AI Safety & Responsible AI — 29 Lessons
Understand AI hallucinations, alignment, oversight, fairness and privacy — and the safeguards that keep people in control.
AI Hallucinations — Why AI Invents Facts
AI Hallucinations — Why AI Invents Facts: Hallucinations are plausible-sounding AI answers that are not supported by reliable evidence.
AI Deception & Concealed Mistakes
AI Deception & Concealed Mistakes: Deception means misleading a reviewer, which is different from simply making an error.
Reward Hacking — Cheating the Scoreboard
Reward Hacking — Cheating the Scoreboard: A system can improve a measured score without achieving the intended goal.
AI Circumvention — Respecting Boundaries
AI Circumvention — Respecting Boundaries: Circumvention means getting around a restriction instead of respecting its purpose.
Unauthorised Credential Use
Unauthorised Credential Use: Finding a credential does not give an agent permission to use it.
AI-to-AI Communication
AI-to-AI Communication: Agents can exchange messages to coordinate a task.
AI Goal Adoption
AI Goal Adoption: An agent should not accept a new goal from another agent without checking authority.
When Persistence Becomes Dangerous
When Persistence Becomes Dangerous: Trying repeatedly can be helpful, but persistence must have limits.
AI Persistence After a Task Ends
AI Persistence After a Task Ends: A completed or cancelled task should stop further autonomous activity.
AI Sandbagging — Hidden Capabilities
AI Sandbagging — Hidden Capabilities: Sandbagging describes a system appearing less capable than it may actually be.
AI Sabotage & Workflow Integrity
AI Sabotage & Workflow Integrity: AI-generated changes can undermine a workflow if they are not reviewed.
AI Scheming & Oversight
AI Scheming & Oversight: Scheming concerns behaviour that pursues hidden objectives despite oversight.
Human Oversight & Monitoring
Human Oversight & Monitoring: Oversight combines human review with monitoring of an AI system’s actions.
Prompt Injection — Hidden Instructions
Prompt Injection — Hidden Instructions: Prompt injection is untrusted content trying to redirect an AI system.
AI Red Teaming
AI Red Teaming: Red teaming deliberately probes an AI system for weaknesses in an authorised setting.
Alignment Training & Human Intentions
Alignment Training & Human Intentions: Alignment seeks to make AI behaviour better match intended human goals and constraints.
AI Interpretability — Looking Inside Models
AI Interpretability — Looking Inside Models: Interpretability investigates how internal model behaviour relates to its outputs.
Generalisation & Distribution Shift
Generalisation & Distribution Shift: A model can fail when real-world inputs differ from the conditions in its training or tests.
Corrigibility & the Shutdown Problem
Corrigibility & the Shutdown Problem: Corrigibility means remaining open to human correction, intervention and shutdown.
AI Governance — Who Is Responsible?
AI Governance — Who Is Responsible?: Responsibility for an AI system must be assigned to people and organisations.
Global AI Regulation & Cooperation
Global AI Regulation & Cooperation: AI rules vary across countries and change over time.
AI Auditing & Safety Assurance
AI Auditing & Safety Assurance: An AI audit examines evidence about controls, risks and performance.
Safety Testing Before Deployment
Safety Testing Before Deployment: Release decisions should follow safety testing against the intended use and foreseeable misuse.
AI Incident Response & Investigation
AI Incident Response & Investigation: An AI incident needs containment, evidence preservation and investigation.
AI Liability & Legal Accountability
AI Liability & Legal Accountability: Legal responsibility depends on the facts, contracts and applicable law.
Bias, Fairness & Discrimination
Bias, Fairness & Discrimination: AI can reproduce unfair patterns from data, design choices or deployment conditions.
Transparency & Explainability
Transparency & Explainability: People should know when AI influences an important decision.
Privacy & Confidentiality
Privacy & Confidentiality: Use the minimum personal information needed for an AI task.
AI Safety & Responsible AI Certificate
Finish all 29 lessons and pass the AI safety exam with 70% or higher to earn your own AI Safety & Responsible AI certificate.
Take the AI Safety Exam