When Algorithms Go Wrong
A visual journey through the hidden flaws in AI systems
"The Past Haunts the Algorithm"
THE PROBLEM:
Tech's history = mostly male hires
THE RESULT:
⚠️ LESSON ⚠️
Historical Bias occurs when AI learns from data reflecting past discrimination.
Amazon scrapped the system in 2017.
"If You're Not in the Data, You Don't Exist"
BUT WAIT...
REAL WORLD TEST:
ACCURACY BY GROUP:
— MIT Media Lab Study, 2018
⚠️ LESSON ⚠️
Representation Bias happens when training data under-represents certain groups.
Solution: Improve representation & equality in datasets!
"Averages Lie"
City A
1,000,000 people
Avg: $50k
City B
100 people
Avg: $30k
City B's 100 struggling people
got ERASED by
City A's million!
They needed help but were invisible to the AI.
"The Test Itself is Broken"
THE REAL PROBLEM:
The benchmark was BIASED!
• 77% white faces
• 83% male faces
• Mostly well-lit photos
Real world ≠ Benchmark
Remember These!
AI learns from biased past data and repeats old discrimination.
Some groups are under-represented in training data.
Averages hide the struggles of minority groups.
Biased benchmarks give false confidence.
It also comes from the humans who build, train, and deploy AI systems.
That means YOU can help fix it.
Based on "Big Data, Big Responsibilities" notes
Real cases: Amazon AI Hiring Tool (2014-2017)
Gender Shades Study by Joy Buolamwini (2018)
🎓 Now YOU know about AI Bias!
Share this knowledge. Build better AI.