Risks and AI
How generative AI works, AI-related risks, the AI Act and standards, using AI well (prompting, verification, data, agents), integrating and operating it (evaluation, incidents, red teaming), safety-critical systems and use in risk management. Lessons, games, quizzes, graded exercises and debates. Course taught entirely in English.
- 8 chapters
- 59 lessons
- 22 exercises
- 9 quizzes
Two tracks (technical and no-programming), graded quizzes, exercises and interactive tools are available in the online course.
Read the first lesson → Create an account Log in to the course
At a glance
- Level
- Introductory
- Estimated time
- about 33 h
- Audience
- Students and professionals, with a technical track and a no-programming track.
- Recommended first
- none
- Learning outcomes
- Explain how generative AI works and its limits.
- Identify AI risks and the applicable texts (AI Act, frameworks).
- Use AI reliably: prompts, checking, data.
- Evaluate and monitor an AI system in operation.
- Assessment
- A graded quiz for each chapter; in a group, the next chapter opens once the lessons are read and the quiz passed (60% by default, threshold set by the trainer).
- Badge
- Digital badge (Open Badges format) issued by the trainer once every quiz of the course is passed and the final exam passed at 70%: 20 random questions, 30 minutes, at most 3 attempts, 24 h between two attempts.
- Updated
Course outline
Lessons marked “free to read” need no account; the other lessons, the exercises, the quizzes and the answers do.
Chapter 1What is AI?5 of 5 lessons free to read
Definitions, history, expert systems, machine learning, neural networks, drift and explainability.
- 1.1Definitions and historyfree to read
- 1.2Expert systemsfree to read
- 1.3Machine learningfree to read
- 1.4Neural networks and deep learningfree to read
- 1.5Model life cycle: drift and explainabilityfree to read
Chapter 2Generative artificial intelligence6 lessons
LLMs, inner workings, RAG, fine-tuning, reinforcement learning, agents and harnesses.
- 2.1What is generative AI?
- 2.2How an LLM works
- 2.3Customising an LLM: RAG and fine-tuning
- 2.4Reinforcement learning
- 2.5AI agents
- 2.6An agent’s harness
Chapter 3AI risks15 lessons
Risk categories, hallucinations and leaks, attacks, agent risks (MCP, OWASP Agentic, access), malicious uses, bias and societal risks, legal risks, loss of control, the human factor and companion chatbots, AI at work, authenticity of content, attacks on models, environment, from risk to scenario.
- 3.1Risk categories, hallucinations and leaks
- 3.2Attacks on machine learning algorithms
- 3.3Attacks on generative AI
- 3.4Risks linked to agents
- 3.5Too much access: whether or not to hand over the keys to your favourite AI
- 3.6AI as a tool for attackers
- 3.7Bias, rights and societal risks
- 3.8Legal risks: personal data and copyright
- 3.9Loss of control: when an AI works around its instructions
- 3.10Human factor: trust, attachment and companion chatbots
- 3.11AI at work: social dialogue and employee monitoring
- 3.12Real or generated? Provenance, watermarking and detectors
- 3.13Attacks on the model: training data, theft and backdoors
- 3.14Environmental risks
- 3.15Summary: from risk to scenario
Chapter 4Main texts and frameworks7 lessons
AI Act, GDPR, CNIL, ANSSI, ISO 42001, NIST, risk analysis sheet, product liability, NIS2, CRA, Council of Europe Convention and non-EU texts.
- 4.1Overview of the texts
- 4.2The AI Act and the Code of Practice
- 4.3CNIL and ANSSI
- 4.4ISO/IEC 42001 and NIST AI RMF
- 4.5Risk analysis sheet for an AI system
- 4.6Who pays in case of damage? Product liability
- 4.7Beyond the AI Act: cybersecurity, fundamental rights, non-EU texts
Chapter 5Using AI well7 lessons
Choosing a system, prompt and context engineering, checking what AI produces, data and confidentiality, using agents, learning with AI and staying competent.
- 5.1Choosing a generative AI system
- 5.2Anatomy of a prompt and techniques
- 5.3Context engineering: configuring your assistant
- 5.4Checking what AI produces
- 5.5Data and confidentiality: what you paste, what you do not paste
- 5.6Using agents
- 5.7Learning with AI and staying competent
Chapter 6Integrating and operating AI7 lessons
Evaluating a prompt or an assistant, evaluation harness, workflows and choice of tool, engineer’s code, integrating a model into a program, operating a system (monitoring, incidents), tooled red teaming.
- 6.1Evaluating a prompt or an assistant
- 6.2Building an evaluation harness
- 6.3Workflows and choice of tool
- 6.4AI and engineer’s code
- 6.5Integrating a model into a program
- 6.6Operating an AI system: oversight, monitoring and incidents
- 6.7Tooled red teaming: attacking your own system before others do
Chapter 7AI in critical systems4 lessons
Safety functions and learned models, the Machinery Regulation and standards, real cases (Uber, Cruise, Epic), architectures that keep AI out of the safety function.
- 7.1When AI enters a safety function
- 7.2The framework: Machinery Regulation, AI Act and standards
- 7.3Real cases: what the investigations show
- 7.4Designing: keeping AI out of the safety function
Chapter 8AI for risk management8 lessons
Uses, DUERP and HAZOP examples and pitfalls, predictive maintenance, anomalies and vision, studies and experience feedback, EasyRisk Pro and Polyodos tutorials, course conclusion.
- 8.1Possible uses
- 8.2Examples: DUERP, HAZOP and pitfalls
- 8.3Beyond generative AI: predictive maintenance, anomalies, vision
- 8.4Should you use it? Should you say so?
- 8.5Tutorial: producing a DUERP with AI in EasyRisk Pro
- 8.6Tutorial: analysing the risks of an industrial installation with AI in Polyodos
- 8.7Tutorial: having AI propose an EBIOS RM analysis in Polyodos (the optician case)
- 8.8Conclusion of the course