AI and Emerging Technology Fundamentals

Introduction

Artificial intelligence (AI) has moved from specialist research laboratories into everyday branding practice. Search engines rank content, streaming services recommend entertainment, retailers personalize offers, social platforms moderate posts, and generative systems produce text, images, audio, and video. For brand managers, the important question is not simply whether an organization uses AI, but how a system works, what it can reasonably do, where it may fail, and how human judgment should govern its use.

AI is best understood as a family of computational methods rather than a single technology. Some systems follow explicit rules; others learn statistical patterns from data. Some classify existing information, while others generate new content. Some optimize a narrow task, while others support many tasks through a general-purpose model. These differences affect brand strategy, customer experience, risk, cost, and accountability.

This chapter establishes the technical vocabulary required for later study. It explains algorithms, machine learning, deep learning, natural language processing, computer vision, generative AI, foundation models, recommendation systems, automation, and human–AI collaboration. It also examines limitations, governance, and the practical implications for brand professionals.

Learning Objectives

By the end of this chapter, you should be able to:

1. Define an algorithm and distinguish rule-based systems from learning systems.
2. Explain supervised, unsupervised, and reinforcement learning.
3. Describe how deep learning uses layered neural networks.
4. Identify the purposes and limitations of natural language processing and computer vision.
5. Distinguish generative AI from predictive and classificatory AI.
6. Explain foundation models, prompting, fine-tuning, and retrieval-augmented generation.
7. Describe how recommendation systems operate and why they can produce bias or filter bubbles.
8. Evaluate automation and human–AI collaboration in branding contexts.
9. Recognize risks involving data quality, privacy, intellectual property, fairness, security, and accountability.
10. Apply a structured approach to selecting and governing AI for a brand problem.

1. Algorithms: The Basic Logic of Computation

An algorithm is a finite, ordered set of instructions for transforming inputs into outputs or decisions. A recipe, a sorting procedure, a pricing rule, and a content-ranking process can all be described as algorithms. An algorithm does not necessarily learn; it may simply execute instructions written by people.

A typical branding example is a rule-based promotional engine:

• If a customer has purchased athletic shoes within the last twelve months, classify the customer as a sportswear prospect.
• If the customer has consented to email communication, send a relevant offer.
• If the customer has not opened three consecutive messages, reduce contact frequency.

Such rules may be transparent and easy to audit, but they can be rigid. They may fail when customer circumstances are complex or when the rules do not anticipate new patterns.

A useful abstract representation is:

Here, x represents the input data, f represents the algorithm or model, θ represents parameters or rules, and ŷ represents the output or prediction. In a fixed rule system, θ is specified by programmers. In machine learning, some or all of θ is estimated from data.

An effective algorithm should be sufficiently accurate for its purpose, computationally feasible, robust to reasonable variation, and appropriate to the legal and ethical context. Accuracy alone is not enough. A highly accurate system that violates privacy or discriminates against customers may be unacceptable.

2. From Algorithms to Machine Learning

Machine learning (ML) is a field of AI in which a system uses data to estimate patterns, relationships, or decision policies rather than relying entirely on manually written rules. The system is trained using examples and then evaluated on data that were not used for training.

2.1 Supervised learning

In supervised learning, each training example contains an input and a known target or label. A brand might train a model using customer interactions labelled “likely to churn” or “not likely to churn.” The model learns a mapping from customer features to an estimated outcome.

Common supervised tasks include:

• Classification: assigning categories, such as positive, neutral, or negative sentiment.
• Regression: predicting a numerical value, such as expected customer lifetime value.
• Ranking: ordering products, messages, or search results by predicted relevance.

A generic training objective is to minimize average loss:

In this expression, n is the number of training examples, xᵢ is the input for example i, yᵢ is its known target, f(xᵢ; θ) is the model output, L is a loss function measuring error, and θ* is the parameter setting that minimizes average training loss. This formulation assumes that examples are represented consistently and that the chosen loss reflects the business objective. Minimizing training loss does not guarantee good performance on new data.

For binary classification, a frequently used loss is binary cross-entropy:

Here, y is either 0 or 1, and p is the model’s estimated probability that the class is 1, with 0 < p < 1. The loss penalizes confident incorrect predictions more heavily than uncertain predictions. It is applicable when the target is binary and the model output can be interpreted as a probability.

