Speaker A
Hello and welcome to the AI Governance Online Course. You're now entering Module 1, where we lay the foundations of your understanding.
Introduction to AI systems, their definitions, and classification axes for effective governance and risk management.
Key Takeaways
- Clear, operational definitions of AI are critical for effective governance and regulation.
- Most existing AI systems are ANI, specialized for narrow tasks, not general intelligence.
- Misclassifying AI systems can lead to governance failures and unaddressed risks.
- Understanding AI system classification axes (scope, learning method, purpose, architecture) is essential for risk assessment and compliance.
- AI agents are architectures combining multiple ANI systems, not examples of AGI.
Summary
- The lesson introduces the foundational question: What exactly is an AI system?
- It presents the OECD definition of AI, highlighting four essential elements: machine-based, human-defined goals, output generation, and influence on environments.
- The importance of clear AI definitions in governance, regulation, and risk management is emphasized.
- Four shared characteristics of AI systems are identified: goal-orientation, ability to learn and adapt, autonomous decision-making, and impact on real or digital worlds.
- The course focuses on classifying AI systems to support governance, risk assessment, and compliance.
- The first classification axis distinguishes between Artificial Narrow Intelligence (ANI) and Artificial General Intelligence (AGI), with ANI being the prevalent type today.
- Clarification that AI agents are not AGI but composed of multiple ANI systems orchestrated together.
- The second axis introduces learning methods within ANI, differentiating Machine Learning (ML) and Deep Learning (DL).
- The lesson includes practical tools like the AI System Classification & Risk-Scoping Canvas for hands-on classification and risk analysis.
- A bonus quiz and encouragement to subscribe for further lessons on AI risks and threats are included.
Full Transcript — Download SRT & Markdown
Speaker A
In this lesson, we'll start with a deceptively simple question: What exactly is an AI system? But we won't settle for surface-level answers. We'll break AI down into parts you can actually analyze and manage. Because in governance, clarity is power.
Speaker A
In this session, you'll discover the true nature of AI, beyond hype and buzzwords. You'll learn the four essential axes we use to classify any AI system: its scope, learning method, purpose, and architecture. Each of these axes affects how we assess risks, determine legal
Speaker A
obligations, and choose control strategies. Most importantly, you'll see how precise classification directly supports real-world governance. Push pin if you're ready to apply what you learn.
Speaker A
There's a hands-on practical lesson available for this lesson, where we'll take a real AI system and deconstruct it step-by-step. You'll classify each of its components, analyze the risks, and explore real governance implications. To guide this process, you'll use a purpose-built tool, the
Speaker A
AI System Classification & Risk-Scoping Canvas. A simple but powerful framework you'll return to throughout the course. Let's get started.
Speaker A
And one more thing, at the end of this lesson, you'll find a quick bonus quiz to test what you've learned. So make sure to stay with us until the very end. And if you don't want to miss the next lesson on AI risks and threats, don't forget to subscribe to the channel.
Speaker A
Let's begin this lesson with a simple but fundamental question. What exactly qualifies a system as "Artificial Intelligence"? In the real world, especially in legal and policy contexts, definitions matter. That's why we turn to one of the most influential definitions
Speaker A
provided by the OECD, the Organisation for Economic Co-operation and Development. The OECD is an international organization that brings together developed nations to coordinate policy, promote economic growth, and establish global standards, including in the field of AI.
Speaker A
According to the OECD, AI is defined as... "...a machine-based system that, for a given set of human-defined objectives, can make predictions, recommendations, or decisions influencing real or virtual environments." This may sound abstract at first, but it gives us
Speaker A
a solid operational definition that regulators and institutions around the world now rely on. What makes this definition powerful is its clarity. It identifies four essential elements of an AI system. It must be machine-based. It must pursue human-defined goals. It must generate outputs,
Speaker A
such as predictions or decisions. And finally, it must influence real or virtual environments. These are the building blocks that will allow us to determine whether a system is truly AI, or something else entirely. Other organizations and countries offer
Speaker A
their own definitions, and we've compiled the most prominent ones for you to explore, highly recommended for your understanding. Despite differences in wording, most definitions converge on a few core characteristics that define an AI system. Understanding these is critical.
