Decolonizing AI: Are we asking the right questions?

This article was originally published on April 12, 2024 on LinkedIn.

Sharing thoughts on Kalika Bali's latest research work at Microsoft with calling out gender bias x Hindi, due for publication soon. While I completely agree decolonization is truly required across the board – from impact metrics frameworks and algorithms to integrating respect for cultural and linguistic diversity – the challenge lies in its connection to power dynamics. AI outputs are just revealing it - how humans carry hidden stereotypes beneath the sophistication of degree and pedigree.

When was the last time we truly encouraged the previous and now the younger generation to write or publish in their native language, let alone speak it at home? Haven't we historically placed a higher value on English, often seeing it as a marker of intelligence? This inter-generational consequence of colonialism is still prevalent and have influenced how we behave at the intersection of tech x social impact x equity.

Bias tolerance is often higher among those with fewer resources, making it a complex issue. It's harder to rally support for change if those most affected are more tolerant of the issue.

And one’s comfort with it is a direct reflection of privileges as it is invisible to those who benefit from it.

Of course, this doesn't mean they are okay. If they were so easy to fix, they would have been addressed in English long ago. Because they are difficult to measure and highly subjective. Even within well-resourced languages like English (if we consider it as the universe), strong gender and racial biases still exist.

For instance, the research findings published by Andrew Sharp, University of Delaware - where a group of scientists fed 8,000 articles from The New York Times and Reuters, offering the headlines as prompts to let LLM (including Grover, Cohere, Meta's LLaMa and several different versions of OpenAI's ChatGPT) create their content. For analysis, used TextBlob to score disrespect, rudeness and profanity at the sentence level unlike earlier studies focused on words frequency.

They found that AIGC (AI generated content) exhibited a 40-60% increase in bias against minorities, particularly Black population and women. Even leading discriminatory prompt to write racially incendiary articles were effortlessly accepted to generate malicious content without safeguarding. If it existed, in newer versions ChatGPT, it was easier to manipulate and circumvent. This is published two days ago.

When I read such case studies, I keep asking myself, isn’t the machine a blank slate? Isn't it built and fed possibly by some of the brightest and smartest people.

And this thought doesn't leave me there. I keep wondering, despite all of the education and global exposure, aren’t these AI outputs mirror of the subconscious which as a society we have either failed to fix or never cared to address?

This is why framing AI bias solely as a tech issue is misleading. It's certainly much deeper, broader and layered than that. Nonetheless, my biggest question (and I suspect others share it) is how will we progress unless we address bias in journalism and publication practices since it essentially feeds the data for LLMs.

I wonder, how focusing on symptoms alone would solve it when the root of linguistic rot is so strong? It begs the question – is monitoring journalism the role of tech companies? Likely not. Perhaps they can provide powerful tools for enhancing precision, source integrity and productivity.

From a commercial standpoint, I understand that a product needs to be tested with bare minimum first and requires large data sets to test functionality at scale. Perfecting a product takes years and sharp focus to let the user's right be at the center of it. Forbes has been quite vocal on how ensuring Diversity, Equity, and Inclusion (DEI) throughout the process is already a challenge, adding another layer of ‘ethical’ complexity would surely demand more resources and intentionality.

Now that AI products have shown significant promise and traction, perhaps it's the right time to pause, reflect, and consider its finer nuances. While profit projections are important, an open dialog to discuss the potential pitfalls and the quagmire of inequity is becoming increasingly equally critical. How are we equipping ourselves to mitigate AI risks proactively during dignified deployment within our respective organizations? Are we fully aware of the scenarios should we fail to get it right?

Unless, there's a deliberate executive mandate or conscious focus from all three – private sector, public administration, and philanthropy through new grants to initiate debias and conduct social audit, it's an uphill battle.

In hindsight, I strongly believe that one of the major contributing factor is the pipeline of the STEM workforce and how the roles are assigned within the tech industry. Repetitive tasks in the assembly line are often assigned to tech workers from the Global South (without agency to question design aspects), while innovative and design-focused tasks are dominated by "disruptive innovators" from the Global North (who are encouraged to question, explore, and embrace failure).

What can be done to address the power dynamics that exist within the STEM workforce, leadership, and the roles assigned to tech workers from the Global South?

Certainly, AI is bringing conversations to the forefront which were running on the assumption that it's sorted and have been mutually agreed upon in the name of inclusive economic growth.

Have you thought about this?

PS: This is the end of my thinking capacity for a Friday afternoon. Will resume when my cup will get filled with more questions and conflicting literature. Till then, take care you all and stay curious.

(Silence = promoting biases, regardless of industry).

(Image credit: Unsplash)