How bad data keeps us from good AI | Mainak Mazumdar

Episode Summary

Executive Summary: Minok Mazumdar argues that AI’s biggest problem is not the algorithm but biased, incomplete data. He shows how underrepresentation in datasets can distort decisions in hiring, lending, public policy, and media, and calls for an urgent reset focused on data infrastructure, data quality, and data literacy to make AI fairer and more inclusive.

Main Topics: AI’s promise and its bias problem (Priority: 5/5): The talk opens by acknowledging AI’s economic and practical benefits while warning that it can amplify existing social bias at scale when used for high-stakes decisions. Biased data, not just biased algorithms (Priority: 5/5): Mazumdar argues that flawed outcomes usually stem from the data used to train systems, not the algorithms themselves, making data collection and representation the real priority. Census undercounting as a data infrastructure failure (Priority: 5/5): He uses census undercounts to show how missing or inaccurate population data can distort public policy and downstream AI systems that rely on those counts. Rural and informal communities are often excluded (Priority: 4/5): Examples from retail measurement in China and India illustrate how ignoring rural populations creates urban bias in business decisions and policy. Hard-to-reach households matter economically and socially (Priority: 4/5): He describes efforts to include minority and antenna-TV households in measurement panels because they represent a large audience and support media ecosystems and democracy. A call for better data infrastructure, quality, and literacy (Priority: 5/5): The proposed solution is an urgent reset: invest in representative data systems, improve measurement practices, and build data literacy so AI serves everyone.

Key Arguments: AI can create enormous economic value, but without representative data it will also scale discrimination and exclusion. The root cause of many harmful AI outcomes is biased or incomplete training data, not the algorithm itself. Public datasets like the Census are foundational infrastructure; if they undercount minorities, AI and policy built on them will inherit that bias. Data quality requires deliberate investment in collection, definition, and measurement, even when it is slower and more expensive than using convenient existing data. Excluding rural, informal, or hard-to-reach populations leads to distorted business decisions, misallocated public resources, and inequitable services. Inclusive measurement is essential not only for fairness but also for markets, media revenue, and democratic access to information.

Data Points: Projected AI economic impact: $16 trillion - AI could add this amount to the global economy in the next 10 years. 2010 U.S. Census omissions: 16 million people - Number of people omitted in the final 2010 Census counts. 2010 undercount of children under 5: about 1 million - Young children were undercounted in the 2010 Census. Australian Census 2016 undercount: 17.5% - Aboriginal and Torres Strait populations were undercounted. China rural population share: 40% - Used to illustrate the importance of rural retail data in China. India rural population share: 65% - Used to show how excluding rural consumers can heavily bias models and decisions. U.S. households using over-the-air TV: 15% - Households receiving TV via antenna, including Hispanic and African American homes, that Nielsen sought to include in panels. People represented by those households: about 45 million - Approximate population size of the 15% of U.S. households using over-the-air TV.

Pivotal Quotes: "It’s not the algorithm, but the biased data that’s responsible for these decisions." — Minok Mazumdar: Core thesis of the talk, explaining the source of harmful AI outcomes. "We need an urgent reset. Instead of algorithms, we need to focus on the data." — Minok Mazumdar: Call to action for shifting investment and attention toward data infrastructure and quality. "Our once-in-lifetime opportunity to reduce human bias in AI starts with the data." — Minok Mazumdar: Closing message emphasizing that ethical AI depends on better data foundations.

Implications: AI fairness depends on representative, high-quality data. Companies and governments must invest in data infrastructure and inclusion now, or AI will keep amplifying inequity in hiring, lending, policy, media, and public services.