Speaker A
Why? Because in the sea of digital tools, we must clearly distinguish what qualifies as AI and apply the appropriate policies and legal frameworks. This answers the essential question: "What is it?" So, why should we care about definitions in the first place? Because in the context of governance,
Speaker A
regulation, and risk management, the first step is always to ask: "Is this system actually AI?" Only once we've answered that question can we apply the right legal frameworks, risk controls, and ethical guidelines. And here's the interesting part. While there
Speaker A
are many definitions of AI from different countries and institutions, they all tend to share a common DNA. Across the board, four shared elements appear again and again. First, AI systems are always goal-oriented. Second, they have the ability to learn and adapt. Third,
Speaker A
they make autonomous decisions, without constant human input. And fourth, their actions affect the real or digital world in some way. These shared elements give us a practical filter. If a system doesn't exhibit these traits, it might not be AI in the governance
Speaker A
sense. And this leads us to the next critical question we'll explore: "What kind of AI system is it?" That's where classification begins—and where governance truly starts.
Speaker A
Why does this classification matter? For an AI governance professional, this is not just academic theory—it's the foundation for everything that follows: risk assessment, compliance obligations, and choosing the right control tools. Misclassifying a system at the
Speaker A
start leads to governance failures and unaddressed risks. Today, we'll walk through four foundational classification axes. For each axis, we'll define clear criteria, look at real-world examples, and, most importantly, explain how each dimension impacts your day-to-day work.
Speaker A
Axis 1: Scope of Intelligence — ANI vs. AGI. The first and most fundamental axis distinguishes between two very different types of AI: Artificial Narrow Intelligence (ANI).
Speaker A
These systems are designed for one specific task or a narrow range of tasks. Think chess engines, facial recognition, or self-driving cars on predefined routes. Keywords: narrow, limited context. 99.99% of existing AI systems today—from your spam filter to industrial robots—fall into
Speaker A
this category. Artificial General Intelligence (AGI), a theoretical concept for now, would have cognitive capabilities comparable to humans, able to learn, reason, and apply knowledge to any intellectual task, including unfamiliar ones. Litmus test: Ask one question. Can this system
Speaker A
perform a task it wasn't trained for? Like a chess bot suddenly composing poetry? If not, it's ANI.
Speaker A
Examples: ANI: Spam filter, Netflix recommender, AlphaFold. AGI: HAL 9000, Skynet, Samantha (from Her)—fictional for now. Implications for Governance: ANI: Focus on tactical risks, model accuracy, data bias, reliability, and user data security. AGI: Strategic, long-term existential
Speaker A
risks, loss of control, misaligned goals, and societal disruption. For most AI governance certifications, understanding AGI is a required part of the theoretical foundation. However, in this course, our primary focus—both in theory and in practice—will be on working with ANI systems.
Speaker A
Let's clarify a common misconception. AI agents are NOT AGI. Even the most advanced "agents" today, like autonomous drones or virtual assistants, are actually built from multiple narrow AI systems, each solving one specific task. For example, a smart assistant may combine an ANI for speech recognition, another ANI for understanding intent,
Speaker A
a third ANI to generate responses, and a few rule-based modules to manage dialogue flow.
Speaker A
These are then orchestrated by a control system that gives the impression of an autonomous, thinking agent. But at the core, there's no general reasoning, no transfer of knowledge between domains, no self-awareness. Understanding this is essential. If you confuse an agent with
Speaker A
AGI, you'll misjudge its risks, capabilities, and governance needs. Always remember: "Agent" is a system architecture, not a level of intelligence. Axis 2: Learning Method. Let's dive deeper.
Speaker A
Within the world of ANI, systems differ in how they learn. The two key terms here are Machine Learning and Deep Learning. Machine Learning (ML) is a broad field where algorithms learn from data without being explicitly programmed. Imagine you are building a model to assess credit risk.
Speaker A
As a human expert, you instruct the machine beforehand: "Focus on these features: age, income, credit history, length of employment." This process of manually selecting and preparing features is called "Feature Engineering." It is the hallmark of classical ML.
