AI & ML Fundamentals

Every ML system is an AI system, but not every AI system learns from data, and that one nested relationship is the mental model the whole domain rests on. Get the AI-ML-deep-learning hierarchy straight and you can read a dashboard, name the problem shape a business question reduces to, and tell when ML earns its keep over plain analytics. Choosing a Google Cloud AI service, or building one, builds on the definitions below.

Artificial intelligence (AI) is the broad field of computer science concerned with building systems that reason, learn, and act in ways that used to require human intelligence. Google describes AI as "a set of technologies that empowers computers to learn, reason, and perform a variety of advanced tasks"^[1] such as understanding language, analysing data, and making suggestions. AI is the umbrella term.

Machine learning (ML) is the application of AI in which a system learns the rules from data instead of having them hand-coded. Google's own definition is precise: "Machine learning is a subset of artificial intelligence that enables a system to autonomously learn and improve ... without being explicitly programmed, by feeding it large amounts of data."^[2] The single most useful mental model for this whole domain is the subset relationship, drawn below as concentric boxes: every ML system is an AI system, but not every AI system learns from data. Deep learning is in turn a subset of ML that trains multi-layered neural networks, and generative AI is a subset of deep learning; you do not need their internals for this exam, only that each nests inside the one before it the same way ML nests inside AI.

The practical dividing line between ML and ordinary software is who writes the logic. In traditional programming a developer codes explicit if-then rules: "if the transaction is over $10,000 and from a new country, flag it." In ML you instead supply examples (labelled fraud and non-fraud transactions) and the training process derives the rule itself. That is why ML earns its keep on problems whose rules are too varied or too subtle to write out by hand, like recognising what is in an image.

ML problems fall into a few recurring problem types, and naming the shape of a business question is the first decision you make. Google groups the learning approaches into four model types: supervised, unsupervised, semi-supervised, and reinforcement learning^[2]. For the Cloud Digital Leader you should be able to recognise the everyday problem shapes those approaches address:

Problem type	Question it answers	Everyday example
Classification	Which category does this belong to?	Is this email spam or not?
Regression	What is the numeric value?	How many units will we sell next month?
Clustering	Which items naturally group together?	Segment customers with no predefined labels
Recommendation	What should we suggest next?	Which movie to show this viewer
Forecasting	How will this series move over time?	Project demand from past sales

Classification and regression learn from labelled examples (supervised learning, which "uses labeled training data ... to map a specific feature to a label"^[2]); clustering finds structure in unlabelled data (unsupervised learning, which "uses unlabeled data to learn patterns"^[2]). If a business question does not reduce to one of these shapes, ML is probably not the answer.

This section draws the line you will be tested on most: when a scenario calls for machine learning versus traditional data analytics or business intelligence (BI). The two are not competitors so much as answers to different questions, and the exam rewards picking the cheaper, simpler one when it fits.

Analytics and BI explain the past. They aggregate and visualise historical data so a human can read what happened and why: sales by region last quarter, the trend in support tickets, a dashboard of key metrics. The logic is written by people: an analyst composes the query, defines the chart, sets the threshold. The output is for human interpretation.

ML predicts and decides on new cases. Instead of summarising known data, an ML model generalises to records it has never seen, producing a per-case prediction, classification, or recommendation. Where a BI report tells you 4% of last month's transactions were fraudulent, an ML model scores each incoming transaction for fraud in real time. The shared mental model from the previous section explains the difference cleanly: BI runs human-written logic over historical data, while ML runs learned logic over future data.

The decision splits on one question, drawn below: do you need a report on what already happened, or a prediction about each new case?

• Choose analytics / BI when the goal is reporting and a person reading a chart: descriptive, backward-looking, structured data.
• Choose ML when you need a prediction about each new case, at a scale or speed no analyst could match, or when the data is unstructured (images, audio, text) that dashboards cannot easily query.

The relationship is also layered, not adversarial: AI as a field includes "many different disciplines including data analytics, statistics, hardware and software engineering"^[2], so analytics often feeds ML. The same warehouse of historical data that powers a dashboard is the training data a model learns from. A common exam stem describes a team that already reports on historical data and now wants to predict or act on each new record; that pivot from describing the past to predicting the future is the signal to move from BI to ML.

