# Global Bias in AI: When Western Data Shapes the World - Masuma Shariff - NDC Copenhagen 2026

## Executive summary

This talk explores systemic bias in AI models caused by uneven global data representation, arguing that when training data overwhelmingly originates from Western nations (the Global North), the resulting systems fail to accurately serve or represent developing populations. The speaker details how biases accumulate across the entire AI pipeline—from data collection and annotation to model deployment—using case studies in image recognition, medical diagnostics, finance, and governance to illustrate real-world consequences like misdiagnosis and financial exclusion.

## Key takeaways

- Bias is not static; it multiplies across the AI pipeline.: Bias does not simply drop into a system; it accumulates and scales at every stage (data collection, labeling, training). The problem is multiplicative, meaning each stage passes the bias forward amplified, like a snowball rolling downhill. This requires mindful design before writing any code.
- Global data gaps lead to systemic failures in critical domains.: Case studies show that Western-centric standards fail globally: Pulse oximeters missed dangerously low oxygen levels in Black patients due to melanin absorption [~1650]; and automated credit scoring models flagged economically active people as uncreditworthy because they rely on communal assets rather than formal bank accounts.
- The AI system is only as good as the data it learns from, which itself is a cultural artifact.: Data is not neutral; it reflects the history and blind spots of those who gathered it. The speaker highlights that many global standards (e.g., Fitzpatrick scale for skin tone) were built for limited demographics, leading to non-inclusive outcomes when applied universally.

## Technical details

- Data Annotation/Labeling: The process of labeling data (e.g., identifying a face, tumor, or profanity) is performed by human 'data annotators,' often located in the Global South (Kenya, Philippines). These workers are paid extremely low wages ($1–$5 per label), raising ethical concerns about labor exploitation.
- Model Collapse: This occurs when a model continuously trains on its own output, creating a feedback loop. The model drifts further from reality and can create an 'AI echo chamber,' leading to issues like hallucinations.
- Algorithmic Colonialism: This term describes the economic form of bias, where Western tools (like automated credit scoring) fail in non-Western economies because they are designed to look for predictors (e.g., formal bank accounts, utility bills) that do not exist or are irrelevant in communal financial systems.
- Bias Benchmarks and Counterwork: To combat bias, researchers are developing specialized tools. Examples include Joy Boamini's 'Gender Shades,' which quantified error rates in AI misgendering based on skin tone, and Adobe Firefly, which is trained on licensed and curated data to control for biases.

## Practical implications

- Design systems with 'fairness' as an architectural requirement, not merely an afterthought.
- When designing data pipelines, explicitly question the demographic source and representation of all training data (data collection phase).
- Implement rigorous validation steps to test model performance across diverse populations and edge cases, rather than relying solely on aggregate metrics.
- Be aware that systems trained in one geographical or cultural context may fail catastrophically when deployed elsewhere (transfer bias).

## Topics

Artificial Intelligence Ethics, Data Bias, Global Development Technology, Algorithmic Fairness, Systemic Discrimination, Invisible Women, The Atlas of AI (or similar works by Sophia Noble), M-Pesa / Mobile Banking Systems

Source: https://www.youtube.com/watch?v=FCgP5xLsHgQ