COMP 3270 / Module 3 / Unit 7 / Lesson 1

Lesson 7.1 · Week 5 · Reading: Chapter 9

Bias, Representation, and Default Assumptions

Treat bias as a sociotechnical issue that emerges across the entire AI-supported workflow—not merely inside a model. Distinguish six types of bias, examine representation and omission, and surface the default assumptions embedded in prompts, sources, and human interpretation.

Estimated study time: 120–150 min Prerequisites: Units 1–6; PCAFE-V, SCOPE, PARTNER, ORBIT Module 3 · Week 5

Table of Contents

1. Lesson Overview

This lesson introduces bias as a sociotechnical issue that emerges across the entire AI-supported workflow, not merely inside a model. You will learn to distinguish multiple types of bias—data, model, interaction, interpretation, workflow, and institutional—and to examine how representation, omission, stereotypes, and default assumptions shape AI-generated content and human decisions.

You will move from identifying bias entry points to conducting a representation audit and redesigning prompts or review processes to reduce unsupported assumptions. The lesson prepares you for Lesson 7.2 (fairness, equity, accessibility, and intersectionality) and provides foundational evidence for Assignment 3 (Responsible AI Case Analysis).

Why this matters: Generative AI systems are trained on human-created data and used in human-centred contexts. Without explicit attention to bias, these systems can reproduce, amplify, or introduce unequal effects—even when the designers had no discriminatory intent. Learning to detect and mitigate bias at every stage is essential for responsible AI use in any discipline.

2. Learning Outcomes

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

  1. Define bias in a sociotechnical context and distinguish it from deliberate prejudice.
  2. Distinguish data, model, interaction, interpretation, workflow, and institutional bias.
  3. Identify representation and underrepresentation in AI-generated content.
  4. Recognize stereotypes, default assumptions, omissions, and unexamined assumptions in prompts, sources, outputs, and human decisions.
  5. Examine how prompts, source selection, AI outputs, and human interpretation interact to produce or amplify bias.
  6. Apply a bias-entry-point map to a bounded AI-supported workflow.
  7. Conduct a preliminary representation audit and recommend mitigations.
  8. Justify a preliminary recommendation concerning whether an AI-supported use should proceed, be modified, or require further review.

3. Before You Begin

4. Retrieval Practice: Concepts from Units 1–6

These questions draw on Units 1–6. Answer each before reading the feedback.

Question 1: What does the V in PCAFE-V represent, and why is it important when evaluating an AI-supported workflow?

Show answer

The V stands for Verification. It emphasises that AI-generated outputs must be checked against authoritative sources, human expertise, and contextual evidence. Verification is essential because AI systems can produce plausible but incorrect information, and reliance on unverified outputs can lead to errors, bias, or harm.

Question 2: List the five elements of the SCOPE framework and briefly state what each addresses.

Show answer
  • S — Sources: Identify the origin, quality, and relevance of information used.
  • C — Context: Understand the situation, audience, and purpose.
  • O — Omissions: Recognise what is missing or underrepresented.
  • P — Purpose: Clarify the legitimate human goal and intended use.
  • E — Evidence: Evaluate the strength and limitations of available evidence.

Question 3: In the PARTNER framework, what role does human review play in an AI-supported workflow, and why is it not a guarantee of fairness?

Show answer

Human review provides oversight, contextual judgment, and accountability. However, it is not a guarantee of fairness because reviewers may have their own biases, lack relevant expertise, be influenced by automation bias (deferring to AI outputs), or operate under time and resource constraints. Effective human review requires clear criteria, training, and contestability pathways.

Question 4: Define automation bias and give one example of how it could affect an AI-supported decision.

Show answer

Automation bias is the tendency to over-rely on automated system outputs, even when contrary evidence exists or the system is known to have limitations. Example: A reviewer accepts an AI-generated performance ranking without questioning its methodology, missing a systematic error that disadvantages certain candidates.

5. Opening Scenario: "AI-Generated Role-Model Examples"

📌 Scenario — AI-Generated Role-Model Examples

A university is developing an AI-powered tool to help students explore career paths. The system generates brief profiles of professionals in various fields, drawn from a curated set of public biographies. The tool is intended to promote diverse role models and broaden students' awareness of possible careers.

During a pilot test, a student notices that the AI consistently generates profiles with the following patterns:

  • Leadership and executive roles are predominantly associated with masculine-coded language and names.
  • Caregiving, teaching, and support roles are predominantly associated with feminine-coded language and names.
  • Technical and engineering roles are dominated by profiles that assume a Western educational background.
  • Profiles for creative fields (e.g., design, writing) rarely include people with visible disabilities or from non-urban settings.
  • Indigenous professionals are absent from all profiles, even when the field has notable Indigenous practitioners.

The tool's developers note that the source biographies were chosen to represent a wide range of fields, but they did not systematically audit for demographic diversity. They also observe that the AI model may amplify patterns present in the source text.

Questions to consider:

  1. What patterns appear in the AI-generated profiles, and why might they matter?
  2. Who is absent, and what might be the effect of that absence?
  3. What alternative explanations could account for the observed patterns?
  4. What evidence would be needed before drawing a broader conclusion about bias in this tool?
Show guided analysis

1. Patterns: The AI reproduces gender- and culture-based role stereotypes. Leadership is coded masculine; care and support are coded feminine; technical fields assume Western, urban, non-disabled identities; Indigenous professionals are entirely omitted. These patterns matter because they can shape students' aspirations, reinforce stereotypes, and marginalise already-underrepresented groups.

2. Absence: Indigenous professionals, people with disabilities, and non-urban professionals are missing. Omission communicates that these groups either do not exist in certain fields or are not worthy of mention—a representational harm.

3. Alternative explanations: The patterns may reflect historical bias in the source biographies (i.e., the real-world underrepresentation of certain groups in certain fields), selection bias in how sources were chosen, or model bias in how the AI generalises from training data. The source curation process may have inadvertently amplified existing imbalances.

4. Evidence needed: A systematic audit of the source biographies (gender, culture, disability, geography, field distribution); comparison with a broader, more diverse corpus; testing the model with varied prompts; and consultation with affected communities to understand the real-world impact of the tool.

This scenario illustrates that bias is not necessarily intentional—it can arise from historical patterns, data selection, and model behaviour. Detecting it requires careful examination of sources, outputs, and effects.