This section covers why an organisation reaches for ML and the single biggest factor in whether the project succeeds. The Cloud Digital Leader exam is a business exam, so it tests the value proposition and the prerequisites, not the algorithms.

ML creates business value in three recurring situations, all variations on "a human cannot keep up":

• Scaling decisions. When the same decision must be made many times, quickly, and consistently, ML applies one learned rule across millions of cases, far beyond what a team of analysts could review. Google lists fraud detection, identifying security threats, personalization and recommendations, and automated customer service as core business functions ML supports^[2].
• Finding patterns in large datasets. As "the rate of data generation is accelerating ... machine learning helps make it possible to analyze and find value in this vast amount of data"^[2], patterns too large or subtle for manual inspection.
• Unlocking unstructured data. Images, audio, video, and free text carry enormous value that row-and-column analytics cannot easily query. ML extracts meaning from them (transcription, translation, image recognition), turning previously untapped data into usable signal.

But ML is only as good as its data, and this is the prerequisite the exam returns to. The diagram below traces why: training data sets a ceiling that flows through training into every prediction the model makes at inference. A model learns whatever its training data teaches it, so the data's quality and representativeness set the ceiling on the model. Google states the dependency directly: "Assuming the training data is of high quality, the more training samples the machine learning algorithm receives, the more accurate the model will become"^[2]. Note the precondition. High-quality data is accurate, complete, consistent, and representative of the real population the model will serve.

The consequence worth remembering: training data that is biased or unrepresentative produces a biased model, no matter how good the algorithm. If the historical data under-represents a group or encodes a past prejudice, the model will reproduce that bias at scale. This is why data preparation and governance, not the choice of model, are usually where an ML project is won or lost, and why a scenario describing "poor, scarce, or skewed data" is a signal to fix the data before investing in modelling.

This section defines the two governance ideas the exam expects every cloud leader to know, explainable AI and responsible AI, and shows how they shape AI decisions. They are requirements you gather up front, not features bolted on after a model ships.

Explainable AI means a model's predictions can be understood and interpreted by people: knowing why the model produced a given output, not just what it output. This matters for four reasons that recur in exam scenarios: building user trust, debugging a model that behaves oddly, checking for unfair bias, and meeting regulatory obligations that demand a decision be justified. A model nobody can interpret is hard to trust and hard to defend, so when a scenario stresses transparency, auditability, or "the bank must explain why a loan was declined," explainability is the concept being tested. (Google offers tooling in this area; the underlying exam concept is the interpretability of model outputs, which Google lists alongside fairness, privacy, and security as responsible-practice areas^[3].)

Responsible AI is the broader practice that contains explainability, drawn below as an umbrella over its pillars: building and operating AI that is fair, accountable, safe, and privacy-respecting. Google frames this through its published AI Principles, a set of commitments that guide how AI should, and should not, be developed and used. Google's generative AI services "are designed with Google's AI Principles in mind,"^[3] and the practice spans fairness (avoiding unjust bias), transparency and interpretability, privacy and security, and safety.

The relationship to keep straight: explainability is one pillar of responsible AI, not a synonym for it. Responsible AI is the umbrella: fairness, accountability, safety, privacy, and explainability together; explainable AI is specifically the "can a human understand this output" pillar. Reflecting the shared-responsibility theme that runs through this exam, the cloud provider supplies tools and safety filters, but the customer "remains responsible" for testing models and adhering to the acceptable-use and prohibited-use policies^[3]. Responsible AI is a shared obligation, not something the platform fully handles for you.

Exam-pattern recognition. Stems that pair AI with the words trust, fairness, bias, transparency, or regulation are testing responsible/explainable AI, and the right answer favours interpretable models, fairness checks, and human oversight. A scenario about predicting per-case outcomes at scale points to ML over analytics. A scenario where data is poor, scarce, or unrepresentative points to fixing data quality first. And the classic distractor, treating analytics/BI and ML as interchangeable, is wrong because one describes the past while the other predicts new cases.