Appendix D: Templates and Worksheets — AI Audit Report

These templates support the progressive AI Audit Report project that runs through all 21 chapters. Each template corresponds to a phase of the audit. You can use them in order or jump to the sections most relevant to your current chapter. Photocopy-friendly versions may be available from your instructor.

Template 1: System Selection Worksheet (Chapter 1)

Use this worksheet when choosing the AI system you will audit throughout the course.

Part A: Identifying Your System

Field	Your Response
System Name
Developed by
Year Launched / Deployed
Primary Function (What does it do?)
Domain (healthcare, education, criminal justice, social media, hiring, finance, other)
Who Uses It? (end users, organizations, governments)
Who Is Affected by It? (even if they don't use it directly)

Part B: Initial FACTS Framework Assessment

Apply the FACTS Framework from Chapter 1 to your chosen system:

Letter	Question	Your Initial Answer
F — Function	What specific task does this system perform?
A — Accuracy	How well does it work, and for whom? (Note: you may not know yet — that is fine. Record what you think and revisit later.)
C — Consequences	Who benefits from this system? Who might be harmed?
T — Training	What data was it trained on? Who curated that data? (Research as much as you can; note what you cannot find out.)
S — Stewardship	Who is responsible when this system makes an error?

Part C: Selection Justification

In 3-5 sentences, explain why you chose this system. Consider: Is it personally relevant? Socially significant? Currently in the news? Does it raise questions you find genuinely interesting?

Part D: Initial Impressions

Record your initial impressions before conducting research. What do you assume about this system? What do you expect to find? You will revisit these assumptions at the end of the course.

Template 2: Technical Analysis Template (Chapters 3–6)

Use this template to document the technical foundations of your chosen AI system.

Learning Approach (Chapter 3)

Question	Your Analysis
What type of machine learning does this system use? (supervised, unsupervised, reinforcement, hybrid, unknown)
What evidence supports your classification?
What does the system's model learn to do? (classify, predict, generate, recommend, detect)
What would overfitting look like for this system?
What would underfitting look like?

Data Profile (Chapter 4)

Question	Your Analysis
What types of data does the system use? (structured, unstructured, labeled, unlabeled)
Where does the training data come from?
Who labeled or annotated the data? Under what conditions?
What populations, perspectives, or scenarios might be underrepresented?
What "ghost data" (relevant information that is absent) might affect performance?

Model Architecture (Chapters 5–6, if applicable)

Question	Your Analysis
Does this system use a large language model? If so, describe its role.
Does this system use computer vision? If so, describe what visual tasks it performs.
What other AI techniques does the system use?
How does the system handle cases it is uncertain about?

Template 3: Data Audit Checklist (Chapter 4)

Use this checklist to systematically evaluate the data foundations of your AI system. Check each item you can confirm; flag items you cannot determine.

Provenance

[ ] The source(s) of the training data are documented
[ ] The time period the data covers is specified
[ ] The geographic scope of the data is documented
[ ] The method of data collection is described
[ ] Any data licensing or consent frameworks are documented

Representativeness

[ ] The data includes a demographic breakdown (if applicable to the system's domain)
[ ] Known gaps or underrepresented groups are acknowledged
[ ] The data reflects the population the system will serve (not just the population that was easiest to collect data from)
[ ] Edge cases and rare events are represented in the data

Label Quality

[ ] The labeling process is documented (who labeled, how many labelers, what guidelines)
[ ] Inter-annotator agreement rates are reported (how often labelers agreed)
[ ] Ambiguous cases are handled with a documented procedure
[ ] Labels are periodically reviewed and updated

[ ] Data subjects consented to their data being used for this purpose
[ ] The data collection passed an ethics review (IRB or equivalent)
[ ] Sensitive or personal data is handled according to applicable regulations (GDPR, CCPA, etc.)
[ ] Data subjects have the ability to opt out or request deletion

Known Limitations

[ ] The dataset developers have published known limitations
[ ] Potential biases are acknowledged in documentation
[ ] The dataset has been audited by independent researchers
[ ] The dataset has been updated or corrected in response to identified issues

Summary: How many items could you confirm? How many could you not determine? What does the pattern of "cannot determine" entries tell you about the system's transparency?

Template 4: Bias Assessment Framework (Chapter 9)

Use this framework to conduct a structured bias audit of your AI system.

Stage 1: Problem Formulation

What is this system designed to do?
What assumptions are embedded in this framing?
Who defined the problem — and whose perspective is centered?
What alternative framings of the problem are possible?

Stage 2: Data Assessment

What sources of historical bias exist in the training data?
What representation biases exist? (Which groups are overrepresented or underrepresented?)
What measurement biases might be present? (Are the proxy variables valid for all groups?)
What aggregation biases exist? (Is a single model appropriate for all subpopulations?)

Stage 3: Model Assessment

Does the system's performance vary across demographic groups?
What fairness metrics have been applied? (demographic parity, equalized odds, calibration)
Are there proxy variables that could encode protected characteristics?
Has the model been tested on adversarial or edge-case inputs?

Stage 4: Deployment Assessment

How is the system used in practice? (Does actual use match intended use?)
Are there feedback loops that could amplify bias over time?
Do users understand the system's limitations?
Is there a mechanism for affected individuals to challenge decisions?