6. What Is Bias in a Sociotechnical Context?

Definition — Sociotechnical Bias

Bias in an AI-supported workflow refers to systematic, non-random deviations from fairness, accuracy, or representativeness that arise from the interplay of human decisions, data, algorithms, interfaces, and institutional structures. Bias is not merely a property of the model; it emerges at every stage—from problem definition to remedy.

Commonly, bias is mistaken for deliberate prejudice or individual malice. In sociotechnical systems, however, bias often results from:

Understanding bias as a sociotechnical issue moves us away from a narrow focus on the model and toward a workflow-wide view. This is crucial because a model that appears neutral in isolation can still produce unequal effects when deployed in a particular context.

Key distinction: Bias ≠ prejudice. Prejudice is a conscious or unconscious attitude held by an individual. Bias is a systematic pattern that can be observed in data, processes, and outcomes—regardless of intent. A system can be biased even if every person involved is well-intentioned.
Section summary
  • Bias is a sociotechnical issue that emerges across the workflow, not just in the model.
  • It is systematic and non-random, and it can occur without any individual's intent.
  • Detecting bias requires examining sources, design, interaction, and institutions.

7. Six Types of Bias

Bias can enter an AI workflow at many points. The table below distinguishes six major categories.

Table 1. Six types of bias in AI-supported workflows (based on Chapter 9)
Type Description Example
Data bias Bias in the training data, including historical, representation, selection, measurement, and label bias. A dataset of engineering profiles contains mostly men, so the AI generates mostly male profiles for engineering roles.
Model bias Bias introduced by model architecture, training process, or optimisation choices. A language model performs better on standardised, formal English than on regional dialects or non-native grammar, producing lower-quality responses for those users.
Interaction bias Bias that arises from how users interact with the system—prompts, feedback, and behaviour. Users who ask vague, leading, or stereotyped prompts receive outputs that reflect those assumptions, which are then used without critical review.
Interpretation bias Bias in how humans understand, evaluate, and act on AI outputs. A reviewer treats an AI-generated candidate ranking as objective, overlooking the model's reliance on proxy variables (e.g., school name).
Workflow bias Bias embedded in the design of the overall process—who is consulted, what is measured, how errors are handled. An AI-supported student-support system provides identical responses to all students, but no alternative is offered for students with low connectivity or visual impairments.
Institutional bias Bias in the policies, norms, and power structures that govern how AI is developed, deployed, and overseen. An organisation adopts an AI tool for hiring but does not require bias audits or provide a contestability pathway for rejected applicants.
🧩 Quick check

Review the opening scenario (Section 5). Which types of bias are likely at play? Jot down your thoughts, then expand the feedback below.

Show possible classification
  • Data bias — historical and representation bias in the source biographies (gender, cultural, disability, Indigenous underrepresentation).
  • Model bias — the AI may amplify existing patterns in the training data.
  • Interpretation bias — developers and students may interpret the profiles as representative if they do not critically examine them.
  • Workflow bias — the source-selection process did not include a diversity audit, and the tool lacks a mechanism for user feedback or correction.
  • Institutional bias — the university may not have policies requiring representation audits for AI-supported career tools.
Section summary
  • Six types of bias can enter an AI-supported workflow: data, model, interaction, interpretation, workflow, and institutional.
  • Each type requires different interventions.
  • A comprehensive fairness review examines all six.

8. Representation and Underrepresentation

Representation refers to the presence and prominence of different groups, perspectives, and experiences in AI-generated content. Underrepresentation occurs when certain groups appear less frequently, in less prominent roles, or not at all.

When examining representation, consider:

Representation is not merely a matter of "counting"—it also involves quality, prominence, and agency. A group may appear frequently but only in subordinate or stereotyped roles; that is not meaningful representation.

Caution: Representation does not automatically create inclusion. A system can include diverse faces in its outputs while still excluding those groups from decision-making, benefit, or remedy. Inclusion requires participation, influence, and accountability, not just visibility.
Section summary
  • Representation is about presence, prominence, and quality—not only counts.
  • Underrepresentation and omission can be as harmful as direct misrepresentation.
  • Representation is a necessary but not sufficient condition for inclusion.

9. Stereotypes, Default Assumptions, and Omissions

Stereotypes are oversimplified, generalised beliefs about a group that are often inaccurate and harmful. In AI-generated content, stereotypes can appear as:

Default assumptions are unexamined beliefs that shape how systems are designed and how outputs are interpreted. Common default assumptions include:

Omissions are absences that matter—missing data, missing perspectives, missing stakeholders, or missing pathways for correction. An omission can be as consequential as a direct misrepresentation.

Practical implication: When you encounter an AI-generated output, ask: Whose experience is treated as the default? Whose experience is exceptional or absent? This helps surface unexamined assumptions.
Section summary
  • Stereotypes are oversimplified generalisations that appear in AI content as role associations, deficit framing, and homogenisation.
  • Default assumptions encode unexamined beliefs about users and contexts.
  • Omissions can be as harmful as misrepresentations.

10. How Prompts, Sources, Outputs, and Human Decisions Interact

Bias is rarely caused by a single factor. Instead, it emerges through interactions across the workflow. A simplified cycle:

  1. Sources — the data and texts used to train or inform the AI reflect historical and cultural patterns.
  2. Prompt — a user poses a question or request; the phrasing, examples, and framing carry assumptions.
  3. Model / retrieval — the AI processes the prompt, drawing on its training and any retrieved information.
  4. Output — the AI generates text, images, or recommendations.
  5. Human interpretation — a person reads, evaluates, and acts on the output, bringing their own assumptions and context.
  6. Decision / action — the interpretation informs a decision, which may affect stakeholders.
  7. Feedback / monitoring — the system may be adjusted based on outcomes, but if monitoring is absent or biased, patterns persist.

At each stage, bias can be introduced, amplified, or mitigated. A biased source can be partially corrected by a well-designed prompt; an unbiased source can be distorted by a leading prompt; a correct output can be misinterpreted by a biased reviewer; a biased output can be challenged if there is a remedy pathway.

Figure 1 (placeholder): [INSERT BIAS ENTRY POINTS DIAGRAM — AI workflow with indicated bias points across sources, prompt, model, output, interpretation, decision, and monitoring].
🧩 Application: Trace a bias pathway

Using the opening scenario, trace one instance of bias from sources through to interpretation. Identify where the bias was introduced and where it could have been caught.

