GAIL Module 2: Unlock Foundational Concepts

Module 2 goes underneath the vocabulary from Module 1 and asks what gen AI is actually made of. The through-line is data: a model’s behavior comes from the data it learned on, so data quality determines capability, data gaps determine limitations, and data handling determines whether the whole thing is secure and lawful. That single dependency explains why one module covers definitions, data pipelines, model limitations, and responsible AI — they’re the same subject viewed at four distances.¹

The hierarchy: AI, ML, deep learning, gen AI

These four terms nest, and the exam tests placement rather than definition recall.

• Artificial intelligence — the broad field of building machines that do tasks normally requiring human intelligence. A field, not a technique.
• Machine learning — a subset of AI where machines learn from data rather than from explicit programming. The consequence worth internalizing: different data creates different models. Change the data and you get a different system, even with identical code.
• Deep learning — a specific ML technique that teaches computers using artificial neural networks. Neural networks can leverage both labeled and unlabeled data, which makes deep learning semi-supervised. This is what gives gen AI the ability to produce text, images, and audio.
• Generative AI — an application of the above, focused on creating new content.

So: AI ⊃ ML ⊃ DL, with gen AI built on top of deep learning.

Exam note: three definition questions map one-to-one onto this list — “learns from data without explicit programming” is machine learning, “the broad field of computer science” is artificial intelligence, and “multi-layer neural networks for complex patterns” is deep learning.

Why data is the lever

An ML model analyzes the data it’s given, identifies patterns, and calculates the likelihood of outcomes when it meets new information. Nothing in that loop corrects for bad input — so the quality and accessibility of data set a ceiling on everything the model can do.

Exam note: a model in ML is “a complex mathematical structure that processes inputs to generate outputs.” The distractors anthropomorphize it or describe it as a database.

Data quality — five factors

• Accuracy — inaccurate data teaches incorrect patterns, producing faulty predictions. Wrong dates and mislabeled details in a training set.
• Completeness — the dataset needs enough size and enough representation within it for the model to predict accurately.
• Representative — data must be inclusive; a skewed sample produces biased outcomes. Distinct from completeness: you can have a large dataset that still represents only one kind of case.
• Consistency — inconsistent formats or labeling confuses the model and hinders learning. The same field recorded as “NY” in some rows and “New York” in others.
• Relevance — the data has to fit the task the AI is designed to perform.

Exam note: factual errors in training data compromise accuracy — not completeness, consistency, or relevance, all of which appear as distractors. Separately: large volume is often beneficial but is not the only data aspect influencing performance.

Data accessibility — three factors

• Availability — no data, no training. The first question, not the last.
• Cost — acquiring high-quality data can itself be the barrier that stops an AI project.
• Format — data must arrive in a form the model can process.

Structured vs. unstructured data

• Structured — organized and easily searchable. Customer ID, name, delivery address, purchase date, order cost — and a 1–5 star rating, which is structured despite being “feedback.”
• Unstructured — no predefined structure, messy and complex, not easily organized. A product image, free-text customer feedback, the body of an email.

Three types of machine learning

The three are distinguished by what the training data looks like, which is the fastest way to sort a scenario.

• Supervised learning — trained on labeled data, where tags assign meaning to each example. The labels let the algorithm learn the input→output relationship and predict on new, unseen inputs. A spam classifier trained on manually labeled emails.
• Unsupervised learning — trained on unlabeled data with no inherent correct answer. The model finds natural groupings in raw data; it’s exploratory, uncovering structure nobody specified in advance. Topic modeling that discovers the themes in a document collection.
• Reinforcement learning — learns through interaction and feedback, receiving rewards or penalties and discovering which actions produce the best outcomes. Useful precisely when you can’t supply explicit instructions or labeled data. A game-playing AI improving by trial and error.

