COMP 3270 / Module 3 / Unit 7 / Cumulative Project

Unit 7 Cumulative Project · Week 5 · Reading: Chapter 9 & FAIRER

FAIRER‑Based Bias, Fairness, and Inclusion Review

Apply the complete FAIRER framework to a bounded synthetic AI-supported case. Produce a structured fairness review that prepares you for Assignment 3 and consolidates all Unit 7 learning.

Estimated effort: 4–5 hours Formative · Ungraded Prerequisite: Lessons 7.1–7.4 & Consolidation

Table of Contents

1. Project Overview

This cumulative project integrates all the knowledge and skills you have developed in Unit 7. You will select a bounded synthetic AI-supported case and conduct a comprehensive fairness review using the FAIRER framework. The review will cover bias, representation, fairness, equity, accessibility, intersectionality, differential performance, proxy variables, disparate impact, participation, contestability, remedy, monitoring, and stop conditions.

The project is designed to prepare you for Assignment 3 (Responsible AI Case Analysis). You will produce a structured report (approximately 1,750–2,000 words or equivalent) that demonstrates your ability to systematically evaluate an AI-supported workflow and make a justified recommendation.

This is a formative, ungraded exercise — it is intended to build your confidence and provide evidence you can draw upon for your assignment. Do not submit this project as Assignment 3; it is a preparatory step.

2. Learning Outcomes

By completing this project, you will be able to:

  1. Define the human purpose and AI role in a bounded synthetic case.
  2. Identify all relevant stakeholders and map bias entry points across the workflow.
  3. Analyze representation, underrepresentation, stereotypes, and default assumptions.
  4. Distinguish equality, equity, fairness, inclusion, and accessibility in context.
  5. Select and justify relevant fairness criteria, recognizing potential conflicts.
  6. Identify intersecting barriers and proxy variables.
  7. Assess differential performance, disparate impact, and multiple forms of harm.
  8. Design meaningful participation, contestability, remedy, and monitoring plans.
  9. Apply the FAIRER framework comprehensively.
  10. Justify a final recommendation with clear evidence and acknowledgment of uncertainty.
  11. Connect your analysis to Assignment 3 and Unit 8.

3. Before You Begin

4. Context Choices

Select one of the following bounded synthetic contexts for your review. All are based on scenarios you have encountered in Unit 7, with additional detail provided below. You may also propose your own context, provided it is synthetic, bounded, and meets the project requirements.

📌 Option A: AI‑Powered Community Health Information Portal

As described in the Unit 7 Consolidation. The system provides personalized health information to residents, using a large language model trained on public health guidelines. It has issues with representation, accessibility, proxy variables (postal code), differential performance, and lack of contestability.

Key details: English-only, text-based, no screen-reader compatibility, postal code used for service recommendations, rural and non-native users report lower satisfaction.

📌 Option B: AI‑Powered Academic Advising System

A university uses an AI system to recommend courses to students based on their declared major, grades, and past enrolments. The system is intended to help students plan schedules efficiently. However, it recommends courses that predominantly attract certain demographic groups, no explanations are provided, and students cannot appeal. The system was not tested with students with disabilities.

Key details: Uses historical enrolment data, no transparency, no contestability, potential for stereotype reinforcement.

📌 Option C: AI‑Generated Public Service Summaries

A government agency uses an AI system to generate summaries of public policies and services (e.g., tax credits, housing assistance) for citizens. The summaries are generated in English and distributed on the agency's website and social media. The system is efficient but has been found to produce summaries that are more detailed and accurate for certain topics (e.g., urban-centric policies) and less accurate for others (e.g., rural or Indigenous-specific programs). No human review is performed, and there is no way for citizens to flag errors.

Key details: Differential performance by policy topic, no contestability, potential for representational harm.

📌 Option D: AI‑Supported Recruitment Summaries

A company uses an AI system to summarize job applications for recruiters. The system generates brief profiles of candidates based on their resumes, highlighting key qualifications. The system has been found to give lower scores to candidates with non-Western names or who attended community colleges, even when qualifications are similar. Recruiters rely heavily on the summaries and rarely review full resumes. There is no appeal process for candidates.

Key details: Proxy variables (name, institution), automation bias, allocation harm, no contestability.

Tip: If you choose your own context, ensure it is synthetic and does not involve real people, data, or sensitive information. Clearly describe the system, its purpose, its users, and the observed fairness issues.