Show example pathway
  • Sources: The curated biographies over-represent men in leadership and Western professionals in technical fields (data bias / selection bias).
  • Prompt: A user asks for "successful professionals in engineering" — the AI draws on the biased sources and generates male-dominated, Western-coded profiles (model bias + interaction bias).
  • Output: The profiles reinforce stereotypes about who belongs in engineering.
  • Interpretation: A student reads the profiles and assumes that engineering is predominantly male and Western (interpretation bias).
  • Decision: The student chooses a different career path, narrowing their aspirations.
  • Monitoring: The tool has no feedback mechanism, so the pattern continues unchecked (workflow bias / institutional bias).

Each stage presents an opportunity for intervention: auditing sources, redesigning prompts, evaluating outputs, training interpreters, and building feedback loops.

Section summary
  • Bias emerges through interactions across the workflow, not from a single factor.
  • Every stage is both a potential bias entry point and a potential intervention point.
  • Tracing a bias pathway helps identify where to intervene most effectively.

11. Why Bias Is Not Solely a Property of the Model

A common misconception is that if a model is "fair" in some technical sense, the overall system is fair. This is misleading because:

Thus, a fairness evaluation must examine the entire workflow—from problem definition to remedy—not just the model.

Core principle: Fairness is a property of the sociotechnical system, not the AI model alone. A model can be mathematically fair according to one metric and still produce unjust outcomes in practice because of interaction, interpretation, or institutional factors.
Section summary
  • Bias exists in data, context, and interpretation—not just the model.
  • A fair model is not sufficient for a fair workflow.
  • Comprehensive evaluation must cover the entire sociotechnical system.

12. Process Walkthrough: Purpose to Sources to Prompt to Representation and Interpretation

This walkthrough traces a complete bias-aware workflow for the career-profile tool introduced in the opening scenario. It shows how to examine each stage for bias.

Table 2. Bias-aware workflow walkthrough for the career-profile tool
Stage Action Bias check Potential intervention
1. Purpose Define the human goal: broaden career awareness for students. Is the purpose legitimate? Does it serve all students equitably? Articulate an inclusion goal: "Support students from all backgrounds to see themselves in diverse careers."
2. Sources Select biographies and profile sources. Are sources diverse? Do they reflect the full range of fields and identities? What is omitted? Conduct a source diversity audit; include biographies from Indigenous, disabled, and non-Western professionals; verify representation across fields.
3. Prompt design Design the user prompt (e.g., "Show me a professional in X field"). Does the prompt assume a dominant language, culture, or background? Does it limit responses? Test prompts with varied wording; include examples that explicitly invite diversity; allow users to specify interests or identities.
4. Model / retrieval AI generates profiles based on sources and prompt. Does the model amplify stereotypes? Does it produce homogenised outputs? Evaluate outputs for representation, roles, and language; adjust retrieval parameters to prioritise diverse sources.
5. Output review Review the generated profiles before deployment. Are stereotypes present? Are omissions evident? Are role assignments balanced? Implement a human-in-the-loop review with diversity criteria; document findings.
6. Human interpretation Students read and react to the profiles. How might students' own backgrounds affect their interpretation? Could they internalise stereotypes? Provide guidance on critical consumption of AI content; include reflective questions about representation.
7. Decision Students make career-related choices based on the tool. Does the tool shape choices in ways that narrow opportunities for some groups? Monitor career-related outcomes (anonymised); ensure non-AI pathways are available.
8. Monitoring Track usage, feedback, and outcomes. Are complaints or disparities emerging? Are certain groups using the tool less or benefiting less? Implement regular fairness audits; collect user feedback; establish a contestability process.
9. Remedy Provide correction and redress where needed. Is there a pathway for users to report issues, correct misinformation, or request human review? Build a feedback and correction mechanism; designate a responsible human owner.
Takeaway: Every stage of the workflow is a potential bias entry point—and a potential intervention point. A systematic, stage-by-stage review is more effective than focusing only on the model.

13. Worked Example: AI-Supported Course-Materials Recommendation

Context: A university uses an AI system to recommend supplementary readings for a large introductory course. The system draws from an institutional repository and generates a list of five recommended articles for each module.

Concern: The recommendations may reflect historical patterns in the repository—for example, over-representation of authors from certain institutions, under-representation of Indigenous and minority scholars, and a narrow range of methodological approaches.

Worked Analysis

  1. Legitimate purpose: To help students identify relevant supplementary readings efficiently.
  2. AI role: Recommender system with retrieval-augmented generation.
  3. Sources: Institutional repository of course materials and open-access articles.
  4. Stakeholders: Students, instructors, librarians, authors, and scholars whose work is included or excluded.
  5. Bias entry points:
    • Data bias: The repository may contain historical imbalances in authorship and topics.
    • Interaction bias: Students may phrase queries in ways that limit retrieval (e.g., using only Western keywords).
    • Interpretation bias: Instructors may trust the recommendation list as comprehensive.
    • Workflow bias: No regular diversity audit of the repository or the recommendation algorithm.
  6. Representation analysis:
    • Present: Many articles from major Western universities, established journals.
    • Absent: Articles from Indigenous scholars, scholars in the Global South, and early-career researchers.
    • Stereotyped roles: The system may recommend certain authors only for topics that align with stereotypes (e.g., Indigenous scholars only for "Indigenous issues" rather than core disciplinary topics).
  7. Default assumptions:
    • All recommended articles are accessible to students (language, format, paywall).
    • Students have the background knowledge to assess the recommendations.
    • The institutional repository is comprehensive and representative.
  8. Mitigations:
    • Expand the source repository with intentional diversity criteria.
    • Design prompts that explicitly invite diverse perspectives (e.g., "include at least one article from a scholar based outside North America").
    • Implement a human review process where instructors can flag omissions or request broader recommendations.
    • Provide students with guidance on evaluating recommendation lists for diversity.
    • Establish a feedback mechanism for authors and students to suggest additions.
  9. Residual uncertainty:
    • Even with a diverse repository, the model may still amplify certain patterns.
    • Student preferences and search behaviour may introduce new biases.
    • Assessing representation requires ongoing, context-sensitive evaluation.
  10. Recommendation: Proceed with safeguards (expanded source audit, human review, user feedback, and regular monitoring). Do not treat the AI recommendations as definitive; always complement with instructor-curated lists.
Show key takeaways from this worked example
  • Bias can enter through the source repository, the query phrasing, the model's retrieval, and the human interpretation of outputs.
  • Representation requires attention to presence, absence, roles, and language.
  • Default assumptions about student background, access, and knowledge can be as problematic as explicit stereotypes.
  • Mitigations are not one-time fixes; they require ongoing monitoring and community input.
  • A justified recommendation acknowledges residual uncertainty and includes conditions for proceeding.

