AI, Algorithms, Ethics and Trust

Introduction: Why AI Governance Matters in OTT and New Media

Artificial intelligence has become part of nearly every stage of the OTT and digital-media value chain. Streaming services use algorithms to recommend programmes, predict demand, schedule content, detect fraud, moderate user contributions, personalise advertising and optimise delivery. Generative AI is increasingly used for scripts, subtitles, dubbing, trailers, thumbnails, metadata, audience research and customer support.

These applications can improve discovery, reduce operating costs and help media organisations serve diverse audiences. They also create substantial risks. An algorithm may systematically disadvantage a language group, optimise sensational content, expose sensitive viewing data, reproduce stereotypes or make a false synthetic video appear authentic. A technically accurate prediction may still be ethically unacceptable if it is obtained through excessive surveillance or used without meaningful accountability.

Trust is therefore not simply a public-relations outcome. It is an organisational capability created through reliable systems, explainable decisions, fair treatment, privacy protection, editorial responsibility and effective remedy. Media managers must understand both what AI systems can do and where their use should be limited.

Learning Objectives

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

1. Explain recommendation systems, personalisation, predictive analytics and generative AI in media operations.
2. Distinguish major algorithmic approaches and select appropriate performance measures.
3. Evaluate AI applications in content development, marketing, audience engagement and operations.
4. Analyse bias, privacy, misinformation, deepfakes, safety and transparency risks.
5. Design a practical responsible-AI governance framework for an OTT or new-media organisation.
6. Make managerial decisions that balance commercial performance, editorial values, user rights and public trust.

1. The AI-enabled Media Value Chain

An OTT service can be understood as a connected system of editorial, technological, commercial and audience processes.

AI does not replace management. It changes the location of managerial judgement. Decisions once made by editors, programmers or customer-service teams may now be embedded in data definitions, model objectives, training examples, thresholds and interfaces. Consequently, governance must cover the entire lifecycle: problem definition, data collection, model development, deployment, monitoring, revision and retirement.

2. Recommendation Systems and Personalisation

A recommendation system estimates which items may be relevant to a user and presents selected items in an ordered interface. Its purpose is not necessarily to identify the objectively best content. Rather, it attempts to match content, context and user needs under uncertainty.

2.1 Main approaches

Content-based recommendation uses attributes of items and the user’s previous interests. If a viewer watches several crime dramas in a particular language, the system may identify similar genres, cast members, themes or production regions. It is relatively interpretable but may over-specialise and repeatedly recommend items similar to those already consumed.

Collaborative filtering uses patterns across users. If viewers with similar behaviour enjoyed a programme, that programme may be recommended to another viewer. Collaborative systems can discover unexpected connections but suffer from the cold-start problem: new users and new titles have little behavioural data.

Hybrid systems combine content, collaborative, contextual and editorial signals. Modern services may also use sequence models that consider the order and timing of interactions, as well as contextual variables such as device, time of day, subscription plan and language preference.

Knowledge-based recommendation uses explicit rules or domain knowledge. For example, a service may recommend age-appropriate educational programmes to a family profile. Such systems are useful when viewing decisions are infrequent or when safety constraints are important.

2.2 Ranking is a value decision

A recommendation interface is not neutral. The objective function determines what the organisation calls success. If the system optimises only immediate clicks or minutes watched, it may favour sensational, repetitive or emotionally extreme content. An alternative objective may include satisfaction, completion, diversity, long-term retention, accessibility and editorial priorities.

A simplified utility function can represent this idea:

Here, U is the system’s overall utility; E is engagement; S is predicted satisfaction; D is content or viewpoint diversity; R is retention or long-term value; and H represents expected harm. The coefficients α, β, γ, δ and λ are organisational weights. This is a conceptual model, not a universal industry formula. The variables must be operationalised, measured on comparable scales and validated for the specific service. A high numerical score does not prove that the objective is ethically appropriate.

