Lab 13: The Ethics of AI — Bias, Hallucinations, Deepfakes

Objective

Understand the ethical challenges that come with deploying AI in the real world. By the end you will be able to:

• Identify and categorise AI harms across multiple dimensions

• Explain why AI hallucinations occur and how to mitigate them

• Describe the deepfake threat landscape

• Apply an ethical framework when building or deploying AI systems

---

Why Ethics in AI Matters

AI systems are not neutral tools. They amplify human decisions — and human biases — at machine scale. A biased human makes biased decisions for the people they interact with. A biased AI system makes biased decisions for millions of people simultaneously, often invisibly, often automatically.

"The discourse on AI ethics has moved from 'should we worry?' to 'what do we do about the harms that are already happening?'" — Kate Crawford, Atlas of AI (2021)

---

The Harm Taxonomy

AI harms fall into several categories:

AI HARMS
│
├── Allocative harms
│   └── AI denies resources/opportunities: loans, jobs, bail, healthcare
│
├── Representational harms  
│   └── AI reinforces stereotypes: image search "CEO" returns mostly men
│
├── Quality-of-service disparities
│   └── AI works worse for some groups: voice recognition fails on accents
│
├── Informational harms
│   └── AI generates false information: hallucinations, misinformation
│
├── Privacy harms
│   └── AI infers sensitive information: sexual orientation from photos
│
└── Autonomy harms
    └── AI manipulates or surveils: Cambridge Analytica, social scoring

---

Algorithmic Bias: Case Studies

Case 1: COMPAS Recidivism Scoring (USA)

COMPAS was used by US courts to predict whether defendants would reoffend, informing bail and sentencing decisions.

ProPublica's 2016 analysis found:

• Black defendants: 44.9% incorrectly flagged as high risk (false positive rate)

• White defendants: 23.5% incorrectly flagged as high risk (false positive rate)

• Nearly 2× disparity in false positive rates

The algorithm never saw race — but neighbourhood, employment history, and social connections are correlated with race due to systemic inequality. Historical injustice was encoded as "risk factors."

Case 2: Facial Recognition in Law Enforcement

MIT Media Lab study (2018, Joy Buolamwini):

• Light-skinned men: 0.8% error rate

• Dark-skinned women: 34.7% error rate

Real-world consequence: Robert Williams (2020) — first documented case of wrongful arrest due to facial recognition. Detroit police arrested him based on a false match despite him having an alibi for the entire period.

Case 3: Amazon Hiring Algorithm (2018)

Amazon trained a CV screening model on 10 years of hiring decisions (predominantly male in tech roles). The model learned to penalise:

• CVs with the word "women's" (women's chess club, women's cricket)

• Graduates of all-women's colleges

Amazon scrapped the system when the bias was discovered. It was never used in production.

---

Hallucinations: When AI Confidently Lies

An AI hallucination is when an LLM generates plausible-sounding but factually incorrect information, stated with complete confidence.

User: "Who is the CEO of InnoZverse?"
LLM:  "The CEO of InnoZverse is Dr. James Harrison, who founded the company 
       in 2019 after leaving Google DeepMind..."
      [HALLUCINATION — Dr. James Harrison does not exist; 
       InnoZverse's actual leadership is entirely fabricated here]

Why hallucinations happen:

1. Training objective mismatch — the model is trained to produce fluent, coherent text, not to produce accurate text. Fluency and accuracy are not the same thing.

2. No factual grounding — LLMs don't retrieve facts from a database; they generate text that matches statistical patterns from training data.

3. Knowledge cutoff — events after the training cutoff are unknown to the model, but it may still generate plausible-sounding responses.

4. Long-tail knowledge — rare facts appear rarely in training data; the model has weak signal and is more likely to confabulate.

# Detecting potential hallucinations: confidence calibration
# Well-calibrated models: when they say 90% confident → actually correct 90% of time

# Mitigation 1: RAG (Retrieval-Augmented Generation)
def grounded_answer(question: str, knowledge_base) -> str:
    # Step 1: retrieve relevant facts
    relevant_docs = knowledge_base.search(question, top_k=5)
    
    # Step 2: answer ONLY from retrieved context
    prompt = f"""
    Answer the question using ONLY the information below.
    If the answer is not in the context, say "I don't know."
    
    Context:
    {relevant_docs}
    
    Question: {question}
    """
    return llm(prompt)

# Mitigation 2: Source citation requirement
prompt = """
Answer the question and cite your source for each claim.
If you are uncertain about a fact, say "I believe..." or "I'm not certain but...".
Never state uncertain information as fact.

Question: {question}
"""

Hallucination rates (2024 benchmark — TruthfulQA):

• GPT-4: ~59% truthful

• Claude 3: ~64% truthful

• GPT-3.5: ~47% truthful

• Best open-source models: ~55% truthful

These numbers are improving rapidly — but no model is hallucination-free.

