All Levels Free 6 modules ~4 hours

AI Safety and Alignment

AI systems are getting more capable every year. Whether that capability is used well depends on safety and alignment, two fields that determine whether an AI does what we actually want, rather than what we literally asked for. This course explains both fields from the ground up, with no math prerequisites.

Start Module 1 View all modules
Your Progress
0 of 7 complete
Is this for me?
This IS for you if...
  • You build or deploy AI systems and want to understand safety risks
  • You are in a leadership role making decisions about AI adoption
  • You are new to AI and want to understand why safety matters
  • You need to explain AI safety to a non-technical audience
  • You want to understand the EU AI Act or NIST AI RMF in plain language
This is NOT for you if...
  • You are looking for deep ML theory on safety (this course is conceptual first)
  • You need a full implementation guide for RLHF (see the referenced papers)
  • You want a legal compliance checklist without context

Prerequisites

Before this course

Comfortable with what AI models do at a high level. Modules 01 to 07 from AJ University (AI Basics through Evaluating AI) are recommended but not required. No code background is needed to follow the concepts. The code examples use Python and are explained line by line.

What you will learn

Six concepts that change how you think about AI

01 Explain AI safety in plain language: the difference between a capable AI and a safe one
02 Describe the alignment problem, Goodhart's Law, and why reward functions are hard to get right
03 Explain how RLHF and Constitutional AI train models to be helpful and avoid harm
04 Identify the main types of jailbreaks and prompt injection attacks, and the defenses used against them
05 Navigate the EU AI Act risk tiers and the NIST AI Risk Management Framework in practice
06 Apply a builder safety checklist: red-teaming, content filters, system prompt hardening, and incident response
Course Modules

Six modules, one capstone

01
What Is AI Safety?
Definition, why it matters now, the difference between capability and alignment, and real examples of AI systems behaving unexpectedly. Includes a capability-vs-alignment interactive and a 2x2 risk matrix SVG.
35 min
› 02
The Alignment Problem
Goodhart's Law in AI, specification gaming, reward hacking, mesa-optimization, and instrumental convergence. Canvas animation showing how a reward function gets gamed.
40 min
› 03
RLHF and Constitutional AI
How reinforcement learning from human feedback works step by step. Reward model training, PPO, Constitutional AI from Anthropic, RLAIF, and the tradeoffs between methods.
45 min
› 04
Jailbreaks and Adversarial Inputs
What jailbreaks are, prompt injection vs jailbreaking, types of attacks (direct, indirect, many-shot, persona), defenses, and why perfect defense is hard. Attack taxonomy SVG included.
40 min
› 05
AI Governance and Policy
EU AI Act risk tiers, NIST AI RMF, ISO 42001, and the US Executive Order on AI. What each framework requires in practice, with a risk-tier classifier interactive.
40 min
› 06
Practical Safety for Builders
Red-teaming, content filters, system prompt hardening, rate limiting, monitoring, and incident response. Includes a deployment safety score canvas and a Python safety wrapper class.
40 min
› ◆
Capstone: Build a Safety Review
Apply everything by writing a full safety review for a real AI system. Radar chart covering six safety dimensions. Certificate of completion included.
45 min
›
Related Courses

Continue learning

Course
Evaluating AI Models
Course
AI Agents
Course
Enterprise AI
Supporting Articles

Read the blog series

Part 1
What Is AI Safety?
Part 2
The AI Alignment Problem, Explained
Part 3
AI Safety for Beginners
Part 4
RLHF: Reinforcement Learning from Human Feedback
Part 5
AI Jailbreaks and Prompt Injection
Part 6
AI Safety in the Enterprise