A responsible ranking system may apply constraints rather than relying only on a weighted score. For instance, illegal content, content unsuitable for a child profile or material that violates platform policy may be excluded before ranking. Diversity rules may prevent a recommendation row from being filled with near-duplicates. Human editorial curation may override automated recommendations in public-interest or crisis contexts.

2.3 Measuring recommendation quality

Common measures include:

• Precision: the proportion of recommended items that are relevant among the items recommended.
• Recall: the proportion of relevant items that appear among the items recommended.
• Click-through rate: the proportion of impressions producing a click.
• Completion rate: the proportion of starts reaching a defined completion point.
• Calibration: whether predicted probabilities correspond to observed outcomes.
• Coverage: the proportion of the catalogue or user population receiving meaningful recommendations.
• Diversity: the degree to which recommendations differ from one another.
• Serendipity: the extent to which useful recommendations are unexpected.

These measures can conflict. Increasing precision may reduce coverage; maximising clicks may reduce satisfaction; increasing novelty may lower immediate completion. Managers should therefore use a balanced scorecard and conduct user research rather than declaring success from one metric.

A basic precision expression is:

The measure assumes that relevance has been defined consistently and that the evaluation sample is appropriate. It does not reveal whether the system excludes certain groups, promotes harmful material or performs differently across languages and devices.

3. Predictive Analytics in Media Management

Predictive analytics uses historical and current data to estimate future outcomes. OTT organisations may predict demand for a title, subscription cancellation, advertising response, content delivery load, customer-support volume or the probability of fraudulent account activity.

A churn model, for example, may use viewing decline, payment events, service complaints and changes in plan usage. The model produces a probability, not a fact. A manager must decide whether intervention is justified, what intervention is proportionate and how false positives will be handled.

3.1 Classification and threshold decisions

Many operational models classify cases, such as “likely to churn” or “requires review.” A threshold converts a probability into an action. A low threshold captures more potential cases but creates more false alarms; a high threshold reduces unnecessary intervention but misses some cases.

Performance should be assessed using a confusion matrix: true positives, false positives, true negatives and false negatives. Accuracy alone can be misleading when the event is rare. Precision, recall, specificity, calibration and group-level performance should be examined. For consequential decisions, the organisation should also assess whether errors are distributed unfairly across demographic, linguistic, geographic or accessibility groups.

3.2 Forecasting and uncertainty

Demand forecasts are affected by release timing, competing services, marketing, cultural events, platform availability and unexpected news. A forecast should therefore be accompanied by uncertainty and assumptions. Managers should use scenario ranges rather than presenting a single number as certainty.

Predictive analytics is especially vulnerable to historical bias. If a platform previously marketed premium content mainly to affluent urban audiences, a model trained on past conversions may learn to ignore other groups. The result may be statistically consistent with history but strategically and socially undesirable.

4. Generative AI in Content Development and Operations

Generative AI systems produce text, images, audio, video or code in response to prompts or other inputs. In media, applications include:

• first drafts of synopses, loglines and promotional copy;
• translation, subtitling, voice adaptation and accessibility descriptions;
• concept art, storyboarding and previsualisation;
• trailer versioning and thumbnail experimentation;
• synthetic datasets and software development;
• conversational customer support and internal knowledge search.

The principal managerial distinction is between assistive use and substitutive use. Assistive use supports a professional who remains responsible for judgement and approval. Substitutive use allows a system to produce or publish material with limited human review. The latter may be appropriate for low-risk routine tasks, but it demands stronger controls when outputs affect reputation, public information, employment, safety or cultural representation.

Generative systems may hallucinate: they can produce fluent but false claims, invented citations, incorrect translations or visually plausible details. They may reproduce copyrighted or private material from training or prompt context. They may also homogenise creative output by favouring dominant styles and high-frequency patterns.

