AI and ML Fundamentals

Study guide for Cloud Digital Leader, Unit 3 · Topic 1. This is the topic's lecture in reading form — every slide's teaching, figures and worked examples, in order — followed by the official Google Cloud pages its claims rest on.

What the exam guide asks. Discuss the main AI and ML concepts, and explain how ML can create business value.

Objectives, quoted from the exam guide:

1. Define artificial intelligence (AI) and machine learning (ML).
2. Differentiate the capabilities of AI and ML from data analytics and business intelligence.
3. Discuss the types of problems that ML can solve.
4. Explain the business value ML creates, including: ability to work with large datasets; scaling business decisions; and unlocking unstructured data.
5. Explain why high-quality, accurate data is essential for successful ML models.
6. Discuss the importance of explainable and responsible AI

AI and ML fundamentals

What the words mean, what the models need, and what they owe

What it is — artificial intelligence, machine learning, and how both differ from analytics. What it is for — the problems machine learning solves and the value it creates. What it needs and owes — high-quality data going in, and explanations and responsibility coming out.

Unit three moves from data to what can be learned from it, and this first topic sets the vocabulary. The guide's summary is to discuss the main artificial intelligence and machine learning concepts, and explain how machine learning can create business value. Its six objectives fall into three questions. What is it: definitions of artificial intelligence, or AI, and machine learning, or ML, and how they differ from data analytics and business intelligence. What is it for: the kinds of problems machine learning solves and the business value it creates. And what does it need and owe: a model is only as good as the data it learns from, and a model people depend on has to be explainable and used responsibly. Keep those three questions in view and the six objectives stop being a list.

Artificial intelligence and machine learning, defined

AI is the goal; ML is one way of reaching it

• AI: machines performing tasks that typically require human intelligence
• ML: a subset of AI that learns and improves from experience
• ML replaces explicit programming with algorithms that learn from data
• Deep learning: a subset of ML that trains neural networks on data

Worked example (synthetic). A bank's rules engine flags any transfer over a fixed amount. It automates a decision, but nobody would call it machine learning: it was programmed, not trained, and it never gets better on its own.

Start with the two definitions the exam tests. Google describes artificial intelligence, or AI, as a field of computer science focused on creating smart machines that can perform tasks that typically require human intelligence, like learning, reasoning and problem-solving. Machine learning, or ML, is narrower. Google calls it a subset of artificial intelligence that automatically enables a machine or system to learn and improve from experience, and says what replaces the programmer: instead of explicit programming, machine learning uses algorithms to analyze large amounts of data, learn from the insights, and then make informed decisions. That phrase — instead of explicit programming — is the line between a machine learning system and ordinary automation. A rule someone wrote is not learning, however clever it is. There is one more level of nesting. Just as machine learning is a subset of artificial intelligence, deep learning is a subset of machine learning, and it works by training neural networks on sets of data; Google says models with more than three layers are what get called deep neural networks, or deep learning. Finally, notice the direction of the difference. Google says artificial intelligence encompasses the idea of a machine that can mimic human intelligence, and machine learning does not — machine learning aims to teach a machine to perform a specific task and provide accurate results by identifying patterns.

One umbrella, several fields under it

Machine learning is inside AI, and deep learning is inside ML

Figure. Nested boxes. The outer box is artificial intelligence. Inside it, a box for machine learning, which itself contains a box for deep learning, and beside it a box listing other fields under the AI umbrella: robotics, expert systems and natural language processing.

Worked example (synthetic). An item asks whether a chatbot built on deep learning is an AI system. It is — deep learning sits inside machine learning, which sits inside AI — so every one of the three labels applies.

Draw the definitions as boxes and the most common confusion disappears, because the terms are nested rather than parallel. The outer box is artificial intelligence, which Google describes as a broad field using technologies to build machines that can mimic cognitive functions associated with human intelligence — seeing, understanding and responding to spoken or written language, analyzing data, making recommendations. Inside it sits machine learning. Google uses exactly this picture: artificial intelligence is an umbrella, and machine learning sits under that umbrella — but so do other major subfields, such as deep learning, robotics, expert systems and natural language processing. Deep learning sits inside machine learning in turn. So an item that asks whether something is AI or ML is often asking a question with two right labels; the discriminating question is whether the system learns from data, which makes it machine learning, or reaches its behaviour some other way, such as rules an expert wrote down.

AI and ML against analytics and business intelligence

Different questions, and a different way of answering them

Worked example (synthetic). A retailer's dashboard shows last quarter's returns by region — business intelligence. A model that scores each new order for its chance of being returned is machine learning. Same data, a different question.