2.2 Unsupervised learning

Unsupervised learning uses data without target labels. Clustering may group customers according to browsing, purchasing, or engagement patterns. Dimensionality-reduction methods may represent many customer variables in a smaller number of dimensions for visualization or analysis.

Unsupervised segments are not automatically meaningful market segments. A mathematical cluster may reflect channel behavior, data collection practices, or purchasing frequency rather than a genuine customer need. Brand managers must interpret clusters using research, context, and validation.

2.3 Reinforcement learning

Reinforcement learning concerns an agent that takes actions in an environment and receives rewards or penalties. The agent seeks a policy that improves cumulative reward over time. This framework can be relevant to dynamic recommendations, conversational interaction, or allocation problems, although many commercial systems use simpler supervised or ranking methods.

A reinforcement-learning system may optimize a short-term metric, such as clicks, while harming a long-term objective, such as trust or retention. Therefore, the reward function must represent the organization’s real priorities rather than an easily measured proxy.

2.4 Training, validation, and test data

A model should be evaluated on data distinct from its training data. Training data are used to estimate parameters; validation data support model selection and tuning; test data provide a final estimate of performance. If a model is repeatedly adjusted using the test set, the test result becomes optimistic because information from the test set has influenced development.

Two central problems are:

• Overfitting: the model memorizes training-specific patterns, including noise, and performs poorly on new cases.
• Underfitting: the model is too simple to represent useful relationships.

Generalization means performing satisfactorily on relevant future data. In branding, future data may differ because of seasonality, economic change, competitor activity, platform changes, or shifts in consumer language. A model should therefore be monitored after deployment.

3. Deep Learning and Neural Networks

Deep learning is a form of machine learning based primarily on neural networks with multiple computational layers. Each layer transforms its input into a new representation. Early layers may detect simple patterns, while later layers combine them into more abstract representations. In an image system, early layers may respond to edges and later layers may contribute to object or scene recognition.

A simplified neuron computes a weighted sum followed by an activation function:

Here, x is an input vector, w is a vector of learned weights, b is a bias term, z is the pre-activation value, φ is an activation function, and a is the output. The dimensions of w and x must be compatible. A network combines many such units in layers.

Training commonly uses gradient-based optimization. The model calculates a loss, determines how each parameter contributed to that loss, and adjusts parameters in a direction expected to reduce it. The details depend on the architecture, optimizer, data, and objective.

Deep learning is powerful because it can learn representations directly from large and complex datasets. It supports speech recognition, image analysis, translation, search, recommendation, and generative systems. Its limitations include high data and computing requirements, sensitivity to distribution shifts, limited interpretability, and the risk that learned patterns reproduce social or historical bias.

Deep learning should not be equated with human understanding. A system may recognize statistical associations without possessing common sense, intentions, lived experience, or reliable knowledge of the physical and social world.

4. Natural Language Processing

Natural language processing (NLP) applies computational methods to human language. NLP tasks include classification, sentiment analysis, named-entity recognition, information extraction, translation, summarization, question answering, speech recognition, and text generation.

Older NLP systems often relied on dictionaries, grammar rules, and manually designed features. Modern systems frequently represent words or tokens as vectors and use neural architectures to model relationships across a sequence. A token may be a word, part of a word, punctuation mark, or another unit selected by the system’s tokenizer.

For brand management, NLP can support:

• analysis of reviews and social conversations;
• routing and summarizing customer-service cases;
• search and question answering;
• multilingual campaign adaptation;
• detection of recurring complaints;
• drafting of variants for human review.

NLP remains difficult because meaning depends on context, culture, irony, ambiguity, domain terminology, and conversational history. Sentiment analysis may misread sarcasm. A model may interpret a reclaimed term as offensive or miss an indirect complaint. Translation may preserve literal meaning while damaging cultural nuance. Language models can also generate fluent but false statements, often called hallucinations.

A brand should not treat linguistic fluency as evidence of factual accuracy or cultural appropriateness. High-risk communications require human review and, where relevant, native-language expertise.

5. Computer Vision

Computer vision enables systems to analyze images and video. Tasks include image classification, object detection, segmentation, optical character recognition, face analysis, quality inspection, and visual similarity search.

The tasks differ. Classification assigns one or more labels to an image. Object detection identifies objects and their locations. Segmentation assigns labels to individual pixels or regions. Optical character recognition converts text in an image into machine-readable characters.