A practical workflow includes:

1. Define the permitted purpose and prohibited uses.
2. Use approved tools and avoid entering confidential, personal or unreleased material into unauthorised services.
3. Label source material, prompts, model versions and significant edits.
4. Require human review for accuracy, rights, cultural sensitivity, safety and originality.
5. Obtain consent and contracts for voices, faces, likenesses and performances.
6. Preserve an audit trail and disclose synthetic or materially altered content where audiences could be misled.
7. Monitor complaints and correct errors after publication.

5. Bias, Fairness and Inclusion

Bias is not limited to prejudiced code. It can enter through problem definition, data collection, labelling, sampling, feature selection, model design, interface choices and institutional use.

Important forms include:

• Representation bias: some populations are missing or underrepresented in training data.
• Historical bias: data reflects unequal past treatment.
• Measurement bias: a proxy, such as clicks, is treated as a measure of quality or interest.
• Aggregation bias: one model is applied to groups with different needs.
• Evaluation bias: tests are conducted mainly in dominant languages or on technically convenient users.
• Deployment bias: a system is used in a context different from the one for which it was designed.

Fairness is multidimensional. Equal error rates, equal access, equal opportunity, appropriate representation and culturally respectful treatment may not all be achievable simultaneously. The correct choice depends on the use case and on whose interests are affected. A content-recommendation system may require diversity and accessibility rather than a narrow parity statistic.

Media organisations should conduct impact assessments before deployment, test results across relevant groups, consult affected communities and provide a route for appeal. Fairness testing must not become a substitute for addressing structural causes, such as inadequate local-language catalogues or exclusionary commissioning practices.

6. Privacy, Data Protection and Surveillance

Personalisation depends on data, but more data is not automatically better management. Viewing histories, search queries, device identifiers, location signals, inferred interests and household relationships can reveal sensitive information. Even when names are removed, combinations of data may permit re-identification.

Sound privacy practice follows several principles:

• Purpose limitation: collect data for specified legitimate purposes.
• Data minimisation: use only what is necessary.
• Transparency: explain collection, inference, sharing and retention in accessible language.
• User control: provide meaningful choices and withdrawal mechanisms.
• Security: protect data in storage, transit and access processes.
• Retention limitation: delete or anonymise data when it is no longer needed.
• Accountability: document decisions and demonstrate compliance.

Consent is not a complete answer. It may be bundled, poorly understood or impossible to refuse without losing access. Children and vulnerable users require heightened protection. Sensitive inferences should not be used for manipulative targeting or discriminatory pricing. Managers should distinguish first-party data from data acquired through brokers and evaluate whether the source is lawful, accurate and ethically acceptable.

7. Deepfakes, Misinformation and Content Authenticity

A deepfake is synthetic or manipulated audio-visual material that may falsely represent a person, event or statement. Generative systems reduce the cost of producing persuasive fabrications. Risks include impersonation, political manipulation, non-consensual intimate imagery, financial fraud, reputational harm and erosion of trust in authentic evidence.

Detection tools are useful but imperfect. They may fail as generation methods improve, produce false accusations or perform poorly across languages and compression formats. Provenance is therefore as important as detection. Provenance systems record how content was captured, edited and published; cryptographic signatures and content credentials can help audiences and platforms inspect that history. They do not by themselves prove that the underlying claim is true.

A media organisation should maintain a rapid-response protocol:

1. Preserve the original file, metadata and chain of custody.
2. Assess authenticity using technical, editorial and source-based evidence.
3. Avoid amplifying suspicious material unnecessarily.
4. Label uncertainty clearly while investigation continues.
5. Consult legal, safety and editorial leaders for high-impact cases.
6. Correct or remove content when policy and evidence require it.
7. Notify affected individuals and document the decision.

Misinformation governance must respect legitimate journalism, satire, artistic transformation and political expression. A blanket removal policy can itself damage trust. The relevant question is not merely whether content is false, but also its likely harm, intent, reach, context and possibility of correction.

8. Transparency, Explainability and Human Oversight

Transparency has at least three audiences. Users need to know why content is recommended or why an account action occurred. Employees need sufficient technical information to operate and challenge a system. Regulators and auditors need evidence of accountability, testing and remediation.

