3.1 AI and ML Fundamentals

Defining Artificial Intelligence and Machine Learning

Artificial intelligence (AI) is the broad field of building machines that mimic intelligent human behavior — perceiving, reasoning, learning, and deciding. Machine learning (ML) is a subset of AI: instead of programming explicit rules, ML systems learn patterns from data and apply those patterns to new inputs. Inside ML sits deep learning, which uses multi-layered neural networks to model complex relationships such as images, speech, and language. Inside deep learning sits generative AI (gen AI) — models that produce new content (text, images, code, audio) rather than only classifying or scoring existing inputs. Foundation models such as Gemini are the large, broadly trained models that gen AI applications build on.

The hierarchy is nested, not parallel:

AI (broad goal: machines that act intelligently)
  └─ ML (subset of AI: learn patterns from data)
       └─ Deep learning (subset of ML: multi-layer neural networks)
            └─ Generative AI (produces new content; built on foundation models)
                 └─ Agentic AI (gen AI that plans and takes actions toward a goal)

The updated exam guide leans on the two lower rows, so be able to separate them: generative AI produces content when asked; agentic AI pursues a goal, deciding which steps to take and calling tools or other systems to take them. A model that drafts a reply to a support ticket is generative; a system that reads the ticket, looks up the order, issues the refund, and then writes the reply is agentic.

A spam filter that learned word frequencies from millions of emails is ML; a rules-based filter where a human wrote "if subject contains 'lottery', block" is not. The exam often tests this distinction: learning from data is the defining trait of ML.

ML Versus Data Analytics and Business Intelligence

Traditional business intelligence (BI) and data analytics answer questions about the past and present: What happened? Why did it happen? They use dashboards, SQL queries, and statistical summaries over historical data. They do not generalize — a BI report tells you last quarter's sales, but it cannot predict next quarter's for a customer it has never seen.

ML, by contrast, answers questions about the future and about new cases: What will happen? What should we do? A trained model generalizes from historical examples to score, predict, or recommend for inputs it has never encountered. ML is therefore predictive and prescriptive, while BI is descriptive and diagnostic.

The two are complementary: BI surfaces the trends that justify building an ML model, and ML predictions feed back into BI dashboards.

Types of Problems ML Solves

ML is not one technique — it is a toolbox mapped to problem families:

• Classification — Assigning inputs to discrete labels. Is this transaction fraudulent or legitimate? Is this tumor malignant or benign? Is this email spam or not?
• Regression — Predicting a continuous number. How much will this house sell for? How many units will sell next month?
• Clustering — Grouping similar items without pre-labeled categories. What customer segments exist in our data?
• Recommendation — Suggesting items to users based on behavior and similarity. Which product should we show next?
• Anomaly detection — Flagging rare, suspicious, or unusual cases. Is this sensor reading abnormal? Is this login attempt from a bot?
• Natural language processing (NLP) — Understanding and generating human language. Translation, sentiment analysis, question answering, chat.
• Computer vision — Extracting meaning from images and video. Detecting objects, reading text in images (OCR), classifying defects on an assembly line.

On the exam, match the problem type to the business question: "group customers with no known segments" = clustering; "predict a price" = regression; "flag fraud" = classification or anomaly detection.

Business Value ML Creates

The Domain 3 blueprint names three concrete value pathways, and these are exam favorites:

1. Work with large datasets. ML is the only practical way to extract signal from datasets too large for humans to inspect — billions of log lines, millions of transactions, petabytes of images. A human cannot read 50 million reviews; a model can classify their sentiment in minutes.
2. Scale business decisions. A credit analyst can underwrite a few loans a day; an ML model underwrites thousands per minute with consistent criteria. ML turns one-off expert judgment into automated, repeatable decisions applied at scale.
3. Unlock unstructured data. Roughly 80–90% of enterprise data is unstructured — text documents, images, audio recordings, video. BI cannot query a PDF or a photo. ML (especially NLP and computer vision) turns that unstructured mass into searchable, classifiable, actionable information.

These three points map directly to exam scenarios. If a use case emphasizes "millions of rows," think large datasets; if it emphasizes "every transaction in real time," think scaling decisions; if it emphasizes "support tickets, call transcripts, product images," think unlocking unstructured data.

Why High-Quality, Accurate Data Is Essential

ML models learn from data — nothing else. If the data is wrong, incomplete, or biased, the model will be wrong, incomplete, or biased. This is the garbage-in, garbage-out principle, and on Google Cloud it is amplified because models often operate at scale: a 1% error rate across a billion decisions is ten million mistakes.

The dimensions of data quality the exam expects you to know:

• Accuracy — Labels and feature values must be correct. A model trained on mislabeled fraud cases will misclassify real fraud.
• Completeness — Missing fields or missing populations distort learning. If training data excludes a customer segment, the model will underperform for them.
• Representativeness — Training data must resemble production data. A model trained on winter sales will not generalize to summer shoppers.
• Consistency — Duplicate, conflicting, or unit-mismatched records confuse the model.
• Bias and fairness — Historical data encodes historical decisions, including biased ones. A hiring model trained on past hires may perpetuate past discrimination unless bias is measured and mitigated.

Data quality is not a one-time checkpoint; it is a lifecycle concern. Data drift (production inputs shifting away from training data) and concept drift (the relationship between inputs and outputs changing) mean models degrade over time and must be retrained.

Updated-exam addition - the six named dimensions of data quality. The exam guide effective August 12, 2026 asks you to identify "the main dimensions of data quality" and names six specific ones. Learn this list as a list, because a question can ask which dimension a described failure violates:

The pairs that get confused are validity versus accuracy - a value can be perfectly valid in format and still be false - and uniqueness versus consistency - duplicates are one record appearing twice, inconsistency is two systems disagreeing about one fact.

Explainable and Responsible AI

A model that predicts but cannot explain why is risky: regulators, customers, and internal reviewers demand transparency. Explainable AI provides interpretability — which features drove a prediction, how confident the model is, and how a single decision was reached — so humans can trust, audit, and challenge model outputs.

Responsible AI broadens the lens beyond accuracy to include:

• Fairness — Does the model disadvantage protected groups? Measured with fairness metrics; mitigated with debiasing techniques and representative data.
• Transparency — Can stakeholders understand what the model does and why it made a specific decision?
• Accountability — Who is responsible when the model errs? Governance, review boards, and clear ownership.
• Safety and reliability — Does the model behave robustly under adversarial or unusual inputs?
• Privacy — Does the model expose or memorize sensitive training data?

Google published its AI Principles in 2018 — a set of seven objectives including safety, fairness, privacy, and societal benefit — and operationalizes them through tools such as Explainable AI, the What-If Tool, Model Cards for documenting model performance and limitations, and Responsible AI practices guidance. The exam expects you to connect these concepts: explainability enables trust, fairness enables equity, and together they make ML deployment responsible.

Key Takeaways

• AI is the broad goal; ML is the subset that learns from data; deep learning is the neural-network subset of ML.
• BI describes the past; ML predicts and prescribes for new cases and generalizes to unseen data.
• Match the problem type to the question: classification labels, regression predicts numbers, clustering groups, recommendation suggests, anomaly detection flags, NLP reads text, computer vision reads images.
• ML creates value through large datasets, scaling decisions, and unlocking unstructured data.
• Garbage-in, garbage-out: accuracy, completeness, representativeness, and bias all determine model quality.
• Responsible AI combines explainability, fairness, transparency, and accountability — supported by Google's AI Principles and tooling.