Applications in branding include:

• monitoring the presence and condition of brand assets in retail environments;
• checking packaging or point-of-sale displays;
• searching product images by visual similarity;
• analyzing creative variations;
• supporting accessibility through image descriptions;
• detecting unsafe or prohibited user-generated content.

Vision systems can fail because of poor lighting, occlusion, unusual viewpoints, low image quality, cultural variation, or under-representation in training data. Face recognition and biometric inference raise especially serious privacy, consent, and discrimination concerns. A system’s confidence score is not a guarantee that its interpretation is correct.

6. Generative AI and Foundation Models

Generative AI creates new outputs such as text, images, audio, video, code, or structured data. Predictive AI estimates an existing target; generative AI produces an output according to learned patterns and a prompt or other conditioning input. The distinction is practical rather than absolute: a generative model may also classify, and a predictive system may generate recommendations.

A foundation model is a large model trained on broad data and designed to support many downstream tasks. It may be adapted through prompting, additional training, fine-tuning, or connection to external tools and data. Large language models are foundation models specialized primarily for language, although many are multimodal.

Important concepts include:

• Prompt: instructions or input supplied to a model.
• Context window: the amount of input and conversational material the model can process in one interaction.
• Fine-tuning: additional training on a task-specific dataset.
• Embedding: a numerical representation of content used to compare semantic relationships.
• Retrieval-augmented generation (RAG): retrieving relevant external material and providing it to a generative model before it produces an answer.
• Multimodal model: a model that processes or generates more than one modality, such as text and images.

A RAG workflow can reduce unsupported answers when the retrieved documents are authoritative and current. It does not eliminate error: retrieval may be incomplete, documents may conflict, and the model may still misinterpret the evidence. Fine-tuning may improve style or task performance but can also encode unwanted bias or cause the model to overfit a narrow dataset.

Generative AI offers brand benefits such as rapid ideation, localization, prototyping, and content variation. It also creates risks involving factual accuracy, brand voice, hidden bias, copyright, personal-data leakage, impersonation, and the production of synthetic content that audiences mistake for authentic human communication.

7. Recommendation Systems

A recommendation system selects or ranks items that may be relevant to a user. Items may include products, videos, articles, advertisements, or messages. The system may use:

• Content-based filtering: recommends items similar to what a user previously engaged with, based on item attributes.
• Collaborative filtering: uses patterns of interaction among users and items; users with similar behavior may receive similar recommendations.
• Hybrid systems: combine content, collaborative, contextual, and rule-based signals.

A recommendation system generally estimates relevance and then ranks candidates. A simplified scoring representation is:

Here, s(u,i) is the score for user u and item i, r(u,i) is predicted relevance, q(i) may represent item quality or business value, c(u,i) is contextual suitability, and α, β, and γ are chosen weights. The variables must be normalized or otherwise made comparable before a weighted sum is interpreted. This is a conceptual scoring model, not a universal recommendation formula.

Recommendation quality should not be measured only by clicks. Other measures include conversion, retention, diversity, novelty, coverage, customer satisfaction, complaint rates, and long-term trust. Optimizing engagement alone may promote sensational or repetitive content, narrow users’ exposure, and create filter bubbles. Cold-start problems occur when a new user or item has little interaction history.

Brand managers should ask whether recommendations serve customer needs, whether the customer can understand and control personalization, and whether commercial incentives distort relevance.

8. Automation and Human–AI Collaboration

Automation occurs when a system performs a task with limited direct human intervention. Automation can be deterministic, such as a rule that sends a confirmation message, or probabilistic, such as a model that prioritizes service cases.

A useful distinction is between:

• Automated execution: the system performs an approved, low-risk action.
• Decision support: the system provides evidence or suggestions while a person decides.
• Human-in-the-loop: a person must approve or modify the system output.
• Human-on-the-loop: the system operates while a person monitors performance and can intervene.
• Human-in-command: people retain authority over objectives, boundaries, and escalation.

The appropriate arrangement depends on risk, reversibility, scale, and the consequences of error. A model might automatically recommend email subject lines, require approval for public claims, and prohibit autonomous decisions about sensitive eligibility or employment.

Human involvement is not automatically effective. Reviewers may suffer from automation bias, accepting machine outputs too readily, or alert fatigue, ignoring too many warnings. Effective collaboration requires clear responsibilities, useful explanations, adequate training, escalation routes, and sufficient time to review.