An explanation should be useful rather than merely technical. “Recommended because you watched two sports documentaries” may be more meaningful to a user than a description of a neural-network architecture. For an internal reviewer, however, the explanation may need feature importance, training-data documentation, model version, threshold settings and known limitations.

Human oversight should be substantive. A person who can only approve an automated output without time, authority or information is not meaningful oversight. Organisations should define:

• which decisions require human approval;
• who owns the decision;
• what evidence the reviewer sees;
• when escalation is mandatory;
• how users appeal;
• how corrections are recorded;
• when a model must be suspended.

Automation should be reversible. A rollback plan, manual fallback process and incident-response team are essential for high-impact systems.

9. Responsible-AI Governance Framework

A practical governance framework links principles to controls across the AI lifecycle.

Governance layers

Strategic layer: The board or senior leadership defines risk appetite, public commitments and accountability. AI objectives must align with editorial standards, privacy obligations and organisational values.

Portfolio layer: An AI register records each system, owner, purpose, data sources, affected groups, risk level, supplier and review date. High-impact applications receive stronger approval requirements.

Operational layer: Product, editorial, legal, security, data-science, accessibility and customer-support teams jointly review deployment. No single technical team should determine social acceptability alone.

Assurance layer: Independent testing evaluates performance, security, bias, privacy, robustness and explainability. Audits should examine real-world outcomes, not only laboratory scores.

Remedy layer: Users and workers require accessible complaint, appeal, correction and compensation mechanisms where appropriate.

A useful risk register includes likelihood, impact, affected stakeholders, controls, residual risk and accountable owner. Risk is dynamic: a low-risk transcription tool may become high risk if it is later used for disciplinary decisions or public-news publication.

10. Illustrative Case: StreamSphere’s Personalised News and Entertainment Service

This is a fictional illustrative case.

Background

StreamSphere operates an OTT service across three countries. Its new recommendation engine combines viewing history, search behaviour, language, device and time of day. Management hopes to increase discovery and reduce cancellations. The service also uses generative AI to create summaries and thumbnails.

Problem

After launch, engagement rises, but complaints increase. Users report seeing repetitive sensational content. Local-language programmes receive less visibility. A generated summary incorrectly attributes a political statement to a public figure. A child using a family account receives recommendations unsuitable for the profile because household behaviour was combined into one user representation.

Analysis

The system was optimised mainly for clicks and short-term viewing. Its training data overrepresented the largest market. The team measured average engagement but not language-level coverage, diversity, child-safety outcomes or summary accuracy. Consent notices did not clearly explain inferred interests. Human reviewers had no formal escalation process.

Approach and solution

StreamSphere creates an AI governance committee including editorial, product, legal, privacy, accessibility and trust-and-safety representatives. It pauses automated political summaries and requires source verification. It separates child and adult profiles, reduces unnecessary data collection and adds user controls for recommendation explanations and preference resetting.

The recommendation objective is revised to include long-term satisfaction, catalogue coverage, language representation, age suitability and harm constraints. A/B tests are supplemented with qualitative interviews and subgroup analysis. Generative outputs receive provenance labels internally, and public disclosure is introduced when synthetic alteration could mislead audiences. A model card records purpose, data, limitations, evaluation results and approved uses.

Outcome

Short-term click-through declines slightly, but complaints fall, completion of locally produced content improves and customer surveys show greater confidence in recommendations. The company also discovers that some apparent “low-demand” programmes had been poorly surfaced rather than genuinely unpopular.

Learning points

1. Engagement is an incomplete measure of media value.
2. Historical data can conceal supply and visibility problems.
3. Personalisation requires profile, privacy and child-safety design.
4. Generative tools need editorial verification and provenance controls.
5. Trust improves when users can understand, challenge and adjust automated decisions.

11. Contemporary Developments and Professional Practice

The field is moving toward multimodal systems, conversational discovery, automated dubbing, synthetic presenters, agentic software and increasingly personalised advertising. These developments create opportunities for accessibility and global distribution, but also intensify rights, labour and accountability questions.