Stage 5: Mitigation Recommendations

For each bias you identified, propose at least one mitigation strategy:

Bias Identified	Stage	Proposed Mitigation	Type (Technical / Organizational / Policy)

Template 5: Stakeholder Impact Map (Chapters 7, 9, 17)

Use this template to map who is affected by your AI system and how.

Stakeholder Identification

List all groups affected by the system, including groups that may not be direct users:

Stakeholder Group	Relationship to System	Potential Benefits	Potential Harms	Power to Influence System
				High / Medium / Low
				High / Medium / Low
				High / Medium / Low
				High / Medium / Low
				High / Medium / Low

Power Analysis

Which stakeholders have the most influence over the system's design and deployment?
Which stakeholders are most affected but have the least influence?
What mechanisms exist for low-power stakeholders to have their concerns heard?

Template 6: Governance Review Checklist (Chapter 13)

Use this checklist to evaluate the governance and regulatory environment around your AI system.

Regulatory Landscape

[ ] I have identified which jurisdiction(s) govern this system
[ ] I have determined whether the system falls under the EU AI Act's risk classification (if applicable)
[ ] I have identified relevant sector-specific regulations (e.g., FDA for healthcare, FERPA for education)
[ ] I have determined whether the system has undergone any required regulatory review

Accountability Structures

[ ] There is a clearly identified party responsible for the system's decisions
[ ] There is a documented process for individuals to appeal or challenge AI-generated decisions
[ ] There is a mechanism for reporting errors or harms
[ ] Audit trails exist that allow decisions to be reviewed after the fact

Transparency

[ ] The system's use of AI is disclosed to affected individuals
[ ] Documentation about the system's capabilities and limitations is publicly available
[ ] The system has been subject to independent audit or review
[ ] Regular performance reports are published

Governance Recommendations

Based on your review, list three specific governance improvements you would recommend:

Template 7: Environmental Footprint Estimator (Chapter 18)

Use this template to estimate the environmental impact of your AI system.

Training Phase

Factor	Estimate or Research Finding	Source / Basis for Estimate
Model size (parameters)
Estimated training time (GPU-hours)
Estimated energy consumption (kWh)
Estimated carbon emissions (kg CO2e)
Data center location and energy source

Inference Phase (Ongoing Use)

Factor	Estimate or Research Finding	Source / Basis for Estimate
Estimated daily/monthly queries or uses
Energy per query (if available)
Estimated annual energy consumption
Estimated annual carbon emissions

Hardware and Infrastructure

Factor	Estimate or Research Finding	Source / Basis for Estimate
Type of hardware (GPUs, TPUs, etc.)
Estimated hardware lifecycle
Water consumption for cooling (if available)
E-waste considerations

Net Environmental Assessment

What environmental benefits (if any) does this system provide? (e.g., optimizing energy use, reducing waste, enabling climate research)
Do the benefits outweigh the costs? Under what assumptions?
What specific changes could reduce the environmental footprint?

Template 8: Final Report Outline (Chapter 21)

Use this template to structure your completed AI Audit Report.

I. Executive Summary (1 page)

System description (2-3 sentences)
Key findings (3-5 bullet points)
Primary recommendation (1-2 sentences)

II. System Overview (2-3 pages)

What the system does and who uses it
Technical foundations (learning approach, data, model architecture)
Historical context (when developed, how it evolved)

III. Technical Analysis (3-4 pages)

Data audit findings (Template 3)
Model performance analysis
Failure modes and limitations

IV. Bias and Fairness Assessment (3-4 pages)

Bias assessment findings (Template 4)
Fairness metrics applied
Identified disparities and their significance

V. Stakeholder Impact Analysis (2-3 pages)

Stakeholder map (Template 5)
Power analysis
Voices and perspectives

VI. Governance and Accountability (2-3 pages)

Regulatory landscape (Template 6)
Accountability structures
Gaps and recommendations

VII. Environmental Footprint (1-2 pages)

Environmental impact estimate (Template 7)
Net assessment

VIII. Global and Cross-Cultural Considerations (1-2 pages)

How the system operates across cultural contexts
Digital sovereignty and data governance implications

IX. Safety and Alignment Assessment (1-2 pages)

Near-term safety concerns
Alignment considerations
Proposed safeguards

X. Recommendations (1-2 pages)

Technical recommendations
Governance recommendations
Policy recommendations
Prioritized action items

XI. Reflection (1 page)

What surprised you during this audit?
How did your understanding of AI change?
What would you investigate further with more time and resources?

XII. References

Organized by citation tier (see Appendix I)

Tips for Using These Templates

Start rough, refine later. Your first pass through any template will have gaps — that is expected. Return to earlier templates as you learn more in later chapters.
"I don't know" is valuable data. If you cannot determine something about your AI system, document that opacity. The inability to find information is itself a finding about the system's transparency.
Cite your sources. Even rough estimates should be traceable. Note whether information came from the company, independent research, news reporting, or your own inference.
Be specific. "The system might be biased" is not useful. "The training data underrepresents rural populations because it was collected from three urban hospitals" is useful.
Revisit your initial impressions. Template 1 asks you to record assumptions before research. Compare those assumptions to your final findings — the gap between them is part of your learning.