5. Required Deliverable

Produce a structured report that addresses all the sections listed below. The suggested length is approximately 1,750–2,000 words or an equivalent structured submission (e.g., a detailed outline with bullet points may also be acceptable if it covers all points thoroughly).

Use the response template in Section 6 to structure your report. You may adapt the template to your writing style, but ensure that every section is addressed.

📁 Required sections (number them as shown):
  1. Use-case title
  2. Human problem
  3. Legitimate purpose
  4. AI role
  5. Intended users
  6. Affected non-users
  7. Stakeholders
  8. Sources and information
  9. Bias entry-point map
  10. Representation analysis
  11. Underrepresentation and omission
  12. Stereotype and default-assumption analysis
  13. Equality analysis
  14. Equity analysis
  15. Fairness criteria
  16. Inclusion analysis
  17. Accessibility audit
  18. Language and technology barriers
  19. Intersectional considerations
  20. Proxy-variable analysis
  21. Differential-performance evidence
  22. Evidence limitations
  23. Representational harms
  24. Allocation harms
  25. Quality-of-service harms
  26. Procedural harms
  27. Relational or capability harms
  28. Disparate-impact analysis
  29. Automation-bias risk
  30. Human authority
  31. Participation plan
  32. Notice and explanation
  33. Correction process
  34. Human-review process
  35. Contestability pathway
  36. Remedy
  37. Monitoring indicators
  38. Escalation
  39. Stop conditions
  40. Complete FAIRER analysis
  41. Residual uncertainty
  42. Final recommendation
  43. Assignment 3 connection
  44. Assignment 1 reflection connection
  45. Preparation for Unit 8

Recommendation options: Proceed, Proceed with safeguards, Pilot with monitoring, Limit the use, Redesign, Require specialist or community review, Provide a non-AI pathway, Pause, Do not proceed, Retire the use.

6. Response Template

Use the following template to structure your report. Copy this into your document and fill in each section. You may add sub-headings as needed. 

1. Use-case title: [Provide a descriptive title]

2. Human problem: [What real-world problem is being addressed?]

3. Legitimate purpose: [What is the legitimate goal?]

4. AI role: [What does the AI do?]

5. Intended users: [Who are the direct users?]

6. Affected non-users: [Who is affected but not using the system?]

7. Stakeholders: [List all affected groups, including indirect stakeholders]

8. Sources and information: [What data or sources does the AI use?]

9. Bias entry-point map: [For each of the ten stages, note potential bias entry points]

10. Representation analysis: [Who appears, who is absent, what roles are assigned?]

11. Underrepresentation and omission: [Describe missing perspectives or groups]

12. Stereotype and default-assumption analysis: [Identify stereotypes and unexamined assumptions]

13. Equality analysis: [What is the same for all users?]

14. Equity analysis: [What differences in needs, barriers, or outcomes exist?]

15. Fairness criteria: [List criteria and potential conflicts]

16. Inclusion analysis: [Who is involved in design and governance? Who is missing?]

17. Accessibility audit: [Identify barriers and propose mitigations]

18. Language and technology barriers: [Specific language, connectivity, device issues]

19. Intersectional considerations: [How do multiple barriers interact?]

20. Proxy-variable analysis: [List possible proxies and their effects]

21. Differential-performance evidence: [Describe performance differences across groups]

22. Evidence limitations: [What is unknown or uncertain?]

23. Representational harms: [Describe]

24. Allocation harms: [Describe]

25. Quality-of-service harms: [Describe]

26. Procedural harms: [Describe]

27. Relational or capability harms: [Describe]

28. Disparate-impact analysis: [Assess whether there is a materially unequal effect]

29. Automation-bias risk: [How might human over-reliance amplify issues?]

30. Human authority: [Who is accountable? Where are the human decision points?]

31. Participation plan: [Who should be involved, at what stages, and how?]

32. Notice and explanation: [How will users be informed and explained?]

33. Correction process: [How can users correct information?]

34. Human-review process: [When and how is human review triggered?]

35. Contestability pathway: [Describe the full pathway from notice to appeal]

36. Remedy: [What remedies are available?]

37. Monitoring indicators: [What metrics will be tracked?]

38. Escalation: [When and how are issues escalated?]

39. Stop conditions: [When would the system be paused or withdrawn?]

40. Complete FAIRER analysis: [Apply all FAIRER elements systematically]

41. Residual uncertainty: [What remains unknown?]

42. Final recommendation: [Choose one option and justify]

43. Assignment 3 connection: [How does this analysis support your Assignment 3?]

44. Assignment 1 reflection connection: [Key insights for your journal]

45. Preparation for Unit 8: [What questions or issues arise about privacy and security?]

7. Assessment Criteria

This project is formative and ungraded, but you can use the following criteria to self-assess the quality of your work. A strong submission will:

Self-assessment checklist
  • I have defined the purpose and AI role clearly.
  • I have mapped bias entry points at every stage.
  • I have analyzed representation, stereotypes, and default assumptions.
  • I have distinguished equality, equity, fairness, inclusion, and accessibility.
  • I have identified relevant fairness criteria and conflicts.
  • I have examined intersectionality and proxy variables.
  • I have assessed differential performance, disparate impact, and harms.
  • I have designed participation, contestability, remedy, and monitoring.
  • I have applied FAIRER comprehensively.
  • My recommendation is justified and includes conditions.
  • I have documented residual uncertainty.
  • I have used only synthetic or public information.

8. Privacy and Respectful‑Use Requirements

⚠️ Critical boundaries
  • Use only synthetic, public, or authorized information. Do not use real personal, medical, financial, employment, or educational records.
  • Do not enter personal, confidential, proprietary, privileged, security-sensitive, or culturally governed information into any AI system for this project.
  • Do not ask AI to infer protected or sensitive characteristics about real or synthetic individuals.
  • Do not generate demeaning, hateful, or harassing stereotypes — your analysis should be respectful.
  • Do not treat one synthetic output as proof of system-wide bias — base your analysis on plausible scenarios and patterns.
  • Do not claim a legal violation without jurisdiction-specific expert review.
  • Do not assume that participation automatically resolves power differences — acknowledge limitations.
  • Do not treat disclosure as remedy — your analysis should include concrete correction and redress mechanisms.
  • Do not provide a complete Assignment 3 submission — this is preparatory work.

9. Self‑Review Checklist

Before submitting (or before using this for Assignment 3), verify that your report meets these criteria:

10. Staged Hints

Use these hints to guide your analysis. Each hint opens a new perspective; read them in order if you need help getting started.

Show Hint 1: Start with the big picture

Define the case clearly. Write a one-paragraph description of the system, its purpose, and the context. Then list all the stakeholders you can think of, including those who might be indirectly affected. This will anchor your analysis.

Show Hint 2: Map bias entry points early

Use the ten-stage map from Lesson 7.1: problem definition, data/sources, prompt, context, model output, retrieval, human interpretation, decision, monitoring, remedy. For each stage, ask: What could go wrong here? This will help you systematically identify risks.

Show Hint 3: Examine representation with a lens

When analyzing representation, look not just at who appears, but how they appear — roles, language, prominence, and whether they are stereotyped. Also, explicitly list who is absent and why that matters.

Show Hint 4: Fairness criteria are not one-size-fits-all

Different stakeholders may prioritise different criteria. List at least three criteria (e.g., equal access, equitable support, comparable error rates) and explain how they might conflict. Then, justify which you would prioritise and why.

Show Hint 5: Intersectionality requires concrete examples

Don't just state that intersectionality matters. Give a specific example of a user who faces multiple barriers (e.g., a non-native English speaker with a visual impairment). Describe how these barriers interact to create a compounded disadvantage.

Show Hint 6: Contestability is a process, not a statement

Design a step-by-step pathway: notice → explanation → correction → human review → challenge → remedy. For each step, specify who is responsible, how the user initiates it, and the expected timeline.

Show Hint 7: Monitoring must be actionable

Select indicators that you can actually measure (e.g., satisfaction scores, complaint rates, override rates). Then define thresholds that would trigger reassessment or escalation.

Show Hint 8: Recommendations should be conditional

Even if you recommend proceeding, include safeguards and conditions. A good recommendation says: "Proceed with safeguards, provided that [conditions] are met, and with a commitment to stop if [triggers] occur."

11. Solution Architecture

This architecture outlines a systematic approach to completing the project. It is not a solution but a guide to the analytical process.

