AI summary by TwinMind
How bad data keeps us from good AI | Mainak Mazumdar
TED
10m 13s
AI Economic Impact and Current Limitations
- AI projected to add $16 trillion to global economy in next 10 years
- Economy will be built by computers and algorithms, not billions of people or millions of factories
- AI benefits: simplifying tasks, bringing efficiencies, improving lives
- AI failures: not delivering fair and equitable policy decision making
- AI becoming gatekeeper to economy: deciding job access and loan approvals
- AI reinforcing and accelerating bias at speed and scale with societal implications
Root Cause Analysis: Data vs Algorithm Problem
- Data scientist perspective: problem is not the algorithm but biased data
- Need urgent reset: focus on data instead of algorithms
- Currently spending time and money scaling AI at expense of designing high quality, contextual data
- Three key focus areas needed:
- Data infrastructure
- Data quality
- Data literacy
Case Study: Duke University PULSE AI Model
- June incident: embarrassing bias in Duke University AI model called PULSE
- Function: enhanced blurry images into recognizable photographs
- Problem: incorrectly enhanced non-white images into Caucasian images
- Cause: African American images underrepresented in training set
- Result: wrong decisions and predictions
- Pattern: AI frequently misidentifies Black persons despite improved AI methodology
Case Study: 2020 US Census Data Bias
- Census foundation for social and economic policy decisions
- Required to count 100% of US population
- Challenges: pandemic and politics of citizenship question
- Expected significant undercounting of minority groups (hard to locate, contact, persuade, interview)
- Undercounting introduces bias and erodes data infrastructure quality
2010 Census Undercounting Statistics
- 16 million people omitted in final counts
- Equivalent to total population of Arizona, Arkansas, Oklahoma, and Iowa combined
- 1 million kids under age 5 undercounted
- Minority undercounting common in national censuses globally
- Causes: harder to reach minorities, mistrust toward government, political unrest areas
International Examples
- Australian census undercounted Aboriginal population by 17.5%
- 2020 undercounting estimated much higher than 2010
- Massive implications for bias
Census Data Infrastructure Importance
- Most trusted, open, publicly available rich population data
- Businesses have proprietary consumer information, but census provides definitive public data
- Census covers: age, gender, ethnicity, race, employment, family status, geographic distribution
- Foundation of population data infrastructure
- When minorities undercounted: AI models for public transportation, housing, healthcare, insurance overlook communities requiring services most
Data Quality Solutions
- First step: make database representative of age, gender, ethnicity, race
- Census importance requires effort to count 100% of population
- Investing in data quality and accuracy essential for making AI possible for everyone, not just few and privileged
Nielsen Data Collection Examples
Shanghai and Bangalore Retail Stores
- Field visits to collect retail sales data from stores outside Shanghai and Bangalore
- Drove miles outside city to find small, informal, hard-to-reach stores
- Could have selected city stores with easy electronic data integration (cheap, convenient, easy)
- Focus on quality and accuracy of rural store data
- Rationale: 40% Chinese and 65% Indians live in rural areas per International Labor Organization
- Excluding 65% of Indian consumption creates massive bias
- Urban bias leads to wrong decisions in pricing, advertising, marketing, and rural policy for health investments
US TV Measurement Panel
- Visited remote trailer park homes in Oregon and NYC apartments
- Invited homes to participate in Nielsen panels (statistically representative sample)
- Mission: include everybody in measurement
- Focus: Hispanic and African homes using over-the-air TV reception via antenna
- These homes constitute 15% of US households (45 million people)
- Commitment to quality means collecting information from hard-to-reach groups
- Impact: without this data, marketers and brands cannot reach important minority populations
- Revenue impact: without ad revenue, broadcasters like Telemundo/Univision cannot deliver free content including news media (foundational to democracy)
Call to Action
- Once-in-lifetime opportunity to reduce human bias in AI starts with data
- Instead of racing to build new algorithms, focus on building better data infrastructure
- Mission: make ethical AI possible through better data infrastructure
