Is-AI-Truly-Impartial.pptx

Is AI Truly Impartial?

How Bias in Data Becomes Algorithmic Discrimination

In your opinion, does AI reduce bias?

✔️ YES

❌ NO

Raise your hand for your answer

The Bias No One Had Seen

Healthcare Algorithm – 200 Million Patients Involved

❌ The problem

The hospital AI used healthcare spend as a proxy for health

❌ The distortion

Black patients spend $1,800 less per year (systemic racism, not a health difference)

❌ The result

The algorithm labelled healthy Black patients as "high risk"

✅ After recalibration

Inclusion rose from 17.7% → 46.5%

This is not a technical error. It is bias embedded in the data.

Bias Starts Before AI

📊 DATA

🎯 MODELS

🧠 HUMANS

• Non-representative samples
• Historical inequalities
• Missing groups

• Proxy variables hide discrimination
• Design choices encode assumptions

"Remove gender = problem solved" ❌

• Confirmation bias in interpretation
• Annotators perpetuate stereotypes
• Selection bias in what to measure

The algorithm did not create the problem. It inherited it — and amplified it.

AI Learned What We Never Taught It

The Timeline

The Result

2014–2018

• It penalised words like "feminine"
• It flagged women’s college names
• It rejected common patterns in women’s CVs

The Problem

10 years of hiring data = 63% men in tech

What the AI Learned

Bias In, Bias Out — At Scale.

"Being male = better candidate"

Three Ways Bias Hides Itself

1️⃣ Historical Bias

📈 Past discrimination → The algorithm learns it as a "pattern"

Examples: CV screening, loans, criminal justice

2️⃣ Hidden Proxies

🔍 "Remove gender = solved" → No. The algorithm finds postcode, name, spending habits

Creates discrimination without mentioning protected attributes

3️⃣ Human Annotation

👥 Annotators (humans) label training data with their own bias

If 90% of annotators come from one demographic group, the model learns their perspective

You cannot remove bias by hiding it. Find another way.

THE HARSH TRUTH

🚫 MYTH

✅ REALITY

"Remove gender/race = problem solved"

Algorithms use proxies (postcode, name, spending). The bias returns.

"High accuracy = fairness"

The COMPAS algorithm was 60% accurate BUT labelled Black defendants as "high risk" 2 times more often. Technically accurate. Morally wrong.

"AI does not eliminate bias. It replicates it at scale."

Bias In. Bias Out. Thousands of times per second.

There Is No Perfect Algorithm

⚖️ Fairness ← → Accuracy

🎯 Demographic Parity ← → Equal Opportunity

💡 These are ETHICAL choices

(helping one group) vs. (overall)

(same outcomes) vs. (same rates)

NOT technical solutions

Fairness is not a mathematical problem. It is an ethical problem.

Conclusion: The Quiz Answer

So, who answered "NO" – AI does NOT reduce bias?

✔️ Bias is inevitable

❌ AI does not eliminate it

📈 AI amplifies it

What We CAN Do

01

02

03

Demand transparency

Diversify data and teams

Continuous monitoring

"How was it tested for bias?"

Not just checks at launch

"AI is not neutral. It is a mirror of our data — and therefore a mirror of society."

"Every bias in your data becomes a bias in your decisions. At scale. Forever."

"AI is only as fair as the data we give it. And we give it the world as it is — not as it should be."