Show solution architecture
  1. Define and scope: Write a clear description of the case, its purpose, and the AI's role. Identify all stakeholders.
  2. Map bias entry points: Use the ten-stage map to systematically identify where bias can enter.
  3. Analyze representation and assumptions: Examine who appears, who is absent, what stereotypes exist, and what default assumptions are embedded.
  4. Apply equality/equity/fairness/inclusion/accessibility: For each concept, describe how it applies and identify gaps.
  5. Select fairness criteria: List the criteria that matter most, and discuss potential conflicts.
  6. Examine intersectionality and proxies: Give concrete examples of intersecting barriers and possible proxy variables.
  7. Assess differential performance and harms: Use the evidence (synthetic) to identify performance differences and classify harms (allocation, quality, representational, procedural, relational, capability).
  8. Assess disparate impact and automation bias: Determine if there is a materially unequal effect and how human over-reliance might amplify it.
  9. Design participation and contestability: Propose who should be involved and how, and design a full contestability pathway.
  10. Define remedy, monitoring, and stop conditions: Specify what remedies are available, what metrics to track, and when to stop.
  11. Apply FAIRER: Synthesize all the above into a structured FAIRER analysis.
  12. Make a recommendation: Choose an option, justify it, and include conditions and residual uncertainty.

This architecture ensures that your analysis is comprehensive, systematic, and grounded in the Unit 7 frameworks.

12. Illustrative Partial Example

Below is a partial example based on a different context (an AI-powered library search system). This example illustrates the depth and style of analysis expected, but it is not a complete submission. Do not copy this — use it as a model for your own work.

Show illustrative partial example

Context: AI-powered library search and recommendation system for a public library.

1. Use-case title: Library Discovery AI

2. Human problem: Patrons need help finding relevant books and resources efficiently.

3. Legitimate purpose: To enhance access to library collections and support discovery.

4. AI role: Recommender system using past borrowing patterns, search history, and metadata.

5. Intended users: Library patrons of all ages and backgrounds.

6. Affected non-users: Patrons who do not use the system due to accessibility barriers or lack of digital literacy.

7. Stakeholders: Patrons, librarians, library administrators, community representatives, and disability advocates.

8. Sources and information: Library catalogue, borrowing records (anonymised), user search logs.

9. Bias entry-point map:

  • Problem definition: Defining "relevance" based on popularity may marginalise niche interests.
  • Data/sources: Borrowing history may reflect historical inequities in collection development.
  • Prompt: Search queries assume certain literacy levels and language.
  • Context: The system is designed for in-library use, not remote access.
  • Model output: Recommendations may be dominated by mainstream authors.
  • Retrieval: Metadata may misclassify or omit Indigenous or non-Western works.
  • Human interpretation: Librarians may trust the system without examining its biases.
  • Decision: Patrons may miss relevant resources if not recommended.
  • Monitoring: No collection diversity audit.
  • Remedy: No mechanism for patrons to suggest additions.

10. Representation analysis: The system's top recommendations are dominated by bestselling Western authors. Authors from Indigenous, immigrant, and non-Western backgrounds are underrepresented. Character descriptions in generated summaries tend to default to gender stereotypes.

11. Underrepresentation and omission: Works in languages other than English, Indigenous knowledge, and self-published works are rarely recommended.

12. Stereotype and default-assumption analysis: The system assumes patrons have high literacy, are fluent in English, and are comfortable with digital interfaces. It also assumes that "popular" equals "relevant".

13. Equality analysis: All patrons see the same search interface and recommendation algorithm.

14. Equity analysis: Patrons with visual impairments or low literacy cannot use the system effectively; no alternative formats are provided.

15. Fairness criteria: Equal access (everyone can search) vs. equitable outcomes (all patrons find relevant materials). Conflict: providing additional support (e.g., audio) requires resources and might reduce universal access if not implemented carefully.

16. Inclusion analysis: The system was designed by IT staff and librarians, but no patrons were involved in testing or design.

17. Accessibility audit: Lack of screen-reader compatibility, no captioning, no audio summaries, text-only.

18. Language and technology barriers: English-only interface; assumes high-speed internet.

19. Intersectional considerations: An older patron with low vision and limited digital literacy faces compounded barriers — they cannot read small text, nor navigate the interface.

20. Proxy-variable analysis: Borrowing history may proxy for socioeconomic background (access to books), which may correlate with race and class.

21. Differential-performance evidence: User satisfaction surveys show lower satisfaction among patrons with disabilities and non-native speakers.