The second objective asks you to tell artificial intelligence and machine learning apart from data analytics and business intelligence, and the cleanest way is by the question each one answers. Google defines business intelligence as using the power of people and technologies to collect and analyze data for an organization's strategic and daily decisions, and says a modern tool can help organizations understand what happened, why it happened, and what needs to happen to optimize operations. That is largely a view of the past, interpreted by people. Machine learning answers a different question. Google's predictive analytics page describes a branch of data science that uses historical data, statistical modeling, and machine learning to answer the question: what might happen next? The mechanism differs too. An analyst asks the data a question; a machine learning system learns patterns from it — Google's phrase is teaching a computer by showing it a million examples instead of writing a million rules. The boundary is not a wall: Google says organizations use business intelligence to move from descriptive to predictive insights. And the reason to cross it is scale — the sheer volume coupled with complexity often makes data difficult to analyze using traditional tools.

From describing the past to predicting what comes next

The questions get harder, and the method changes

Figure: A flow of four questions. The first two, what happened and why did it happen, are grouped as business intelligence, where people analyze data. The last two, what might happen next and act on the prediction, are grouped as machine learning, where a model learns patterns.

Worked example (synthetic). A logistics firm's report shows deliveries ran late in March because of storms. Knowing that is business intelligence; a model that forecasts next week's late deliveries from weather data is the step past it.

This figure puts the two sides of the second objective on one line. On the left, business intelligence: understanding what happened and why it happened — Google's own description of what a modern business intelligence tool does. On the right, the question Google gives predictive analytics: what might happen next, answered using historical data, statistical modeling and machine learning. Google treats the move between the halves as a progression organizations make on purpose — they use business intelligence to move from descriptive to predictive insights. What changes as you cross the line is who finds the pattern. On the left, a person asks the data a question and reads the answer. On the right, a model learns the pattern from examples and applies it to cases nobody has seen yet. An exam item that describes a dashboard, a report or an explanation of last quarter is on the left of this line; an item that describes a forecast, a score or a recommendation for something that has not happened yet is on the right.

The kinds of problems machine learning solves

Match the problem to what the training data looks like

Worked example (synthetic). A retailer has years of purchases but no customer segments defined. There are no labels to learn from, so the problem is unsupervised: let the model find the groups.

The third objective asks which types of problems machine learning can solve, and Google answers it two ways — by business function and by learning approach. By function, Google lists fraud detection, identifying security threats, personalization and recommendations, automated customer service through chatbots, transcription and translation, and data analysis. By approach, the table sorts problems by what the training data looks like, which is usually the fastest way to classify an exam scenario. Supervised learning uses labeled training data to map a specific feature to a label — the right answer is known for every example, as when pictures are labeled apples. Unsupervised learning uses unlabeled data to learn patterns, and Google says it is good at descriptive modeling and pattern matching. Semi-supervised learning is the mixed case, in which only some data is labeled. Reinforcement learning can be described as learn by doing, through a series of trial and error experiments. So the question to ask of any scenario is: does the business already know the right answers for its past examples? If it does, the problem is supervised; if it wants groups nobody has defined, it is unsupervised.

Business problems machine learning is used for

Each one is a pattern too large or too fast for people to find

Figure. Six cards naming business problems machine learning is used for: fraud detection, maintenance forecasting, personalization, security threats, customer service chatbots, and transcription and translation.

Worked example (synthetic). A factory's machines fail without warning twice a quarter. The machine learning problem is maintenance forecasting — predicting the failure before it happens — not a better report on the last failure.

These six cards are the business form of the third objective, and every one comes from Google's own lists. Fraud detection is described as examining network actions in real time to pinpoint abnormalities. Maintenance forecasting is predicting equipment failure before it occurs to reduce downtime. The other four are the functions Google's machine learning page names directly: personalization and recommendations, identifying security threats, automated customer service through chatbots, and transcription and translation. What they share is the reason machine learning fits: in each, the useful pattern is buried in more examples, arriving faster, than people could review. Notice too that each is one specific task. Google describes today's artificial narrow intelligence models the same way — designed to perform a single, specific task, such as identifying images, engaging in chat, or filtering emails. An exam scenario that asks machine learning to do everything at once is usually the distractor.