14. Guided Learning Activities

🧩 Activity 7.1.A: Bias Entry-Point Map
  • Purpose: To systematically identify where bias can enter an AI-supported workflow.
  • Learning outcome: Distinguish multiple bias entry points and explain how they interact.
  • Estimated time: 20–25 minutes.
  • Materials: The opening scenario or a case of your choice.
  • Instructions: For each of the ten stages below, describe a possible bias entry point and a potential intervention. Use a scenario you have encountered or the opening scenario.
  1. Problem definition
  2. Data or sources
  3. Prompt design
  4. Context
  5. Model output
  6. Retrieval
  7. Human interpretation
  8. Decision
  9. Monitoring
  10. Remedy
Show example map (based on opening scenario)
  1. Problem definition: Defining "successful career" in narrow, Western terms. Intervention: Broaden the definition and include community-defined success.
  2. Data / sources: Biographies over-represent certain groups. Intervention: Conduct a diversity audit and expand sources.
  3. Prompt: Default prompts do not invite diversity. Intervention: Design prompts that explicitly request varied perspectives.
  4. Context: The tool is deployed without considering students' diverse backgrounds. Intervention: Contextualise the tool and provide guidance.
  5. Model output: The model generates stereotyped profiles. Intervention: Evaluate outputs for representation and role stereotypes.
  6. Retrieval: Retrieval prioritises certain sources. Intervention: Adjust retrieval parameters to prioritise diversity.
  7. Human interpretation: Students internalise stereotypes. Intervention: Teach critical AI literacy.
  8. Decision: Students narrow career choices based on biased outputs. Intervention: Offer human-supported career guidance.
  9. Monitoring: No feedback mechanism. Intervention: Build a user-feedback and correction process.
  10. Remedy: No pathway to report omissions or stereotypes. Intervention: Design a contestability and remedy pathway.
🧩 Activity 7.1.B: Representation Audit
  • Purpose: To practise examining AI-generated content for representation, absence, and stereotypes.
  • Learning outcome: Identify representation, underrepresentation, and omission.
  • Estimated time: 25–30 minutes.
  • Materials: Provided synthetic output or a live AI system (with synthetic inputs only).
  • Instructions: Examine the synthetic outputs below (or generate your own) and answer the audit questions.

Synthetic output (excerpt from career profiles):

  • "Dr. James K. is a distinguished software engineer at a leading tech firm in Silicon Valley. He holds a PhD in Computer Science from Stanford and has published over 50 papers."
  • "Maria S. is a dedicated elementary school teacher in a suburban district. She has a master's degree in education and is known for her creative lesson plans."
  • "Dr. Aisha R. is a public health researcher working on global health initiatives. She is based in Geneva and collaborates with international organisations."
  • "John M. is a construction project manager with 20 years of experience in commercial building projects."

Audit questions:

  • Who appears? What fields, roles, and backgrounds are represented?
  • Who is absent? Which fields, identities, or perspectives are missing?
  • Are there stereotypes in role assignment (e.g., gender, culture, profession)?
  • Is language homogenised (e.g., all profiles use a formal, Western academic style)?
  • What default assumptions are embedded in the profiles?
Show sample audit findings
  • Who appears: Profiles include engineers, teachers, public health researchers, and construction managers. All are presented as successful and accomplished.
  • Who is absent: No Indigenous professionals, no profiles with visible disabilities, no non-binary or gender-diverse individuals, no profiles from non-urban or non-Western settings (except Geneva, which is still Western-affiliated).
  • Stereotypes: Engineering is coded male (James), teaching is coded female (Maria). Public health is international but still Western-centric (Geneva). Construction is male. There is an implicit gender-role stereotype.
  • Homogenised language: All profiles use a formal, credential-focused style that assumes a Western educational and professional norm.
  • Default assumptions: Success is measured by prestigious institutions, publications, and international recognition. The profiles assume readers value these markers and have the background to interpret them.
🧩 Activity 7.1.C: Default-Assumption Rewrite
  • Purpose: To practise revising prompts or outputs to remove unsupported assumptions while retaining relevant context.
  • Learning outcome: Recognise and revise default assumptions.
  • Estimated time: 15–20 minutes.
  • Instructions: Choose one of the synthetic profiles from Activity 7.1.B. Identify two default assumptions embedded in it. Then rewrite the profile to be more inclusive without erasing the professional context.
Show sample rewrite

Original: "Dr. James K. is a distinguished software engineer at a leading tech firm in Silicon Valley. He holds a PhD in Computer Science from Stanford and has published over 50 papers."

Identified assumptions:

  • Software engineering success requires a PhD from a prestigious university.
  • Leading tech firms are located in Silicon Valley.
  • Publication count is a primary measure of success.

Rewrite: "Dr. James K. is a software engineer whose work focuses on accessible design and community-centred technology. He has collaborated with diverse teams across North America and has contributed to open-source projects used by non-profit organisations. His background includes both formal training and self-taught expertise."

This rewrite retains the professional identity but broadens the indicators of success, shifts the geographic focus, and emphasises community impact.

🧩 Activity 7.1.D: Bias Type Classification
  • Purpose: To classify examples of bias into types.
  • Learning outcome: Distinguish data, model, interaction, interpretation, workflow, and institutional bias.
  • Estimated time: 15–20 minutes.
  • Instructions: For each synthetic example below, classify the primary type(s) of bias at play. Explain your reasoning.
  1. Example A: An AI transcription tool consistently performs worse for speakers with non-North American accents, leading to lower-quality captions for international students.
  2. Example B: A company uses an AI resume-screening tool that was trained on historical hiring data. The tool favours candidates who attended a small set of elite universities.
  3. Example C: A manager uses an AI-generated performance score as the primary basis for promotions, even when team members raise concerns about the AI's criteria.
  4. Example D: A university's AI-powered chatbot provides different levels of detail depending on the language used in the query, with more comprehensive responses in English than in other languages.
  5. Example E: An organisation adopts an AI tool for hiring but does not require bias audits, provide a contestability process, or train reviewers on automation bias.
Show sample classifications
  • Example A: Model bias (performance variation across accents) and data bias (training data may be dominated by certain accents).
  • Example B: Data bias (historical bias in training data) and model bias (the model amplifies historical patterns).
  • Example C: Interpretation bias (the manager over-relies on the AI) and automation bias (deference to the system).
  • Example D: Model bias (differential performance by language) and interaction bias (user language shapes the quality of response).
  • Example E: Institutional bias (policies and practices that allow bias to go unchecked) and workflow bias (no monitoring or contestability).

