How to keep human biases out of AI | Kriti Sharma

Episode Summary

Executive Summary: Kriti Sharma argues that the most urgent AI danger is not job loss or robot takeover, but algorithmic bias that reproduces sexism, racism, and elitism in everyday decisions. She calls for awareness, diverse builders, and diverse data, while showing AI’s potential to improve healthcare, safety, and access if built responsibly.

Main Topics: The real AI risk is bias, not science-fiction takeover (Priority: 5/5): Sharma reframes public fear away from sentient robots and toward the immediate harm caused when AI makes consequential decisions using biased human data. Algorithms shape high-stakes life outcomes (Priority: 5/5): AI is used in hiring, lending, insurance, credit scoring, and performance reviews, meaning biased models can affect employment, finances, and opportunity. Bias is learned from human behavior and past data (Priority: 5/5): She explains how machine learning absorbs patterns from biased historical choices, then amplifies them into discriminatory outputs. Everyday interfaces reinforce gender stereotypes (Priority: 4/5): Female-voiced assistants and male-coded professional systems teach children and users that service roles are feminine and authority roles masculine. Diversity in AI teams is essential (Priority: 5/5): Sharma emphasizes that people from varied genders, races, and backgrounds are needed to challenge blind spots and build fairer systems. AI can be a force for social good (Priority: 4/5): She closes by highlighting real-world uses of AI in maternal health and domestic violence support, arguing the technology can help make the world more equal.

Key Arguments: AI is already making consequential decisions about people, often without accountability, and those decisions can encode discrimination. The most pressing issue is not hypothetical superintelligence but present-day bias in algorithms used for hiring, lending, and assessment. If humans would reject a racist or sexist decision-maker, they should reject the same outcome when produced by a machine. AI systems reflect the data and assumptions they are trained on, so biased human history can become automated bias. Voice assistants and product design reinforce stereotypes by making women the default servant-like AI and men the default authoritative AI. Diverse teams improve AI because they bring different perspectives, challenge elitist assumptions, and identify real-world problems worth solving. AI should be built with diverse data and values so it can address important needs such as healthcare access and safety for vulnerable populations.

Data Points: Walk time to prenatal clinic: 17 hours - A pregnant woman in the Democratic Republic of Congo may need to walk this far to reach a rural prenatal clinic, illustrating a use case for mobile AI health support. Women facing domestic violence in South Africa: 1 in 3 - Sharma cites this as a population that could benefit from discreet AI support and emergency assistance. Code acceptance increase when gender hidden: 4% more than men - She references a study where women coders’ code was accepted more often when their gender was concealed.

Pivotal Quotes: "There is a more pressing danger, a bigger risk with AI that we need to fix first." — Kriti Sharma: She introduces her central thesis that AI bias is more urgent than sensational fears about robots. "A black man is more likely to be a repeat offender than a white man." — Kriti Sharma: An example of biased machine outputs that reflects discriminatory patterns learned from human data. "You don't need to look like a Mark Zuckerberg. You can look like me." — Kriti Sharma: Her closing message encouraging broader participation in AI creation beyond a narrow tech stereotype.

Implications: Listeners should treat AI bias as a present-day governance issue, push for diverse teams and datasets, and demand accountability for algorithmic decisions. The industry’s future depends on building AI that serves everyone, not just the historically privileged.