22. Evidence limitations: Satisfaction data is sparse; we lack qualitative feedback.

23. Representational harms: The system reinforces the dominance of Western, English-language culture.

24. Allocation harms: Some patrons may miss out on relevant resources, affecting their access to information.

25. Quality-of-service harms: Recommendations are less useful for patrons with diverse interests.

26. Procedural harms: No way to flag missing works or incorrect metadata.

27. Relational or capability harms: Trust in the library may erode if certain communities feel underrepresented.

28. Disparate-impact analysis: The system has a disparate impact on patrons with disabilities and non-English speakers, who are less likely to find what they need.

29. Automation-bias risk: Librarians may rely on the system's recommendations rather than their own expertise to suggest books.

30. Human authority: Librarians have final say, but they are not required to review the system's outputs.

31. Participation plan: Form a patron advisory group including seniors, disability advocates, and multicultural representatives. Conduct usability testing and incorporate feedback.

32. Notice and explanation: Add a disclaimer that the system uses borrowing history and popularity, and provide a link to more information.

33. Correction process: Allow patrons to suggest books and flag irrelevant recommendations.

34. Human-review process: A librarian reviews flagged recommendations and updates the system if needed.

35. Contestability pathway: Patron flags → librarian review → if unresolved, appeal to head librarian → final decision with explanation.

36. Remedy: Correction of metadata, addition of missing works, and a personal recommendation from a librarian.

37. Monitoring indicators: User satisfaction by disability status and language; diversity of recommended works; complaint rates.

38. Escalation: If complaints from a group increase by 20%, escalate to a full audit.

39. Stop conditions: If the system consistently fails to recommend works from underrepresented groups, pause and redesign.

40. FAIRER analysis: (Summarised) Fairness goals: equitable discovery. Stakeholders: patrons, librarians. Inclusion: involve affected communities. Differential effects: performance varies by ability and language. Evidence: satisfaction surveys, complaint logs. Redress: contestability and remedy as above.

41. Residual uncertainty: We do not know the long-term impact on patron engagement; we lack data on how the system affects different groups over time.

42. Final recommendation: Redesign and pilot with monitoring. The system has significant gaps, but with the proposed participation, contestability, and monitoring, it can be improved. Maintain a non-AI pathway (librarian assistance).

43. Assignment 3 connection: This analysis will provide evidence for the Assignment 3 case on library AI, covering stakeholder mapping, bias entry, and FAIRER.

44. Assignment 1 reflection: I have learned that even well-meaning systems can exclude users if accessibility and representation are not prioritised from the start.

45. Preparation for Unit 8: Privacy concerns arise from using borrowing history; I need to consider how data is stored and anonymised.

End of partial example.

13. Key Takeaways

Remember these points
  • A comprehensive fairness review requires systematic analysis — from purpose to remedy.
  • Bias, fairness, inclusion, and accessibility are interconnected — none can be addressed in isolation.
  • Participation and contestability are non-negotiable — they operationalize accountability.
  • Monitoring and stop conditions turn fairness into a continuous process — not a one-time check.
  • Recommendations must be justified, conditional, and transparent about uncertainty.
  • This project prepares you for Assignment 3 — the skills and content are directly transferable.

14. Key Terms

FAIRER framework
A structured approach to evaluating fairness, inclusion, and remedy in AI-supported workflows. [INSERT AUTHOR-APPROVED EXPANSION].
Contestability
The ability of affected people to challenge decisions, correct information, and obtain human review and remedy.
Remedy
Action taken to correct harms, including correction, reconsideration, restoration, compensation, or system changes.
Monitoring
Ongoing observation of performance, complaints, and outcomes to detect emerging issues.
Stop condition
A predefined circumstance under which a system should be paused, restricted, or withdrawn.
Disparate impact
A materially unequal effect that occurs even with a neutral process and no discriminatory intent.
Proxy variable
A measurable attribute that correlates with or indirectly reveals a protected characteristic.
Intersectionality
The recognition that people experience overlapping social positions and barriers that interact.

15. Looking Ahead

After completing this cumulative project, you will have a solid foundation for Assignment 3. You can use the analysis and evidence from this project as a starting point for your case study, but remember: do not submit this project as your assignment. Your Assignment 3 should be a polished, refined analysis that integrates feedback from this formative work.

Next, you will move to Unit 8: Privacy, Security, and Data Protection, where you will learn about the PROTECT framework, information classification, purpose limitation, data minimization, and incident response. The privacy and security considerations you identified in this project will be useful there.

Continue to Unit 8: Privacy, Security, and Data Protection Return to Unit 7 Home
👤