Automation should redesign work rather than simply remove human labor. In customer experience, AI may handle routine classification while employees focus on empathy, complex judgment, and relationship repair. In creative work, AI may expand exploration while professionals retain responsibility for strategy, cultural meaning, legal clearance, and final approval.

9. A Brand-Management Framework for Evaluating AI

Before adopting an AI system, a brand team should proceed through six questions.

9.1 Define the decision or customer problem

Specify the business objective, affected stakeholders, decision owner, and acceptable outcome. “Use AI for personalization” is too vague. “Reduce irrelevant product recommendations while preserving customer control and consent” is more actionable.

9.2 Assess data fitness

Examine provenance, consent, relevance, representativeness, accuracy, timeliness, security, and retention. Data collected for one purpose may not be appropriate for another. Historical data may reflect unequal treatment rather than neutral customer preference.

9.3 Select the least complex adequate method

A transparent rule or simple statistical model may be preferable to a complex model when the task is narrow and explainability matters. More sophisticated technology is not automatically better.

9.4 Test performance and harm

Evaluate overall performance and subgroup performance. Test unusual cases, adversarial inputs, accessibility needs, language variation, and distribution shifts. Assess false positives and false negatives, not only aggregate accuracy.

For a binary classifier, precision and recall are useful:

Here, TP is the number of true positives, FP false positives, and FN false negatives. Precision answers, “Of the cases predicted positive, how many were actually positive?” Recall answers, “Of the actual positive cases, how many did the system identify?” These measures apply to a defined positive class and should be interpreted with the business cost of each error.

9.5 Establish governance

Governance includes approval authority, documentation, access controls, audit logs, incident response, vendor management, content review, and mechanisms for customer inquiry or appeal. Policies should define prohibited uses and the information that must never be entered into an external tool.

9.6 Monitor and improve

After launch, track performance, drift, complaints, fairness indicators, security incidents, and unintended effects. A model is not finished when it is deployed. New data, cultural changes, platform changes, and campaign objectives can alter its suitability.

10. Limitations, Ethics, and Regulation

AI systems inherit limitations from their data, objectives, design, deployment environment, and users. Major risks include:

• Bias and discrimination: unequal performance or harmful outcomes across groups.
• Privacy loss: collection, inference, or disclosure of personal information without an appropriate basis.
• Opacity: difficulty explaining how a complex output was produced.
• Hallucination and misinformation: plausible but unsupported generated content.
• Security threats: prompt injection, data poisoning, model extraction, fraud, and account compromise.
• Intellectual-property disputes: uncertain rights concerning training data, generated outputs, or reused creative material.
• Automation complacency: excessive trust in a system that appears objective.
• Environmental cost: energy and infrastructure requirements associated with training and operating large models.
• Brand dilution: inconsistent voice, low-quality synthetic content, or loss of distinctive human meaning.

Ethically responsible practice requires proportionality. The more consequential the decision, the stronger the requirements for transparency, human oversight, documentation, security, and contestability. Legal requirements vary by jurisdiction and change over time; organizations should obtain current legal advice rather than treating a general AI policy as a substitute for compliance.

Transparency does not require revealing proprietary model details in every interaction. It does require avoiding deception and giving people meaningful information about automated processing where it affects their experience or rights. Consent should be specific and understandable, and personalization should not exploit vulnerability.

11. Contemporary Developments

Several developments are reshaping AI-enabled branding. Multimodal models combine text, visual, audio, and video capabilities. Smaller and more efficient models can operate on local devices, reducing latency and potentially improving data control. Agentic systems can plan and call tools, but their increased autonomy makes authorization, monitoring, and failure containment more important. Synthetic data can supplement scarce datasets, although it may reproduce or amplify the biases of the original data. Provenance tools and content credentials seek to communicate how digital media were created or modified, but adoption and technical coverage remain incomplete.

These developments do not remove the need for fundamentals. A larger model is not a strategy. A brand still needs a clear customer problem, reliable evidence, coherent positioning, legal safeguards, and accountable decision-making.

Case Study: Illustrative Scenario—Aurora Outdoor’s Personalization Program

Background

Aurora Outdoor, a fictional mid-sized retailer, sells clothing and equipment through stores, a website, and a mobile application. Its marketing team wants to use AI to recommend products and generate personalized email content. Management hopes to increase relevance without making the brand feel intrusive.