The business value machine learning creates

Large datasets, decisions at scale, and data nobody could read

• Large datasets: find value in more data than ever before
• Scaling decisions: insight quickly as data grows, not as headcount grows
• Fewer errors: the same process followed every single time
• Unstructured data: text and images turned into business-ready data

Worked example (synthetic). An insurer receives thousands of scanned claim forms a day. Optical character recognition turns them into structured records, and a model scores each for review — work that previously scaled only by hiring.

The fourth objective names three kinds of business value, and each has a Google sentence underneath it. The first is working with large datasets. Google says the rate of data generation is accelerating, creating more data than ever before, and machine learning helps make it possible to analyze and find value in this vast amount of data. The second is scaling business decisions. The problem Google names is that building, testing, iterating and deploying analytical models by hand eats up employees' time in a way that scales poorly; the answer is that machine learning can enable an organization to derive insights quickly as data scales. Google adds the quality side of the same benefit: artificial intelligence, or AI, can minimize manual errors through automation and algorithms that follow the same processes every single time, for more informed, faster decision-making. The third is unlocking unstructured data. Here the guide's phrasing is broader than any single Google sentence, so be precise. Google credits AI and ML together with analyzing and activating a wider range of unstructured and structured data sources, and gives optical character recognition, or OCR, as the example — pulling text and data from images and documents, which transforms unstructured content into structured, business-ready data. And the value compounds: machine learning systems continuously adjust and enhance themselves as they accrue more experiences.

Each value, and the mechanism under it

A benefit the exam names always has a reason Google gives

Worked example (synthetic). Offered "machine learning is cheaper than analysts" as the benefit of scaling decisions, prefer the answer that names insight growing with the data rather than with headcount — that is the mechanism Google gives.

Laid out as a table, the fourth objective becomes three problem-and-answer pairs. Large datasets: the sheer volume coupled with complexity often makes data difficult to analyze using traditional tools, and machine learning helps find value in that vast amount of data. Scaling decisions: building and deploying analytical models by hand eats up employees' time in a way that scales poorly, and machine learning derives insights quickly as data scales, with algorithms that follow the same processes every single time. Unstructured data: scanned forms, images and documents cannot be queried as they are, and artificial intelligence, or AI, working with machine learning — optical character recognition is Google's example — transforms unstructured content into structured, business-ready data. The middle column is the one exam items test. A benefit stated with no problem behind it, or with the wrong one, is usually the distractor.

Why high-quality, accurate data decides the model

A model learns whatever the data teaches it — including its mistakes

• More samples make a model more accurate — if the data is high quality
• A biased dataset produces biased results
• Poor data is critical in loans, hiring and policing
• Quality is judged on six dimensions, and governance protects it

Worked example (synthetic). A lender trains an approval model on ten years of past decisions. If those decisions were skewed against one group, the model learns the skew and applies it faster and at scale.

The fifth objective asks why high-quality, accurate data is essential, and Google's answer has a condition built into it. Assuming the training data is of high quality, the more training samples the machine learning algorithm receives, the more accurate the model will become. The whole sentence depends on its first clause. Google's artificial intelligence page says the same from the other side: training data serves as the material the model learns from, and its quality and quantity are crucial for the AI's performance. When the data is wrong, the model is wrong in the same way. Google puts it bluntly — machine learning is often only as good as the data it is being fed, and if an algorithm is fed a biased dataset, it will deliver biased results. It also names where that matters most: poor training data can lead to biased or inaccurate outputs, which can be critical in applications like loan approvals, hiring decisions and predictive policing. So quality has to be managed, not hoped for. Google's governance page judges data quality on six dimensions — accuracy, completeness, consistency, timeliness, validity and uniqueness — and says modern data governance ensures data scientists and engineers can access high-quality data to build accurate models. Even gathering it is work: acquiring and cleaning data to train algorithms can be difficult if the data sits in siloed locations across the organization.

The training loop, and where bad data enters it

The loop corrects the model toward the data — not toward the truth

Figure: A training loop. Training data feeds fitting the model; a loss function measures the model's errors; an optimization step adjusts parameters; the model is re-trained until its output fits the expected outcome, producing a trained model. A dashed arrow shows biased or inaccurate data entering the training data, which the loop learns as well.

Worked example (synthetic). A model trained on sensor readings from a miscalibrated device converges perfectly — to the miscalibration. Its loss is low because it matches the data, not reality.