Google Cloud examples, as the slides pair them:

Use case	Learning type	Tool
Predictive maintenance — predict machine failure from sensor data	Supervised	Agent Platform
Anomaly detection — flag transactions deviating from the norm	Unsupervised	BigQuery ML
Product recommendations — maximize engagement and sales	Reinforcement	Agent Platform

Exam note: agents in reinforcement learning learn primarily by interacting with their environment and receiving feedback — not from labeled examples or from being programmed with rules.

The ML lifecycle

Five stages, each with the Google Cloud tools the exam associates with it. This same five-stage frame returns later in the module as the skeleton for security controls, so it’s worth learning once properly.

1. Gather — determine what data you need based on the desired outcome. Pub/Sub for real-time streaming, Cloud Storage for unstructured data, Cloud SQL and Spanner for structured data.
2. Prepare — clean and transform raw data into a usable format, including formatting and labeling. BigQuery for analysis, BigQuery universal catalog for governance.
3. Train — create the model from the prepared data. Agent Platform as a managed training environment.
4. Deploy and predict — make the trained model available for use. Agent Platform.
5. Manage — maintain models over time: versioning, performance tracking, drift monitoring. Agent Platform Feature Store for data management, Model Garden for storage, Agent Platform Pipelines for automation.

Identity and Access Management (IAM) runs across all of it: create and manage user accounts, assign roles, grant and revoke permissions to resources, audit user activity, and monitor your security position.

Foundation models

Foundation models use deep learning and train on massive datasets, which lets them learn complex patterns and perform a variety of tasks across domains. Module 1 covered what makes them flexible and adaptable; this module adds the two families and the selection criteria.

• Large language models (LLMs) — understand and generate human language: translation, writing, question answering.
• Diffusion models — generate high-quality images, audio, and video by iteratively refining data and patterns. The iterative refinement is the mechanism that distinguishes them, and it’s why they dominate image generation.

Exam note: the best model type for producing photorealistic images from a text description is a diffusion model. An LLM is the reflexive wrong answer.

Google Cloud’s gen AI models:

• Gemini — multimodal understanding, advanced conversational AI, content creation, question answering.
• Gemma — lightweight, user-friendly, and customizable; built for local deployments and specialized applications.
• Nano Banana — generates high-quality images from text descriptions.
• Veo — generates video from text descriptions or still images.

Choosing a model — eight factors

• Modality — the data types the model can process and generate.
• Context window — how much information the model can consider at once when generating a response.
• Security — model security features and industry standards, critical with sensitive data.
• Availability and reliability — uptime guarantees, redundancy, disaster recovery for production use.
• Cost — pricing model and cost effectiveness; match model size to task rather than defaulting to the largest.
• Performance — accuracy, speed, efficiency, evaluated on relevant benchmarks.
• Fine-tuning and customization — whether the model can be specialized.
• Ease of integration — well-documented APIs and SDKs, fit with existing systems.

Foundation model limitations

Each limitation traces back to the training data, which is why the fixes in the next section are all about supplying better or more current information.

• Data dependency — performance is bounded by training data quality; bias or incompleteness seeps straight into outputs.
• Knowledge cut-off — the last date the model saw new information. Anything after it simply isn’t in the model.
• Bias — training data contains societal biases, and subtle ones can be magnified in outputs rather than merely reproduced.
• Fairness — different people interpret it differently, and fairness assessments have inherent limits: they typically target specific bias categories and can overlook others. The point is that “we tested for fairness” is never a complete claim.
• Hallucinations — outputs that aren’t accurate or grounded in real information. The model has no internal signal distinguishing recall from invention.
• Edge cases — rare or atypical scenarios expose weaknesses, producing errors, misinterpretations, and unexpected results.

Techniques to overcome limitations

Four techniques plus human oversight. They escalate in cost and effort, and the exam cares about knowing which to reach for when.