Problem

The existing system sends broad promotions based mainly on past purchases. Click-through rates are acceptable, but customers complain that recommendations are repetitive and sometimes inappropriate—for example, suggesting advanced climbing equipment to customers who purchased beginner accessories once. The organization also has incomplete consent records across older systems.

Analysis

The team first defines the objective as improving relevant product discovery while protecting customer control and avoiding sensitive inference. It inventories data sources and discovers that some historical browsing data lack clear purpose and consent documentation. The team excludes those records until their lawful use is established.

A baseline rule-based system is created using current purchase category, explicit preferences, inventory availability, and season. A hybrid recommender is then tested against this baseline. The team measures conversion, product diversity, complaint rate, unsubscribe rate, and performance for new customers. It does not optimize clicks alone.

The team also identifies a cold-start problem. For new customers, it uses selected popular products, stated preferences, and contextual information rather than pretending to know detailed interests. Recommendations include a visible preference-control option.

Approach and solution

Generative AI is used to draft subject-line alternatives, but factual claims, price statements, safety information, and final copy require human approval. A retrieval system supplies approved product specifications and brand guidelines. The model is prevented from using unapproved external facts. The customer-service team receives summaries of recommendation complaints, with an escalation route for suspected unfairness or privacy concerns.

Before launch, the team conducts subgroup and language testing, reviews data-retention practices, documents vendor responsibilities, and establishes a rollback plan. A monthly governance meeting examines performance and incidents.

Outcome

In the illustrative scenario, the pilot improves product discovery without increasing complaint or unsubscribe rates. The most important success is not a single campaign metric: the organization gains a documented process for data selection, testing, human approval, and ongoing monitoring. It also learns that customers value control and explanation more than maximum personalization.

Learning points

1. Begin with a customer and brand problem, not a technology label.
2. Establish data rights and quality before building a model.
3. Compare AI with a transparent baseline.
4. Use multiple performance and harm measures.
5. Reserve human authority for consequential or externally visible decisions.
6. Treat recommendation and generation as governed services, not one-time software purchases.

Best-Practice Checklist

A brand team introducing AI should:

• define the objective and acceptable risk;
• document data sources, consent, and retention;
• select the simplest adequate technology;
• separate experimentation from production deployment;
• test accuracy, robustness, subgroup outcomes, and cultural fit;
• label or disclose synthetic content where appropriate;
• protect confidential and personal information;
• require human approval for high-impact communications and decisions;
• maintain audit trails and incident procedures;
• monitor post-launch performance and customer feedback;
• provide correction, appeal, or opt-out mechanisms where relevant;
• review vendor contracts, security, intellectual-property terms, and regulatory duties.

Chapter Summary

AI is a collection of methods for transforming data into predictions, decisions, recommendations, or generated outputs. Algorithms may be explicit rules or learned models. Machine learning includes supervised, unsupervised, and reinforcement learning, while deep learning uses multilayer neural networks to learn complex representations. NLP works with language; computer vision works with images and video. Generative AI produces new content, and foundation models provide broad capabilities that can be adapted through prompting, retrieval, or fine-tuning.

Recommendation systems use content, collaborative, contextual, and business signals, but should be evaluated beyond clicks. Automation can improve scale and consistency, yet human–AI collaboration requires clear authority, review, escalation, and monitoring. Across all applications, data quality, privacy, fairness, security, intellectual property, transparency, and accountability are central. For brand managers, responsible AI begins with a well-defined customer problem and ends with continuous governance—not merely with model deployment.

Key Terms

Algorithm: A finite sequence of instructions that transforms inputs into outputs.

Artificial intelligence: Computational systems that perform tasks associated with capabilities such as perception, prediction, reasoning, language, or generation.

Machine learning: Methods that estimate patterns or decision policies from data.

Supervised learning: Learning from examples that contain inputs and known targets.

Unsupervised learning: Learning structure from data without target labels.

Reinforcement learning: Learning actions through interaction and rewards or penalties.

Deep learning: Machine learning using neural networks with multiple learned layers.

Natural language processing: Computational analysis and generation of human language.

Computer vision: Computational analysis of images and video.

Generative AI: AI that creates new text, images, audio, video, code, or other content.

Foundation model: A broadly trained model adaptable to multiple downstream tasks.

Prompt: Input or instructions supplied to a generative model.