Speaker A
Deep Learning (DL) is a subset of ML. It is based on multi-layered artificial neural networks. And here I
Speaker A
layer by layer, extracts a hierarchy of features, from simple lines and gradients to complex objects like cats or cars. Your key differentiator is this: if the system learns on its own from "raw" material, it's most likely DL. Examples. ML: That same credit risk system that
Speaker A
uses features like age and income. DL: A system that identifies cancer cells in medical images, which learns on its own to see the textures and patterns indicative of the disease, without any guidance from a human. What does this mean for AI governance?
Speaker A
The implications here are massive. Resource Intensity. DL systems are voracious. They require vast datasets and immense computational power. This directly impacts a project's budget, its carbon footprint, and its IT infrastructure requirements. The "Black Box" Problem. This is perhaps the most significant challenge. Due to millions of parameters and a complex,
Speaker A
multi-layered structure, the decisions made by DL systems are extremely difficult to interpret. Why was the loan application denied? There is no precise answer. This creates a colossal challenge for explainability and for compliance with regulations, such as the "right to explanation" under GDPR. Amplified Data Risks. DL models are like "sponges" for data. They
Speaker A
can learn and then reproduce not only useful patterns, but also hidden, undesirable ones, including biases and even confidential information from the training dataset.
Speaker A
Beyond classic machine learning and deep learning, there are several other ways AI systems can "learn" or acquire behavior. In this segment, we'll cover three of the most common approaches, and then I'll give you a preview of more advanced methods we'll study later in the course.
Speaker A
Rule-based / Expert Systems. These are the original AI methods. You encode explicit "if-then" rules based on human expertise. For example, an expert system for medical diagnosis follows a decision tree like: "If symptom A and symptom B, then suggest test X." There is
Speaker A
no statistical learning or data-driven model here, just transparent, handcrafted logic. Reinforcement Learning (RL). Here, an AI "agent" learns by trial and error, guided by rewards and penalties. Imagine training a robot to navigate a maze: each correct move earns a point, each collision costs points. Over many episodes, the agent maximizes its cumulative
Speaker A
reward and discovers optimal strategies. RL is widely used for robotics, game-playing agents like AlphaGo, and dynamic resource allocation. Unsupervised Learning. In unsupervised learning, the system finds patterns in unlabeled data. Common techniques include clustering (grouping customers by purchasing behavior) and dimensionality reduction,
Speaker A
which compresses large feature sets into simpler representations. Because there are no predefined "answers," unsupervised methods excel at exploratory analysis and anomaly detection.
Speaker A
There are also semi-supervised, self-supervised, transfer learning, fuzzy logic, and evolutionary algorithms, each with its own strengths and governance considerations. We'll dive into these advanced methods in later modules of this course. Axis 3: Purpose — Analytical vs. Generative. The
Speaker A
third axis is the most relevant one today. It categorizes AI based on its primary purpose: does it analyze existing information, or does it create something new? Classical, or Analytical AI, is used to analyze existing data. It answers questions like: "What is this?",
Speaker A
"Which category does this belong to?", or "What will happen next?" The output of its work is a conclusion, a label, or a prediction. Generative AI (GenAI), in contrast, is used to create new, original content that resembles the data it was trained on. It responds to the prompt, "Create for
Speaker A
me..." The output of its work is a new artifact: a piece of text, an image, code, or music.
Speaker A
The criterion for identification here is the most straightforward: is the system's output an insight about existing data, or is it the creation of a new, previously non-existent artifact?
Speaker A
Examples. Classical AI: a fraud detection system at a bank, a customer churn prediction model. GenAI: ChatGPT, Midjourney, DALL-E, GitHub Copilot.