Many examples involve multiple types of bias, reflecting the sociotechnical nature of the issue.

15. Exercises

📝 Exercise 7.1.1: Identify Bias Entry Points
  • Difficulty: Introductory
  • Estimated time: ~10–12 min
  • Learning outcome: Identify bias entry points in an AI workflow.

Scenario: A public library uses an AI chatbot to answer patron questions. The chatbot was trained on the library's historical FAQ data and is deployed without any explicit fairness testing.

Task: List at least four distinct stages in the workflow where bias could enter, and for each, describe one possible bias.

Show sample solution
  1. Data / sources: The FAQ data may reflect historical patterns—e.g., questions from certain communities may be over-represented or under-represented.
  2. Prompt: Users with limited literacy or non-native language skills may phrase questions in ways the chatbot misinterprets.
  3. Model output: The chatbot may generate longer, more detailed responses for certain topics and shorter, less helpful responses for others.
  4. Human interpretation: Librarians may rely on the chatbot's answers without verifying them, especially for routine questions.
  5. Institutional: The library may not have a process for users to report incorrect or biased responses.
📝 Exercise 7.1.2: Conduct a Representation Audit
  • Difficulty: Intermediate
  • Estimated time: ~15–18 min
  • Learning outcome: Conduct a representation audit of synthetic content.

Materials: The synthetic profiles from Activity 7.1.B.

Task: Write a brief audit report (100–150 words) that identifies patterns of representation, underrepresentation, stereotyping, and homogenisation. Include one recommendation for improving representation.

Show sample audit report

Audit report: The profiles represent a narrow range of professions and backgrounds. Men appear in engineering and construction; women appear in teaching and public health. Geographic diversity is limited to North America and Europe. All profiles use a formal, credential-based narrative that assumes a Western educational framework. Indigenous professionals, people with disabilities, and non-binary individuals are absent. The language is homogenised and does not reflect the varied paths that lead to professional success.

Recommendation: Expand the source material to include profiles from non-Western settings, Indigenous professionals, and individuals with diverse educational and career paths. Include biographies that highlight community impact, mentorship, and collaborative work alongside individual achievement.

📝 Exercise 7.1.3: Identify Default Assumptions
  • Difficulty: Intermediate
  • Estimated time: ~10–12 min
  • Learning outcome: Recognise default assumptions in AI-generated content.

Scenario: An AI-powered study-guide generator produces the following text: "Most students will find it helpful to review the textbook chapters before attending the lecture. Use the library's online resources to access additional readings."

Task: Identify at least three default assumptions embedded in this text. For each, explain why it might not apply to all students.

Show sample analysis
  • Assumption 1: Students have a textbook. Why it may not apply: Some students may not be able to afford textbooks or may use alternative editions.
  • Assumption 2: Students attend lectures. Why it may not apply: Some students may be in distance-education programs, have scheduling conflicts, or prefer recorded lectures.
  • Assumption 3: Students have access to online library resources. Why it may not apply: Some students may have limited internet access, use older devices, or face accessibility barriers with digital platforms.

Each assumption reflects a "typical student" profile that may exclude many learners.

📝 Exercise 7.1.4: Classify Forms of Bias
  • Difficulty: Intermediate
  • Estimated time: ~12–15 min
  • Learning outcome: Classify bias examples into appropriate types.

Task: For each scenario, identify the primary bias type(s) and justify your classification.

  1. An AI model for medical diagnosis is trained on data from urban hospitals and performs less accurately for rural patients.
  2. A hiring manager disregards an AI-generated candidate ranking that conflicts with her intuition, favouring candidates who attended her alma mater.
  3. A language model generates marketing copy that uses predominantly male pronouns for technical roles and female pronouns for caregiving roles.
  4. A company deploys an AI tool for performance reviews but does not provide employees with an explanation of how the tool works or a way to appeal its assessments.
Show sample classifications
  • (1) Data bias (training data not representative of rural populations) and model bias (performance varies by setting).
  • (2) Interpretation bias (the manager overrides the AI based on personal preference) and automation bias (the manager may over-trust or under-trust the AI selectively).
  • (3) Data bias (training data contains gender-role associations) and model bias (the model reproduces and amplifies them).
  • (4) Workflow bias (no explanation or appeal) and institutional bias (organisational policy does not require transparency or contestability).
📝 Exercise 7.1.5: Redesign a Prompt or Review Process
  • Difficulty: Advanced
  • Estimated time: ~15–20 min
  • Learning outcome: Redesign a prompt or review process to reduce bias risk.

Scenario: A team is using an AI tool to generate summaries of academic articles for a multidisciplinary journal. The current prompt is: "Summarise the key findings of this article." The summaries often omit contributions from non-Western scholars and overemphasise Western methodological approaches.

Task: Redesign the prompt and/or the review process to reduce these biases. Explain why your changes would help.

Show sample redesign

Redesigned prompt: "Summarise the key findings of this article. Identify the primary methodological approach, the geographic and cultural context of the research, and any contributions from Indigenous or non-Western scholars. Highlight how the findings relate to the global literature on this topic."

Redesigned review process: Instead of one reviewer accepting the AI summary, use two reviewers from different disciplinary or cultural backgrounds. Include a checklist that prompts reviewers to consider representation, language, and methodological diversity. Maintain a log of omissions and corrections to inform future training and prompt adjustments.

Why this helps: The revised prompt explicitly asks for context and diversity, reducing the AI's tendency to default to Western, mainstream sources. The review process introduces human checks that catch omissions and ensures that multiple perspectives inform the final output.

16. Homework Problems

Formative · Ungraded · Recommended preparation for Assignment 3.

🧠 Problem 7.1.1: Bias Entry Points
  • Estimated time: ~8–10 min

Task: A university uses an AI tool to generate course descriptions from a database of past syllabi. Describe two distinct bias entry points in this workflow and a possible mitigation for each.