Media managers should develop the following professional capabilities:

• data literacy sufficient to question datasets, metrics and uncertainty;
• editorial judgement about accuracy, context and public interest;
• knowledge of privacy, intellectual property, contracts and platform regulation;
• ability to conduct stakeholder and impact assessments;
• competence in incident response and vendor management;
• communication skills for explaining automated systems to non-technical audiences;
• commitment to continuous monitoring rather than one-time compliance.

Vendor contracts should specify data use, retention, model training, security, audit access, incident notification, intellectual-property responsibilities, service continuity and deletion. “The supplier’s model is proprietary” should not prevent the buyer from understanding risks or obtaining evidence of performance.

Challenges and Limitations

Responsible AI has practical limits. Fairness definitions may conflict, explainability may be incomplete, audits can be expensive and small-language datasets may be scarce. Privacy-preserving methods may reduce analytical utility. Human review can introduce its own inconsistency and bias. Labels such as “AI-generated” may be misunderstood or ignored. Regulation differs across jurisdictions and evolves rapidly.

There is also a political-economy problem: organisations may claim that personalisation serves users while primarily optimising advertising revenue or market power. Ethical governance must therefore examine incentives, ownership, labour effects and concentration, not only technical performance.

The appropriate response is not to reject all automation or to treat innovation as automatically beneficial. It is to match controls to risk, document trade-offs, involve affected communities and preserve meaningful human responsibility.

Chapter Summary

AI is transforming OTT and new-media management through recommendation, personalisation, predictive analytics, generative production, marketing optimisation and trust-and-safety operations. Recommendation systems may use content, collaborative, knowledge-based or hybrid methods, but their objectives reflect organisational values. Engagement metrics must be balanced with satisfaction, diversity, coverage, accessibility, safety and long-term trust.

Predictive analytics produces uncertain estimates and can reproduce historical inequalities. Generative AI offers efficiency and creative assistance but introduces hallucination, rights, privacy, labour and authenticity risks. Bias can arise at every stage from problem definition to deployment. Privacy requires purpose limitation, minimisation, transparency, security and meaningful user control. Deepfakes and misinformation require a combination of provenance, editorial verification, proportionate moderation and rapid response.

Responsible-AI governance connects strategy, risk registers, cross-functional review, testing, documentation, human oversight, appeals and remediation. Trust is earned when media organisations are reliable, transparent, fair, privacy-conscious and willing to correct mistakes. The central managerial task is to use AI to expand media value without allowing opaque optimisation to displace editorial responsibility or human rights.

Key Terms

Artificial intelligence: Computational systems performing tasks associated with human intelligence, such as prediction, classification or generation.

Algorithm: A defined procedure for processing inputs to produce outputs or decisions.

Recommendation system: A system that selects or ranks content likely to be relevant to a user.

Personalisation: Adaptation of content, interface or communication to an individual or group.

Collaborative filtering: Recommendation based on behavioural similarities among users or items.

Content-based recommendation: Recommendation based on item attributes and observed user interests.

Predictive analytics: Use of data and statistical or machine-learning methods to estimate future outcomes.

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

Hallucination: A plausible-looking but inaccurate or unsupported generative-AI output.

Bias: A systematic tendency that produces unfair, inaccurate or unrepresentative outcomes.

Fairness: The ethically and contextually appropriate treatment of affected people or groups.

Personal data: Information relating to an identified or identifiable person.

Data minimisation: Limiting collection and use to what is necessary for a specified purpose.

Deepfake: Synthetic or manipulated media that can falsely represent a person, event or statement.

Provenance: Information about the origin, history and alterations of a piece of content.

Explainability: The degree to which an AI output or process can be understood by relevant people.

Human oversight: Meaningful human authority to review, challenge, correct or stop an AI system.

Model card: Documentation describing a model’s purpose, data, performance, limitations and approved uses.