Embedding: A numerical representation that captures relationships among items or meanings.

Retrieval-augmented generation: Generation supported by retrieved external information.

Recommendation system: A system that predicts or ranks items for a user or context.

Overfitting: Learning training-specific patterns that do not generalize well.

Automation: Execution of tasks with limited direct human intervention.

Human-in-the-loop: A design in which a person reviews or approves a system action.

Algorithmic bias: Systematic unfairness arising from data, design, deployment, or use.

Model drift: Decline in suitability as real-world data or conditions change.

Review Questions

Short Answer Questions

1. What is an algorithm, and how does it differ from a machine-learning model?
2. Distinguish supervised learning from unsupervised learning.
3. What are two common applications of natural language processing in branding?
4. What is a foundation model?
5. Why is optimizing clicks alone inadequate for evaluating a recommendation system?

Long Answer Questions

1. Explain the machine-learning development process from data preparation through post-deployment monitoring.
2. Compare natural language processing, computer vision, and generative AI in terms of capabilities and limitations.
3. Explain how deep learning differs from traditional rule-based automation.
4. Discuss the benefits and risks of using generative AI in brand-content production.
5. Describe the governance requirements for responsible human–AI collaboration in a customer-facing brand system.

Analytical / Application Questions

1. A retailer wants to automatically approve personalized discounts using purchase history and inferred income. Analyze the data, fairness, privacy, and human-oversight risks, and propose a safer design.
2. A streaming brand’s recommendation system increases viewing time but also increases complaints about repetitive content. Design a revised evaluation framework using at least four performance or harm measures.
3. A global brand uses a language model to translate and localize a campaign. Develop a workflow that addresses factual accuracy, cultural meaning, intellectual property, privacy, human review, and accountability.

Multiple Choice Questions with Answers

Question 1

Which statement best defines machine learning?

A. A system that always follows manually written rules

B. A method that learns patterns or decision policies from data

C. A technology limited to producing images

D. A process for encrypting customer records

Correct Answer: B. A method that learns patterns or decision policies from data

Question 2

Which task is an example of supervised learning?

A. Grouping customers without predefined labels

B. Learning a product ranking from labelled purchase outcomes

C. Randomly sampling website visitors

D. Storing customer data in a database

Correct Answer: B. Learning a product ranking from labelled purchase outcomes

Question 3

What is a major limitation of sentiment analysis?

A. It cannot process any written language

B. It always requires image data

C. It may misinterpret sarcasm and cultural context

D. It produces only numerical recommendations

Correct Answer: C. It may misinterpret sarcasm and cultural context

Question 4

What does generative AI primarily do?

A. Creates new content based on learned patterns and inputs

B. Only sorts existing files alphabetically

C. Removes the need for human governance

D. Guarantees that every factual statement is correct

Correct Answer: A. Creates new content based on learned patterns and inputs

Question 5

What is retrieval-augmented generation designed to do?

A. Replace all training data with random examples

B. Provide retrieved external information to support generation

C. Make a model incapable of producing text

D. Convert every image into a spreadsheet

Correct Answer: B. Provide retrieved external information to support generation

Question 6

Which recommendation approach uses similarities among users’ interaction patterns?

A. Collaborative filtering

B. Optical character recognition

C. Rule compilation

D. Image segmentation

Correct Answer: A. Collaborative filtering

Question 7

If a classifier has 80 true positives and 20 false positives, its precision is:

A. 20%

B. 25%

C. 80%

D. 100%

Correct Answer: C. 80%

Question 8

What is overfitting?

A. A model performing well on new data but poorly on training data

B. A model memorizing training-specific patterns and generalizing poorly

C. A model using no data at all

D. A model being reviewed by a human

Correct Answer: B. A model memorizing training-specific patterns and generalizing poorly

Question 9

Which arrangement requires a person to approve or modify a system output before action?

A. Human-in-the-loop

B. Fully autonomous execution

C. Unsupervised clustering

D. Content-based filtering

Correct Answer: A. Human-in-the-loop

Question 10

Why should recommendation systems be evaluated using more than click-through rate?

A. Clicks are never measurable

B. Other outcomes such as diversity, complaints, retention, and trust may matter

C. Recommendations cannot be ranked

D. Click-through rate measures only image quality

Correct Answer: B. Other outcomes such as diversity, complaints, retention, and trust may matter