Show sample answer
  • Entry point 1 — Data / sources: Past syllabi may reflect historical patterns in course offerings (e.g., fewer courses on Indigenous topics or global perspectives). Mitigation: Audit the database and add supplementary sources that represent diverse content.
  • Entry point 2 — Human interpretation: Course coordinators may accept the AI-generated descriptions without verifying their accuracy or inclusivity. Mitigation: Require a human review step with a checklist for representation, language, and accessibility.
🧠 Problem 7.1.2: Representation Analysis
  • Estimated time: ~10–12 min

Task: Provide two examples of underrepresentation in AI-generated content that you have encountered (in coursework, professional life, or everyday use). For each, describe who is absent and what effect the absence may have.

Show sample answer
  • Example 1: An AI travel-recommendation system suggests destinations primarily in Europe and North America, with few recommendations from Asia, Africa, or South America. Effect: Users may develop a narrow view of travel possibilities and miss out on cultural and economic diversity.
  • Example 2: An AI-powered news aggregator surfaces articles from a limited set of major outlets, omitting regional, Indigenous, and independent media. Effect: Users receive a homogenised perspective and may be unaware of important local or community-based issues.
🧠 Problem 7.1.3: Omission Analysis
  • Estimated time: ~8–10 min

Task: Explain why omission is a form of bias. Give one example of an omission in an AI-supported process and describe its potential consequence.

Show sample answer

Explanation: Omission is a form of bias because it systematically excludes certain groups, perspectives, or information. This exclusion can be as harmful as a direct misrepresentation because it denies visibility, voice, and recognition. Omission signals that certain people or experiences do not matter.

Example: An AI-supported history-learning tool omits Indigenous perspectives on colonisation, presenting only European accounts. Consequence: Students receive an incomplete and skewed understanding of history, and Indigenous students may feel excluded or misrepresented.

🧠 Problem 7.1.4: Stereotype Identification
  • Estimated time: ~8–10 min

Task: Identify two stereotypes that could be reinforced by an AI-powered career-recommendation system. Explain how they might affect users' aspirations.

Show sample answer
  • Stereotype 1: Nursing and caregiving roles are recommended to women; engineering and technical roles are recommended to men. Effect: Students may be steered toward careers aligned with gender stereotypes rather than their interests or abilities.
  • Stereotype 2: Creative roles are recommended to people from urban backgrounds; technical or manual roles are recommended to people from rural backgrounds. Effect: Students may internalise assumptions about what is "available" to them based on geography.
🧠 Problem 7.1.5: Prompt Design
  • Estimated time: ~10–12 min

Task: Design a prompt for an AI system that generates inclusive, representative examples of professionals in a field of your choice. Explain how your prompt encourages diversity and reduces stereotypes.

Show sample answer

Field: Environmental science.

Prompt: "Generate three profiles of environmental scientists. Include at least one scientist who works in a non-Western context, one who centres Indigenous knowledge, and one who uses community-based participatory methods. For each profile, describe their professional background, a key contribution, and the impact of their work on communities. Avoid using gendered language or stereotypes."

Explanation: The prompt explicitly requests diversity across geography, methodology, and community engagement. It asks for impact on communities rather than only individual achievement, broadening the definition of success.

🧠 Problem 7.1.6: Human Interpretation
  • Estimated time: ~10–12 min

Task: Describe a situation in which a human reviewer might misinterpret an AI output in a way that introduces bias. What safeguards could prevent this?

Show sample answer

Situation: A reviewer uses an AI-generated list of recommended candidates for a grant program. The AI ranks candidates using a model trained on past successful applications, which historically favoured applicants from certain institutions. The reviewer treats the ranking as objective and does not examine whether the model's criteria align with the program's equity goals.

Safeguards:

  • Train reviewers on automation bias and the limitations of AI-generated rankings.
  • Provide a transparency report that explains the model's criteria and known biases.
  • Require reviewers to justify any decision that diverges from the AI ranking.
  • Establish a contestability process for applicants who believe the process was biased.
🧠 Problem 7.1.7: Institutional Workflow
  • Estimated time: ~10–12 min

Task: An organisation adopts an AI tool for internal performance evaluations. The tool is used without any formal fairness audit or training for managers. Identify two institutional biases that may arise and suggest one policy change to address them.

Show sample answer
  • Institutional bias 1: Managers may use the AI tool to justify existing biases (e.g., rating certain employees lower based on historical performance trends that reflect discrimination).
  • Institutional bias 2: Without a contestability process, employees cannot challenge inaccurate or biased evaluations, so errors persist and accumulate.
  • Policy change: Require an annual fairness audit of the AI tool's outputs, including analysis of performance score distributions across demographic groups (using aggregate, anonymised data). Establish a formal appeal process for employees.
🧠 Problem 7.1.8: Uncertainty
  • Estimated time: ~8–10 min

Task: Why is it important to acknowledge residual uncertainty when assessing bias? What practical implications does this have for decision-making?

Show sample answer

Importance: Bias assessment is not a one-time, definitive determination. Data is incomplete, contexts change, and models evolve. Acknowledging uncertainty prevents over-confidence and encourages ongoing monitoring.

Practical implications:

  • Decision-makers should treat AI outputs as provisional rather than conclusive.
  • Systems should include feedback loops, monitoring, and regular reassessment.
  • Decisions that affect people's rights or opportunities should have human review and contestability pathways.
  • Uncertainty should be documented and communicated to stakeholders.
🧠 Problem 7.1.9: Transfer
  • Estimated time: ~10–12 min

Task: Choose a domain outside technology (e.g., education, health, law, public administration) and describe how the concepts of bias, representation, and default assumptions apply to AI use in that domain. Provide a concrete example.

Show sample answer

Domain: Health care.

Application: AI systems are used for patient triage, diagnostic support, and treatment recommendations. Bias can enter through training data that over-represents certain populations (e.g., clinical trials that historically included mostly white men). Default assumptions may include that all patients have access to the same health care infrastructure or that patients understand medical terminology.

Concrete example: An AI-powered symptom checker may perform less accurately for patients with rare genetic conditions or for those who describe symptoms in non-medical language. If the system is used to triage patients, it may direct some patients to lower-priority care, delaying diagnosis.

