AI Systems, Concepts and Governance Workshop

Please contact me if you would like me to deliver this workshop for your institution. Otherwise, develop this for your own educational use with attribution. 

Core Learning Outcomes

By the end of the workshop, participants should be able to:

1. Explain how modern AI systems (especially foundation models) are trained and deployed at a conceptual level
2. Distinguish between model capabilities, training data, and inference behaviour
3. Identify common failure modes (hallucination, bias, distribution shift, reward hacking as conceptually relevant analogue)
4. Understand the basic landscape of AI governance (technical safety, policy, corporate governance, regulation)
5. Critically evaluate claims about AI risk, capability, and alignment
6. Form a grounded, non-hype mental model of what current AI systems can and cannot do

MODULE 1 — What AI Is (and What It Is Not)

Objective

Build a precise conceptual foundation for modern machine learning systems without requiring technical prerequisites.

Topics

• Historical framing: symbolic AI → machine learning → deep learning → foundation models
• What "AI" means in contemporary usage (and why it is ambiguous)
• The core paradigm shift: from rules → data-driven pattern learning
• What a model actually is (parameters, function approximation intuition)
• Training vs inference (critical conceptual distinction)

Key Concepts Introduced

• Dataset
• Model parameters
• Loss function (intuitive only)
• Generalisation vs memorisation

Activity

Participants map "common public claims about AI" (e.g. "AI thinks", "AI is objective", "AI is sentient") onto correct/incorrect conceptual categories.

Output

A shared glossary of foundational terms in plain language.

MODULE 2 — How Modern AI Systems Work (Foundation Models)

Objective

Develop a working mental model of large-scale AI systems such as large language models.

Topics

• Neural networks as function approximators (intuitive geometry, not maths-heavy)
• Tokenisation and language modelling intuition
• Training pipeline: data collection → pretraining → fine-tuning → alignment layers
• What makes foundation models different (scale, transferability, generality)
• Emergent behaviour (what it means and what it doesn't mean)

Key Concepts Introduced

• Tokens
• Transformer architecture (conceptual level only)
• Pretraining vs fine-tuning
• Reinforcement learning from human feedback (RLHF) at a high level
• Embeddings (intuitive similarity space framing)

Activity

"Deconstruct a chatbot": participants trace how a prompt becomes output step-by-step conceptually.

Output

A diagram of the lifecycle of a prompt through a model system.

MODULE 3 — Failure Modes, Limits, and Misconceptions

Objective

Develop critical literacy around AI reliability, safety, and epistemic limits.

Topics

• Hallucination as probabilistic text generation (not "lying")
• Bias and representational distortion in training data
• Distribution shift and brittleness in deployment
• Why scale does not equal understanding
• Misleading anthropomorphism in public discourse
• Safety vs capability distinction

Key Concepts Introduced

• Confidence vs correctness separation
• Epistemic opacity
• Model uncertainty (informal framing)
• Robustness failure
• Alignment problem (conceptual introduction only)

Activity

Case analysis:

• AI-generated misinformation example
• Biased classification example
  Participants identify failure type and classify root cause.

Output

A "failure taxonomy" sheet created collaboratively.

MODULE 4 — AI Governance, Power, and Societal Integration

Objective

Situate AI systems within institutional, political, and governance structures.

Topics

• What "AI governance" actually covers (technical + institutional + regulatory layers)
• Key actors: labs, governments, standards bodies, civil society
• Regulation approaches (principles-based vs rules-based systems)
• Safety vs innovation trade-offs
• Compute, scaling, and concentration of power
• Introduction to alignment as a governance-relevant problem

Key Concepts Introduced

• Regulatory capture (AI context)
• Model evaluations (evals)
• Frontier models
• Dual-use systems
• Institutional incentives in AI development

Activity

Scenario simulation:
"Should a frontier model be released?"
Participants role-play different stakeholders (lab, regulator, civil society, enterprise).

Output

A structured policy memo outline (1 page).

FINAL CAPSTONE

"AI Literacy + Governance Brief"

Participants produce a short structured output plan or actual piece:

• Explain a chosen AI system
• Identify at least two failure modes
• Assess one governance implication
• Make a reasoned recommendation

I am willing to provide feedback at a later date too, amnakhan-22@outlook.com.

Pedagogical Design Principles

• Concept-first, jargon-second (terms introduced only after intuition is built)
• No mathematics dependency
• Systems thinking over tool usage
• Epistemic humility as a core skill (knowing uncertainty boundaries)
• Policy relevance embedded throughout, not tacked on at the end

Optional Extension Module (Advanced Cohort)

If you expand the workshop:

"Frontier AI and Alignment Thinking"

• Scaling laws (conceptual)
• Alignment problem deeper dive
• Interpretability vs control
• AI existential risk discourse (carefully balanced framing)
• Forecasting limits