---

Deepfakes: The Synthetic Media Crisis

Deepfakes are AI-generated media — video, audio, images — that depict real people doing or saying things they never did.

The technology progression:

Year	Capability
2017	Face-swap videos (poor quality, obvious artefacts)
2019	Real-time face-swap; voice cloning from 5 minutes of audio
2021	Full-body pose transfer; any face on any body
2023	Video generation from text (not yet mainstream)
2024	Photorealistic video + voice in seconds; Sora-level quality

Documented harms:

1. Non-consensual intimate imagery (NCII): 96% of deepfake videos online are NCII targeting women. Victims lose employment, relationships, mental health.

2. Political disinformation: Deepfake of Ukrainian President Zelenskyy "surrendering" distributed on social media (2022). Quickly debunked but viewed millions of times.

3. Financial fraud: CEO voice clone used to authorise £200,000 wire transfer (UK, 2019). Estimated $25M lost to deepfake fraud in 2023.

4. Eroding reality: "Liar's dividend" — knowing deepfakes exist allows bad actors to deny authentic footage. Real evidence becomes deniable.

Detection approaches:

# Deepfake detection signals (conceptual)
detection_features = {
    "facial_inconsistencies": [
        "eye blinking rate (deepfakes often blink unnaturally)",
        "facial boundary artefacts at hair/skin edges",
        "skin texture inconsistency across frames",
        "gaze direction not tracking correctly"
    ],
    "audio_visual_sync": [
        "lip movement doesn't match audio phonemes",
        "audio background noise inconsistent with video setting"
    ],
    "metadata": [
        "compression artefacts inconsistent with claimed source",
        "GAN fingerprints (spectral analysis of pixel statistics)"
    ]
}

# Current detection models: ~80-90% accuracy on lab datasets
# ~50-60% accuracy on in-the-wild deepfakes
# Arms race: detectors improve → generators improve to evade detection

Defensive measures:

• C2PA (Coalition for Content Provenance and Authenticity) — digital watermarking standard

• Content credentials embedded at capture time by cameras and devices

• Adobe, Microsoft, Google, Sony all implementing C2PA

---

Accountability: Who Is Responsible?

When an AI system causes harm, who is liable?

Patient harmed by AI diagnosis
         │
    Who's responsible?
         │
    ┌────┴───────────────────────────┐
    │                                │
Hospital (deployed it)   AI Company (built it)
         │                          │
    Did they validate       Was the training data
    it for this use case?   representative?
    Did they get consent?   Did they document limits?
         │                          │
    Doctor (used it)         Data labellers
                             (who labelled training data?)

The EU AI Act (2024) — the world's first comprehensive AI law:

Risk Level	Examples	Requirements
Unacceptable	Social scoring, real-time biometric surveillance	Banned
High	Medical AI, credit scoring, recruitment AI, CCTV	Conformity assessment, human oversight, transparency
Limited	Chatbots	Must disclose AI nature
Minimal	Spam filters, AI video games	No requirements

---

Building Ethically: A Framework

# Ethical AI checklist (before deployment)

ETHICAL_CHECKLIST = {
    "fairness": [
        "Audit model performance across demographic groups",
        "Document any disparate impact",
        "Consider whether the use case is appropriate at all",
    ],
    "transparency": [
        "Document training data sources and known biases",
        "Provide explanations for consequential decisions",
        "Disclose when users are interacting with AI",
    ],
    "privacy": [
        "Collect minimum necessary data",
        "Obtain informed consent for AI use",
        "Enable data deletion (right to erasure)",
    ],
    "accountability": [
        "Designate responsible AI owner",
        "Establish incident response process",
        "Create feedback channel for affected individuals",
    ],
    "robustness": [
        "Test on distribution shifts (out-of-domain inputs)",
        "Monitor for performance degradation in production",
        "Define human override mechanism",
    ]
}

for category, checks in ETHICAL_CHECKLIST.items():
    print(f"\n[{category.upper()}]")
    for check in checks:
        print(f"  ☐ {check}")

---

Summary

Issue	Root Cause	Mitigation
Algorithmic bias	Biased training data; proxy variables	Diverse data; fairness auditing; human review
Hallucinations	Training objective ≠ accuracy	RAG; source citation; uncertainty quantification
Deepfakes	Generative model capability	C2PA watermarking; media literacy; detection tools
Lack of accountability	No clear legal framework	EU AI Act; internal governance; impact assessments

---

Further Reading

• Atlas of AI — Kate Crawford

• Algorithmic Justice League

• EU AI Act Full Text

• Partnership on AI — Responsible AI Practices

• Deepfake Detection Challenge