17. Mini-Project: Bias and Representation Audit

📁 Project: Bias and Representation Audit of an AI-Supported Use Case
  • Purpose: To apply the concepts from this lesson to a bounded, synthetic AI-supported use case, identifying bias entry points, representation issues, and preliminary mitigations.
  • Learning outcomes: Define bias entry points; conduct a representation audit; identify stereotypes and default assumptions; recommend preliminary mitigations.
  • Estimated effort: 90–120 minutes.
  • Context choices: Choose one of the following synthetic use cases, or use a case from your own experience (with syntheticised information).

Option A: AI-Generated Educational Content
An online course provider uses AI to generate practice questions and explanations for a large introductory course. The system draws from a repository of past exams and textbooks.

Option B: AI-Powered Public Health Information
A public health agency uses an AI chatbot to answer questions about vaccine eligibility, symptoms, and clinic locations. The system is trained on official communications and FAQ pages.

Option C: AI-Supported Recruitment Summaries
A company uses AI to summarise job applications for recruiters. The system generates brief profiles of candidates based on their submitted materials.

Required deliverable: A structured audit report (approximately 500–800 words or equivalent) containing:

  1. Human purpose — what is the legitimate goal?
  2. AI role — what does the AI do?
  3. Sources or context — what data or information does the AI use?
  4. Intended audience — who uses the AI output?
  5. Representation present — who or what appears in the outputs?
  6. Representation absent — who or what is missing?
  7. Default assumptions — what unexamined beliefs are embedded?
  8. Possible stereotypes — what stereotypes might the system reinforce?
  9. Bias entry points — at which stages can bias enter?
  10. Human interpretation risks — how might users misinterpret the outputs?
  11. Possible stakeholder effects — who benefits, who bears burdens?
  12. Evidence required — what additional evidence is needed?
  13. Preliminary mitigation — one or two concrete changes you would recommend.
  14. Residual uncertainty — what remains unknown?
  15. Escalation or stop condition — when should the use be paused or stopped?
  16. Initial recommendation — proceed, proceed with safeguards, redesign, or do not proceed?

Privacy and respectful-use boundaries: Use only synthetic, public, or generalised information. Do not use real personal, medical, financial, or employment records. Do not ask the AI to infer protected characteristics.

Show solution architecture (staged approach)
  1. Scope: Clearly define the AI use case and its boundaries. What is the system meant to do, and what is it not meant to do?
  2. Stakeholder mapping: Identify all affected groups, including end users, non-users, people whose data is used, and people who may be indirectly affected.
  3. Data/source audit: Examine the sources for diversity, omissions, and historical patterns.
  4. Output review: Generate a sample of outputs (or use supplied ones) and analyse representation, roles, language, and assumptions.
  5. Bias mapping: Use the ten-stage map (Section 14) to identify where bias can enter.
  6. Human factors: Consider how different users might interpret the outputs differently.
  7. Mitigation design: Propose changes to sources, prompts, review, or monitoring.
  8. Residual uncertainty: Document what you cannot know without further study or stakeholder input.
  9. Recommendation: Make a conditional recommendation based on the evidence.

This architecture helps ensure that your audit is systematic and thorough, covering the full sociotechnical workflow.

Show illustrative partial example (different context)

Context: AI-powered course-recommendation system for university students.

Human purpose: Help students discover courses that align with their interests and academic goals.

AI role: Recommender system using historical enrolment data and course descriptions.

Sources: Past enrolment records, course descriptions, student evaluations.

Intended audience: Current university students.

Representation present: Courses from all faculties, with emphasis on popular courses and high-enrolment programs.

Representation absent: Courses from smaller programs, interdisciplinary offerings, and Indigenous studies courses are under-represented.

Default assumptions: Students want high-enrolment courses; students have similar schedules and workload capacities; students are full-time, on-campus learners.

Possible stereotypes: The system may recommend stereotypically "masculine" or "feminine" subjects based on historical enrolment patterns.

Bias entry points: Data bias (historical enrolment patterns), model bias (recommendation algorithm may amplify popularity), interaction bias (students who search for niche topics may receive fewer recommendations).

Human interpretation risks: Students may treat the recommendations as authoritative and miss courses that are less popular but relevant to their interests.

Possible stakeholder effects: Students benefit from discovery; under-represented programs lose visibility; advisors may rely on the system rather than their own expertise.

Evidence required: Analysis of recommendation diversity, student feedback, and program-enrolment trends.

Preliminary mitigation: Adjust the recommendation algorithm to include a diversity factor that promotes less-visible courses; add a human-curated "hidden gems" list.

Residual uncertainty: We do not know how well the system serves students with different academic backgrounds or interests.

Escalation / stop condition: If the system consistently fails to recommend courses from certain programs, pause and review the algorithm.

Initial recommendation: Proceed with safeguards (monitoring, human review, and student feedback).

18. Self-Check Quiz

Eight questions · Formative · Ungraded. Expand each item to see feedback.

Question 1: Which of the following best describes bias in a sociotechnical context?

Select one:
Show answer and explanation

Correct answer: b.

Explanation: Bias in a sociotechnical context is systematic and emerges from the interaction of human decisions, data, algorithms, interfaces, and institutional structures. It is not limited to intentional prejudice (a) or to model properties alone (d). Not every error is bias (c) — bias is a systematic pattern, not a random error.

Question 2: Which of the following are types of bias in an AI-supported workflow? Select all that apply.

Select all that apply:
Show answer and explanation

Correct answers: All of the above (a–f).

Explanation: Bias can arise at every stage: in the data, in the model, in user interaction, in human interpretation, in the workflow design, and in the institutional context. A comprehensive fairness review addresses all six types.

Question 3: Match each scenario to the most relevant type of bias.

Table 3. Matching bias scenarios
Scenario Bias type
1. An AI model performs worse for speakers with regional accents. A. Representation bias
2. A reviewer accepts an AI ranking without questioning its criteria. B. Interpretation bias
3. A dataset contains mostly male engineers. C. Model bias
Show answer and explanation
  • 1 — C. Model bias: The model's performance varies across input characteristics (accents).
  • 2 — B. Interpretation bias: The reviewer uncritically accepts the AI output.
  • 3 — A. Representation bias: The data lacks diversity, leading to underrepresentation of certain groups.

Question 4: Why is it important to examine omissions in AI-generated content, not just the content that is present?

Show answer and explanation

Answer: Omissions can be as consequential as direct misrepresentations. When certain groups, perspectives, or experiences are systematically absent, it signals that they do not matter, which can lead to representational harm, missed opportunities, and unequal access to visibility and voice. Omissions also reflect the values and assumptions embedded in the data and workflow.