Responsible AI: Development and use of AI consistent with fairness, safety, privacy, accountability and human values.

Review Questions

Short Answer Questions

1. What is the difference between content-based and collaborative filtering?
2. Why can click-through rate be an inadequate measure of recommendation quality?
3. Identify three ways bias can enter a media-AI system.
4. What is a deepfake, and why is provenance useful?
5. State four principles of responsible data governance.

Long Answer Questions

1. Explain how an OTT service could design a hybrid recommendation system while preserving editorial and user control.
2. Discuss the opportunities and risks of generative AI in localisation, marketing and content development.
3. Analyse why historical predictive data may reproduce inequality even when a model is technically accurate.
4. Explain the relationship between transparency, explainability, human oversight and user trust.
5. Design a responsible-AI governance framework for a national streaming platform.

Analytical / Application Questions

1. An OTT platform increases watch time after optimising for emotionally intense trailers, but complaints and cancellations also rise. Diagnose the problem and recommend metrics, constraints and governance actions.
2. A churn model identifies a minority-language audience as unlikely to renew. Explain how you would test whether the result reflects genuine demand, poor catalogue visibility or discriminatory data practices.
3. A synthetic video falsely depicts a celebrity endorsing a financial product. Develop an immediate response plan and a longer-term prevention strategy covering technology, editorial policy, legal issues and audience communication.

Multiple Choice Questions with Answers

Question 1

Which recommendation approach primarily uses similarities in behaviour among users?

A. Content-based filtering

B. Collaborative filtering

C. Rule-free generation

D. Cryptographic provenance

Correct Answer: B. Collaborative filtering

Question 2

What is a major weakness of optimising an OTT recommendation system only for clicks?

A. It guarantees excessive catalogue diversity

B. It eliminates the need for editorial review

C. It may favour sensational content over satisfaction and safety

D. It prevents all personalisation

Correct Answer: C. It may favour sensational content over satisfaction and safety

Question 3

In the chapter’s conceptual utility model, what does H represent?

A. Historical viewing

B. Expected harm

C. Hardware capacity

D. Household size

Correct Answer: B. Expected harm

Question 4

Which measure asks whether recommended items cover a broad portion of the catalogue or user population?

A. Coverage

B. Precision

C. Specificity

D. Latency

Correct Answer: A. Coverage

Question 5

Why can accuracy be misleading when evaluating a churn classifier?

A. Accuracy cannot be calculated for classifiers

B. Rare events can make a model appear accurate while missing most important cases

C. Accuracy measures only data-storage speed

D. Accuracy always penalises true negatives

Correct Answer: B. Rare events can make a model appear accurate while missing most important cases

Question 6

Which is an example of representation bias?

A. A model is updated too frequently

B. A training dataset contains very little material from a language community

C. A user changes a profile password

D. A platform labels synthetic content

Correct Answer: B. A training dataset contains very little material from a language community

Question 7

What does data minimisation require?

A. Collecting every available behavioural signal

B. Using only data necessary for a specified legitimate purpose

C. Deleting all data immediately after collection

D. Sharing data with all suppliers

Correct Answer: B. Using only data necessary for a specified legitimate purpose

Question 8

Why is human oversight meaningful only when reviewers have authority and information?

A. Because reviewers must be able to challenge, correct or stop decisions

B. Because humans are always more accurate than models

C. Because automated systems cannot process data

D. Because oversight removes the need for documentation

Correct Answer: A. Because reviewers must be able to challenge, correct or stop decisions

Question 9

Which control is most directly associated with content authenticity?

A. Increasing thumbnail brightness

B. Reducing subscription prices

C. Recording content provenance and editing history

D. Maximising autoplay duration

Correct Answer: C. Recording content provenance and editing history

Question 10

What should a responsible AI register normally record?

A. Only the model’s marketing slogan

B. System purpose, owner, data sources, risks and review information

C. Only the number of software engineers

D. Users’ private passwords

Correct Answer: B. System purpose, owner, data sources, risks and review information