Speaker A
And again, the key question: What does this mean for AI governance? GenAI is not just the next step. It introduces entire new vectors of risk that were not a concern for classical AI. Intellectual Property. There are risks of copyright infringement both during training
Speaker A
(if protected data was used) and during generation (if the output is a derivative work). Disinformation and Fakes. The ability to mass-produce plausible, yet completely false, content. This includes deepfakes, fake news, and propaganda. Hallucinations. This is a risk specific to GenAI. The model can generate information that is factually incorrect but sounds
Speaker A
highly convincing. Security. The use of GenAI to create malicious code, convincing phishing emails, and to facilitate social engineering. The takeaway for you is this: standard data and model governance policies are no longer sufficient. You will need specialized policies for
Speaker A
acceptable use, fact-checking, the implementation of watermarks, and prompt governance. Axis 4: Architecture — Symbolic vs. Connectionist. And for our final axis today, the fourth one. It looks under the system's hood and answers the question: "On what principles is
Speaker A
its logic constructed?" Symbolic AI, also known as Good Old-Fashioned AI or GOFAI, is based on manipulating symbols according to explicit, predefined rules. This is the kind of AI that we can "read." Its operational logic is fully transparent. Connectionist AI,
Speaker A
on the other hand, is based on training artificial neural networks on large datasets. We've already touched upon this when discussing ML and DL. Here, knowledge is represented not in the form of rules, but as the numerical weights of the connections between neurons. Its logic is emergent,
Speaker A
meaning it arises on its own during the training process and is frequently opaque. Your key question for differentiation is: "Can the entire decision-making process be traced through a chain of explicit, human-readable rules?" If yes, it's a symbolic approach. If the decision is the
Speaker A
result of a complex mathematical transformation inside a neural network, it's connectionist. Examples. Symbolic: classic expert systems for diagnostics, rule-based grammar checkers.
Speaker A
Connectionist: almost all modern systems for pattern recognition, translation, and text generation. It's also important to mention the hybrid approach, which combines both worlds. For instance, a neural network (connectionism) might recognize objects in a photo, while a symbolic engine reasons about
Speaker A
their interactions based on a set of rules. What does this mean for AI governance?
Speaker A
The primary distinction lies in transparency and explainability (XAI). Symbolic systems are "white boxes." Their decisions are inherently explainable. Connectionist systems are "black boxes." They require specialized, complex post-hoc analysis methods, like LIME or SHAP, to obtain even a partial explanation. And implementing these methods is a distinct and
Speaker A
critical task for governance. Furthermore, they differ in their reliability. Symbolic systems are predictable within the confines of their rules, but very brittle beyond them. Connectionist systems are more flexible but can sometimes yield completely unpredictable and illogical results.
Speaker A
In this lesson, you've learned how to define and classify AI systems, including the four key axes used in global frameworks. And that's more than enough to pass most certification exams. But before we wrap up, there's a bonus mini-quiz to help you check your understanding. Stick around,
Speaker A
the questions are coming up in just a moment. And if you're the kind of learner who prefers doing over just watching, we've prepared an optional hands-on practice lesson. It's designed for those who want to apply AI governance principles from day one. In the practice session, you'll use a
Speaker A
real governance tool, the Classification and Risk Canvas. Break down an actual AI system into components. Classify each one across all four axes. Analyze and document your results, just like in a real audit. You'll be working with Microsoft 365 Copilot as your case study. If that
Speaker A
sounds helpful, the link is in the description. Let's wrap up with a quick knowledge check.
Speaker A
I'll read each question aloud. Take a moment to think through your answers. What are the four axes of AI system classification?
Speaker A
Which of the following is NOT one of the OECD criteria for identifying an AI system?
Speaker A
Which classification would best fit a voice assistant like Siri? What does "Deep Learning" primarily rely on? What is the main purpose of a generative AI system? How would you classify Google Translate (the standard web version)? Now's your turn. Pause the video and write
Speaker A
your answers in the comments. It's a great way to share your thoughts with others and test your understanding of the key concepts from today's lesson.
Speaker A
Great job making it to the end of this lesson! If you found it valuable, don't forget to give it a like and subscribe to the channel so you don't miss the next session. In our next lesson, we'll dive into a critical topic: AI risks and threats—exploring how AI systems can cause
Speaker A
harm to individuals, organizations, and entire societies. And if you know someone who's working with AI or thinking about it, send them this lesson. It might be exactly what they need to understand how to govern AI systems responsibly. Thanks for watching, and see you in the next one.
Topics:AI governanceArtificial IntelligenceAI classificationArtificial Narrow IntelligenceArtificial General IntelligenceMachine LearningDeep LearningOECD AI definitionAI risk managementAI policy