Question 5: True or false: If an AI system does not use explicit demographic labels, it cannot produce biased outcomes.

Select one:
Show answer and explanation

Answer: False.

Explanation: Bias can arise through proxy variables (e.g., postal code, language, device type) that correlate with demographic characteristics. Removing explicit labels does not remove bias if the data and model still encode patterns that produce unequal effects. Bias can also arise through interpretation, workflow, and institutional practices.

Question 6: Define automation bias and give one example from an AI-supported workflow.

Show answer and explanation

Answer: Automation bias is the tendency to over-rely on automated system outputs, even when contradictory evidence exists or the system is known to have limitations. Example: A doctor accepts an AI-generated diagnosis without considering the patient's unique symptoms and family history, leading to an incorrect treatment plan.

Question 7: Which of the following is not a default assumption that can appear in AI-generated content?

Select one:
Show answer and explanation

Correct answer: c. Users have a PhD in the relevant subject.

Explanation: While this could be an assumption in some specialised contexts, it is not a common default assumption in most AI-generated content. The other options (language, connectivity, disability access) are frequent and problematic default assumptions that exclude many users.

Question 8: Why is it important to distinguish between data bias and model bias?

Show answer and explanation

Answer: Data bias and model bias require different interventions. Data bias is addressed by improving the quality and diversity of training data, while model bias may require changes to the architecture, training process, or retrieval mechanisms. Distinguishing them helps allocate responsibility and design appropriate mitigations. It also clarifies that even a well-designed model can produce biased outcomes if the data is biased.

19. Assignment Connections

Assignment 3: Responsible AI Case Analysis

Lesson 7.1 provides foundational evidence for Assignment 3: Responsible AI Case Analysis (20% of final grade). As you work through this lesson, preserve the following evidence:

  • Bias entry points — your maps and classifications.
  • Representation audit — your findings on presence, absence, and stereotypes.
  • Default assumptions — your identification of unexamined beliefs in prompts, outputs, or review processes.
  • Stakeholder identification — who is affected, who is missing, who has authority.
  • Mitigations — your proposed changes to prompts, sources, review, or monitoring.

You will use these insights to develop a complete fairness analysis for your chosen case in Assignment 3. Do not submit this lesson work as your assignment — it is preparatory evidence.

📝 Assignment 1: Reflective Journal

Assignment 1 is a reflective journal that continues throughout the course. This lesson offers several prompts for your Unit 7 journal entry. Do not provide model journal responses — write your own reflections.

Show reflection prompts and planning checklist
  1. How has your understanding of bias changed after distinguishing it from deliberate prejudice?
  2. What surprised you about the number and variety of bias entry points in an AI workflow?
  3. Think of an AI-generated output you have encountered recently. Who was represented? Who was absent?
  4. What default assumptions have you noticed in AI-generated content, and how might they affect different users?
  5. How might you apply the bias-entry-point map to a context in your field of study or work?
  6. What responsible-use boundaries have you become more aware of after this lesson?

Reflection planning checklist:

  • I have identified at least one specific example of bias from my own experience or the lesson.
  • I have distinguished between data, model, interaction, interpretation, workflow, and institutional bias.
  • I have examined representation and omission in at least one AI output.
  • I have identified at least one default assumption and explained its potential effect.
  • I have connected my reflections to responsible AI use in my discipline or professional practice.

Use these points to structure your journal entry. Write from your own perspective; there is no single "correct" response.

20. Key Takeaways

Remember these points
  • Bias is sociotechnical: It emerges from the interaction of data, models, interfaces, human interpretation, and institutional structures—not solely from the AI model.
  • Six types of bias: Data, model, interaction, interpretation, workflow, and institutional bias each require distinct attention and interventions.
  • Representation matters: Who appears, who is absent, and who appears in stereotyped roles all affect how AI systems shape understanding and opportunity.
  • Default assumptions are everywhere: Unexamined beliefs about language, access, culture, and ability can exclude users and produce unequal effects.
  • Bias enters at every stage: From problem definition to remedy, each step of the workflow is a potential entry point—and a potential intervention point.
  • No single metric guarantees fairness: Even a mathematically "fair" model can produce unjust outcomes when used in context; fairness requires ongoing, human-centred evaluation.
  • Detection requires systematic audit: Identifying bias demands structured examination of sources, prompts, outputs, and human practices, not just intuition.

21. Key Terms

Bias (sociotechnical)
Systematic, non-random deviations from fairness, accuracy, or representativeness that arise from the interplay of human decisions, data, algorithms, interfaces, and institutional structures.
Data bias
Bias in the training data, including historical, representation, selection, measurement, and label bias.
Model bias
Bias introduced by model architecture, training process, or optimisation choices, independent of the data.
Interaction bias
Bias that arises from how users interact with the system—prompts, feedback, and behaviour—and how the system responds.
Interpretation bias
Bias in how humans understand, evaluate, and act on AI outputs, including automation bias.
Workflow bias
Bias embedded in the design of the overall process—who is consulted, what is measured, how errors are handled.
Institutional bias
Bias in policies, norms, and power structures that govern AI development, deployment, and oversight.
Representation
The presence and prominence of different groups, perspectives, and experiences in AI-generated content.
Underrepresentation
The systematic absence or marginalisation of certain groups, perspectives, or experiences.
Omission
The absence of relevant groups, perspectives, or information from AI-generated content or workflows.
Default assumption
An unexamined belief that shapes how systems are designed and how outputs are interpreted, often reflecting dominant cultural or social norms.
Stereotype
An oversimplified, generalised belief about a group that is often inaccurate and harmful.
Automation bias
The tendency to over-rely on automated system outputs, even when contradictory evidence exists or the system has known limitations.

22. Looking Ahead: Lesson 7.2

In Lesson 7.1, you have learned to identify bias, representation, and default assumptions in AI-supported workflows. This is the first step in a four-lesson progression:

After completing all four lessons, you will be well prepared for Assignment 3 and for Unit 8: Privacy, Security, and Data Protection.

Continue to Lesson 7.2: Fairness, Equity, Accessibility, and Intersectionality Return to Unit 7 Home

23. Sources & Further Reading

[VERIFICATION REQUIRED] — Please confirm all source details with the course coordinator and textbook publisher.

⚠️ Note

Some textbook content, definitions, examples, and figures are marked with placeholders. These will be supplied by the course coordinator or subject-matter expert.

👤