is what fuels the algorithm. Any algorithm shall be useless lines of code, if not trained and fed with a large volume of data. The core reason behind any algorithm's 'biased' behavior is the 'biased' datasets used to train them. Simply put, if the dataset is all about white males, AI will learn to prioritize white males and discriminate against others.
Additionally, if the team of engineers and coders that develop the AI is homogenous and lacks diversity, it is easy to overlook the biases in the data and they may fail to predict evident patterns of discrimination.
To add to that, the core philosophy and principles behind the development of AI are also very important. If the owner, investor, or board of an AI company is right-leaning, the final product coming out of that company is likely to reflect the philosophies of its patrons; and vice-versa.
Once there are elements or patterns of bias inside AI, often it becomes very difficult to neutralise such biases.
How does Bangladesh tackle AI bias?
While the entire world is trying and often failing to eradicate AI bias, a few immediate steps can go a long way. To begin with, there should be increased collaboration between the government and these large AI companies. Discussion regarding neutral and credible datasets for Bangla language should start right now.
There should be a collaborative effort to promote and patronise content creators; having a substantial volume of accurate content on the internet is critical. Educational curricula must recognise the existence of AI and take adequate actions to prevent potential AI related plagiarism.
Most importantly, collective awareness regarding the pros and cons of AI is absolutely essential. We must recognise that there are risks associated and take steps from both institutional and personal levels.
Sinha Ibna Humayun works in technology marketing and is a tech enthusiast.
Disclaimer: The views and opinions expressed in this article are those of the author and do not necessarily reflect the opinions and views of The Business Standard.