This figure shows why data quality cannot be fixed later by training harder. Google describes fitting a model as a loop: the process uses a loss function to measure the model's errors, and an optimization technique, like gradient descent, to adjust the model's parameters and minimize those errors — and if the outcome does not fit the expected outcome, the algorithm is re-trained again and again until it outputs the accurate response. Look at what the loop measures errors against. It is the training data. The loop drives the model toward agreement with its examples, so when those examples are biased or inaccurate, the loop faithfully learns the bias and the inaccuracy, and reports success. That is the mechanical meaning of Google's warning that a machine learning algorithm fed a biased dataset delivers biased results, and why Google says having big data does not guarantee results unless the data is accurate, relevant and properly organized for analysis.

Explainable and responsible AI

People must be able to see why a model decided — and trust how it is used

• Models are often black boxes, even to their designers
• Explanations show which features drove an inference
• Responsible AI: fairness, interpretability, privacy and security
• Risks depend on the use case, users and business context

Worked example (synthetic). A lender's model declines an applicant. Without an explanation, nobody can say whether income drove the decision or a proxy for a protected characteristic did — and nobody can fix it.

The last objective asks why explainable and responsible artificial intelligence, or AI, matter, and the two ideas answer different worries. Explainability is about seeing inside. Google says machine learning models are often black boxes: even their designers cannot explain how or why a model produced a specific inference. Knowing how a model behaves, and how its training dataset influences it, gives anyone who builds or uses machine learning the ability to improve models, build confidence in their inferences, and understand when and why things go awry. One concrete technique is feature attribution, which indicates how much each feature in a model contributed to each inference — and can help detect issues in the data that standard model evaluation techniques would usually miss. One product note: Google's page for Vertex Explainable AI says it is deprecated as of March sixteenth, twenty twenty-six, so learn the idea, not the product name. Responsible AI is about consequences. Google says the evolving capabilities of artificial intelligence create potential for misapplication, misuse, and unintended or unforeseen consequences; its generative AI services are designed with Google's AI Principles in mind; and customers are encouraged to promote fairness, interpretability, privacy and security practices, and to understand and test their models to deploy safely and responsibly. The two meet at bias: if the data reflects human biases, the AI will learn and perpetuate them, and explanations are how you find out.

Explainable against responsible

One asks why the model decided; the other asks whether it should

Worked example (synthetic). A hiring tool passes every accuracy test. Explainability reveals it leans on postcode; responsible-AI practice asks whether postcode should be allowed at all. Both questions are needed.

Side by side, the two halves of the last objective — explainable and responsible artificial intelligence, or AI — are complementary rather than overlapping. Explainable AI starts from a black box nobody can explain, offers feature attributions that show which inputs drove each inference, and catches issues in the data that standard model evaluation techniques would usually miss. Responsible AI starts from the potential for misapplication, misuse and unintended consequences, offers practices for fairness, interpretability, privacy and security, and watches for what Google calls bias amplification — models inadvertently amplifying existing biases in their training data into outputs that reinforce unequal treatment of certain groups. The last row is the one exam items like to test. Google does not treat responsibility as something the platform settles alone: beyond built-in technical safeguards, it says it is important to consider other risks specific to your use case, users and business context. A scenario that assumes the provider's safeguards discharge the customer's responsibility is the distractor.

What this topic actually tests

Four discriminations, six objectives

Learned or programmed? only learning from data is ML. Past or next? BI explains what happened; ML predicts what might. Labeled or not? it picks supervised or unsupervised. Good data in, and an explanation out? without both, accuracy is not trust.

Close the topic on the four discriminations the exam leans on. First, learned or programmed: machine learning is the system that learns from data instead of explicit programming, and a rule someone wrote is automation, not machine learning — while both sit under the artificial intelligence umbrella. Second, past or next: business intelligence explains what happened and why, and machine learning answers what might happen next. Third, labeled or not: labeled examples make a problem supervised, unlabeled data looking for groups makes it unsupervised, and trial and error makes it reinforcement learning. Fourth, what goes in and what comes out: a model is only as good as the data it is fed, and a model people rely on must be explainable and used responsibly, because a biased dataset delivers biased results. Topic two turns from concepts to Google Cloud's own AI and ML products, and every one of these words comes with it.

Official sources for this topic

• Cloud Digital Leader exam guide — Section 3
• What is artificial intelligence? — Google Cloud
• AI vs. machine learning — Google Cloud
• What is machine learning? — Google Cloud
• What is business intelligence? — Google Cloud
• What is predictive analytics? — Google Cloud
• What is data governance? — Google Cloud
• Introduction to Vertex Explainable AI
• Responsible AI — Gemini Enterprise Agent Platform
• What is big data? — Google Cloud