Lecture 10 — Bias, Ethics & Human-in-the-Loop (HITL) in Multimodal AI

~3–4 hours (human-centered AI lecture)

🌍 Why This Lecture Matters More Than Any Other

Power without ethics is danger.

Modern AI systems:

see humans
read private documents
make recommendations
influence decisions
act autonomously

Without ethics:

bias scales
harm multiplies
trust collapses

Ethics is not optional. It is engineering.

⚖️ What Is Bias in AI?

Bias is systematic unfairness that disadvantages individuals or groups.

Bias can appear in:

data
models
evaluation
deployment
human usage

AI does not create bias — it amplifies it.

🧠 Types of Bias (Must-Know)

1️⃣ Data Bias

Skewed demographics
Missing populations
Historical inequality

Example:

Face recognition trained mostly on light-skinned faces.

2️⃣ Annotation Bias

Subjective labels
Cultural assumptions
Inconsistent annotators

3️⃣ Model Bias

Shortcut learning
Spurious correlations
Overgeneralization

4️⃣ Deployment Bias

Model used outside training context
Different population
High-stakes environment

🖼 Bias in Multimodal Systems

Multimodal AI adds new bias risks:

Vision stereotypes
Language prejudice
Accent discrimination
Cultural misinterpretation

Example:

Describing professions differently based on gender in images.

🎥 Temporal & Contextual Bias (Video)

Video models may:

Misinterpret behavior
Infer intent incorrectly
Over-police certain actions

Seeing is not understanding.

⚠️ Ethical Risks of Multimodal AI

Risk	Example
Surveillance	Facial recognition misuse
Privacy	Reading personal documents
Manipulation	Deepfakes
Automation bias	Blind trust in AI
Exclusion	Accessibility gaps

🧠 Ethics ≠ Rules

Ethics involves:

Values
Context
Trade-offs
Human judgment

Ethical AI is not “always right” — it is accountable.

👥 What Is Human-in-the-Loop (HITL)?

HITL = Humans actively guide, verify, and override AI systems.

HITL is used when:

Stakes are high
Errors are costly
Context matters
Accountability is required

🔁 HITL Interaction Modes

1️⃣ Human-in-the-Loop

Human approves or corrects outputs.

2️⃣ Human-on-the-Loop

Human monitors and intervenes if needed.

3️⃣ Human-out-of-the-Loop

Fully autonomous (⚠️ risky).

🧩 Where HITL Fits in AI Pipelines


Data → Model → Prediction → Human Review → Decision

Examples:

Medical diagnosis
Legal document review
Loan approval
Content moderation

🐍 Python: HITL Pattern (Conceptual)

prediction = model(input)

if confidence < threshold:
    send_to_human(prediction)
else:
    accept(prediction)

Uncertainty is a signal, not a failure.

🧠 Designing HITL Systems Well

Good HITL systems:

Are transparent
Minimize human fatigue
Respect human expertise
Log decisions
Learn from corrections

Bad HITL systems:

Treat humans as rubber stamps
Overload reviewers
Hide model uncertainty

⚖️ Fairness Metrics (High-Level)

Common notions:

Demographic parity
Equal opportunity
Equalized odds

⚠️ Important:

You cannot satisfy all fairness definitions simultaneously.

Ethics requires choices.

🧠 Accountability & Responsibility

Key questions:

Who is responsible for errors?
Who audits the system?
Who can appeal decisions?
Who benefits?

AI shifts power — ethics decides where it goes.

📜 Regulation & Governance (Brief)

Trends:

AI Act (EU)
Model cards
Data sheets
Audit trails

Purpose:

Transparency
Safety
Human rights protection

🧠 Research Insight

The most dangerous AI is not malicious — it is confident, biased, and unchecked.

Future AI research must integrate:

Ethics-by-design
Value alignment
Continuous monitoring
Human agency

🧪 Student Knowledge Check (Hidden)

Q1 — Objective

Does AI create bias?

Answer

No. It amplifies existing bias in data and systems.

Q2 — MCQ

Which is NOT a type of bias?

A. Data bias B. Annotation bias C. Hardware bias D. Deployment bias

Answer

C. Hardware bias

Q3 — MCQ

When is HITL most important?

A. Low-risk chatbots B. Image filters C. High-stakes decisions D. Games

Answer

C. High-stakes decisions

Q4 — Objective

What is automation bias?

Answer

Humans over-trusting AI outputs without critical thinking.

Q5 — Objective

Why is uncertainty important?

Answer

It signals when human review is needed.

🌱 Final Reflection

If AI becomes very powerful, what must always remain human?

Values, responsibility, empathy, and moral judgment.

✅ Key Takeaways

Bias is systemic, not accidental
Multimodal AI increases ethical risk
HITL is a design principle, not a patch
Fairness requires trade-offs
Humans must remain accountable

Last updated on 2026