• Prompt engineering — crafting precise prompts to guide the model toward the output you want. Cheapest intervention; try it first.
• Grounding — anchoring the model in specific data, such as company documents, so responses are accurate, relevant, and enterprise-specific. This is the direct fix for hallucinations: instead of relying on what the model absorbed during training, you hand it source material at generation time, which makes the answer both more accurate and more trustworthy.
• RAG (retrieval-augmented generation) — a grounding method that uses search. Three steps: retrieve relevant information from a knowledge base by meaning, augment the prompt with it, then generate the response. Covered in full, with the retrieval tooling, in Module 5.
• Fine-tuning — further training a foundation model on a new, task-specific dataset so it excels in a particular area or output format. The heavyweight option, reached for when prompting isn’t enough.
• Humans in the loop (HITL) — integrating human expertise where judgment or context is required. Use cases: content moderation, sensitive applications, high-risk decision making. Applied at pre-generation review and post-generation review.

Exam note: the four techniques are grounding, prompt engineering, fine-tuning, and HITL — distractors offer more compute or better hardware, which address speed, not limitation. And fine-tuning is specifically right when prompt engineering alone doesn’t achieve the desired outcome and the model needs specializing for a task or output format with a new dataset.

Building AI securely

The security section reuses the ML lifecycle, applying controls stage by stage. The logic is that an AI system has more attack surface than a conventional app — the data, the trained parameters, and the outputs are each separately worth protecting.

1. Gather — secure data is the foundation; protect it at all times and control who can access, add to, and input data.
2. Prepare — pay special attention to confidential data in the training set: anonymization, validation, secure processing, logging, and real-time monitoring.
3. Train — safeguard both the training data and the model parameters from unauthorized access or modification. The parameters are the asset here, not just the inputs.
4. Deploy and predict — control access to the model, and verify sources and check for vulnerabilities in any pre-built models you adopt.
5. Manage — stay current on best practices, update regularly, monitor performance and outputs for anomalies or tampering, and review access permissions on a schedule.

Across all stages: guard against adversarial attacks, and monitor outputs to prevent leaks and harmful content.

SAIF (Secure AI Framework) establishes security standards for building and deploying AI systems responsibly. Google Cloud supports it with secure-by-design infrastructure, encryption, IAM, Security Command Center, and monitoring tools. The Google Threat Intelligence Group’s global insights combined with Mandiant’s frontline expertise shift protection from reactive to proactive, AI-driven operational readiness.

Exam note: SAIF’s goal is to establish security standards for building and deploying AI responsibly, addressing threats unique to the AI landscape. It is not about restricting innovation, and not solely about external attacks. Separately, the key aspect of securing the training phase is safeguarding training data and model parameters from unauthorized access.

Building AI responsibly — four foundations

• Transparency — users need to understand how their information is used and how the system works.
• Privacy — anonymize or pseudonymize data, and safeguard against models inadvertently leaking sensitive training data. Note the second half: the model itself can be the leak.
• Data quality, bias, and fairness — ethical AI requires high-quality data and responsible use of it. AI inherits societal biases and produces unfair outcomes as a result, so fairness has to be designed in from the start rather than audited on at the end.
• Accountability and explainability — fairness requires someone accountable. Explainable AI makes decision-making transparent and understandable, and you need to know how your own application uses and interprets the model’s output.

Exam note: for an ethically sound AI system — the loan-assessment scenario — the two correct actions are regular audits of model performance to identify and mitigate emerging biases and training on a diverse dataset representing different demographics and socioeconomic backgrounds. Black-box models and unsupervised deployment are the distractors. The primary goal of ethical AI development is to ensure AI systems are used responsibly and do not cause harm.

Legal implications

Four legal areas carry the risk: data privacy, non-discrimination, intellectual property, and product liability. AI laws require responsible data handling, bias mitigation, transparency, and model compliance — and the landscape is still evolving, so trustworthy AI requires ongoing vigilance and legal counsel rather than a one-time review.

Exam note: the key legal responsibility is adherence to data privacy laws, non-discrimination principles, and the specific licensing terms of AI models. Model licensing is the clause people forget.

Footnotes

1. Module 2 slide deck (Generative AI Leader course) ↩