AI Safety & Ethics

AI Safety & Ethics

From Circuits to Secrets in a day

Understand how AI works, why it’s unpredictable, and how to keep it safe in a world that’s playing catch-up.

AI is no longer a niche research field—it’s a mainstream force shaping politics, industry, and everyday life. Yet the people building and regulating it often understand it only partially. This course blends an accessible technical primer with a candid look at geopolitical tensions, regulatory loopholes, and ethical dilemmas. Led by an instructor with over 30 years of industry experience, we’ll combine serious insights with engaging, relatable examples—from AI-driven global rivalries to deepfake mischief—to explore how to use AI responsibly in the real world.

Learning Outcomes

By the end of the day, participants will:

• Understand the basics of how AI systems—especially large language models (LLMs)—are trained and operate.
• Explain why AI behaves unpredictably and remains a “black box” despite impressive capabilities.
• Describe key global AI political and regulatory positions, including the EU’s military exemption and the Trump administration’s deregulatory stance.
• Recognize the role of U.S.–China AI rivalry in shaping technology access and standards.
• Evaluate the ethical and legal challenges of AI training data collection without consent.
• Identify threats from deepfakes and apply safety measures to avoid being fooled.
• Engage critically and creatively with AI ethics issues through interactive activities.

Prerequisites

• Basic familiarity with AI concepts and current events.
• No coding knowledge required—just curiosity and critical thinking.

Training Outline

1. AI 101: Under the Hood

• What AI Really Is:
  • Narrow AI vs. General AI.
  • Neural networks, machine learning, and deep learning—how they differ.
• How AI Learns:
  • Data collection, preprocessing, and labeling.
  • Training using backpropagation and gradient descent.
  • Role of GPUs/TPUs in computation.
• Large Language Models (LLMs):
  • Tokenization, embeddings, attention mechanisms, and transformers.
  • The scaling laws that made modern AI possible.
  • Why bigger data + more compute ≠ always better results.

2. AI’s Mystery Sauce

• The Black Box Problem:
  • Why models can solve problems without us fully knowing how.
  • Complex feature representations that humans can’t intuitively interpret.
• Geoffrey Hinton’s Concerns:
  • His suspicion of backpropagation’s long-term viability.
  • Calls for rethinking AI foundations before unexpected failures emerge.
• Non-Determinism in Action:
  • Why repeated prompts to the same model can produce different outputs.
  • Impact on safety, reliability, and accountability.
• Implications:
  • Limits of interpretability tools like SHAP and LIME.
  • Risk assessment in high-stakes environments.
  • Overview of recent AI safety incidents: deceptive LLM behavior and evaluation shortcomings

3. Geopolitics & Regulation

• Key global initiatives:
  • EU's AI Act (in force since August 1, 2024) and applicability timelines
  • UNESCO’s AI Ethics recommendations
  • California’s policy report on irreversible AI risks and trust‑but‑verify approach
  • Other governance efforts (AI Safety Institutes, global summits)
• EU’s Double Standard: Civilian AI oversight vs. military exemptions.
• Trump’s Deregulation Push: Prioritizing AI dominance over safety rules.
• U.S.–China Cold War: Chips, data, and digital infrastructure supremacy.

4. Data Ethics

• The Consent Problem: How most LLMs used public data without permission.
• Legal Grey Zones: GDPR’s “legitimate interest” loopholes vs. privacy rights.
• Cultural Impacts: Artists, journalists, and domain experts losing control of their work.

5. Deepfakes

• How They Work: GANs, diffusion models, and real-time manipulation.
• Real Dangers: Scams, political disinformation, revenge porn.
• Personal Safety: Verification habits, detection tools, skepticism training.
• Fun Demo: Light-hearted but safe deepfake examples to illustrate risk without harm.

6. Building Ethical Assurance into AI Systems

• Frameworks for ethics‑based auditing and continuous oversight.
• Implementing transparency, explainability, accountability, and governance layers aligned with best practices and standards.
• Discussion: evaluating tools like
  • HELM Safety,
  • AIR‑Bench,
  • FACTS for factuality and safety testing

7. Interactive Wrap-Up

• Scenario Game: Choose your role—regulator, engineer, activist, or everyday citizen—and craft a mini AI safety plan.
• Meme-Making: Teams create ethical memes from course takeaways.
• Group reflection on AI’s future and where responsibility truly lies.