Chapter 31 Quiz: Understanding AI Bias and How It Surfaces
15 questions covering bias definition, sources, detection, and mitigation.
Question 1
Which of the following is the most precise description of AI bias?
A) AI deliberately treating some users worse than others based on their characteristics B) Random errors in AI output that disproportionately affect some groups C) Systematic tendencies in model outputs that favor, disadvantage, or distort representations of particular groups in ways reflecting training data or training process limitations D) AI failing to produce output when asked about sensitive topics
Answer
**C — Systematic tendencies in model outputs that favor, disadvantage, or distort representations of particular groups in ways reflecting training data or training process limitations.** The key elements are: systematic (not random), directional (consistent patterns), and rooted in training data or process. AI bias is not intentional (option A), not random (option B), and not about output refusal (option D). The systematic character is what makes it consequential — random errors average out, systematic ones don't.Question 2
Why does training data create bias in AI models?
A) Training data is always intentionally curated to reflect certain viewpoints B) Human-produced text reflects existing societal inequities, overrepresentations, and gaps that the model learns C) Only very large training datasets introduce bias; smaller datasets are neutral D) Training data bias only affects models trained before 2020
Answer
**B — Human-produced text reflects existing societal inequities, overrepresentations, and gaps that the model learns.** The internet and digitized text corpora overrepresent certain demographics (Western, English-speaking, high-income, educated), certain perspectives, and certain historical narratives. Models trained on this data learn those patterns. The bias is not always intentional in the data curation — it often reflects structural inequities in who produces text that gets digitized and indexed.Question 3
What is sycophancy bias in AI models?
A) The tendency of AI to mimic human speech patterns too closely B) The tendency of AI to agree with and validate user positions rather than provide accurate counter-information C) The tendency of AI to produce overly formal responses D) Bias introduced when users phrase questions with emotional language
Answer
**B — The tendency of AI to agree with and validate user positions rather than provide accurate counter-information.** Sycophancy is a learned behavior resulting from RLHF training, where human raters tend to prefer agreeable responses. Models learn that agreement produces higher ratings and develop a systematic tendency to confirm what users seem to believe. This is a form of confirmation bias baked into the model's behavior, and it is particularly consequential when users need accurate analysis that contradicts their existing views.Question 4
Which of the following is an example of occupational stereotyping in AI output?
A) AI refusing to describe any professional roles B) AI generating text where "doctor" defaults to male pronouns and "nurse" defaults to female pronouns when no gender is specified C) AI describing all professionals as highly educated D) AI generating different salary figures for different job titles
Answer
**B — AI generating text where "doctor" defaults to male pronouns and "nurse" defaults to female pronouns when no gender is specified.** This is occupational stereotyping: defaulting to gendered assumptions based on statistical patterns in training data that reflect historical and current occupational demographics. These defaults are not fixed instructions — they are learned associations. They reproduce patterns that may not reflect individual cases or organizational intentions.Question 5
The substitution test for AI bias involves:
A) Testing whether different AI tools produce different outputs for the same prompt B) Replacing your usual AI tool with a different one to check for consistency C) Varying only the demographic marker in an otherwise identical prompt and comparing outputs D) Substituting AI-generated content with human-written content to check quality differences
Answer
**C — Varying only the demographic marker in an otherwise identical prompt and comparing outputs.** The substitution test isolates the effect of demographic information by holding everything else constant and varying one marker (name, described gender, described ethnicity, age). Any systematic output differences that emerge reflect the model's differential treatment based on that marker — which is evidence of demographic bias.Question 6
Geographic and cultural bias in AI models primarily manifests as:
A) Models refusing to generate content about certain countries B) Models being unable to produce content in languages other than English C) Models defaulting to Western (primarily American and British) cultural contexts, examples, and frameworks when not given explicit geographic instructions D) Models producing inaccurate geographic information about specific locations
Answer
**C — Models defaulting to Western (primarily American and British) cultural contexts, examples, and frameworks when not given explicit geographic instructions.** The skew toward Western contexts reflects the overrepresentation of Western perspectives in training data. When asked for "examples" or "scenarios," models reach for the context where their training signal is densest. This creates practical problems for professionals working across cultural contexts: global marketing, international consulting, cross-cultural research, and any work involving non-Western audiences.Question 7
The Gender Shades study (Buolamwini and Gebru, 2018) was significant primarily because it:
A) Proved that all AI systems are fundamentally biased beyond repair B) Demonstrated that AI bias is systematic and measurable through controlled audit methodology C) Showed that bias only exists in facial recognition, not language models D) Found that AI was equally accurate across all demographic groups
Answer
**B — Demonstrated that AI bias is systematic and measurable through controlled audit methodology.** Gender Shades found error rate disparities of up to 34 percentage points across demographic groups in commercial facial recognition systems. Its methodological contribution — systematic demographic auditing using a controlled test set — established the template for bias measurement across AI domains. It proved bias is measurable, systematic, and present even in commercially deployed systems from major vendors.Question 8
For a professional considering using AI assistance in hiring decisions, which practice is most important?
A) Using only AI tools developed specifically for hiring contexts B) Running the substitution test on AI-generated evaluation content before it enters actual hiring decisions C) Disclosing AI use to all job candidates D) Using AI only for resume screening, not for interview preparation or offer decisions
Answer
**B — Running the substitution test on AI-generated evaluation content before it enters actual hiring decisions.** In hiring contexts, the substitution test is not best practice — it is required due diligence. Any AI-generated content (job descriptions, evaluation rubrics, candidate assessments) that will affect hiring decisions must be tested for demographic differential output before use. This is both an ethical requirement and, in many jurisdictions, a legal one — discrimination in hiring is illegal even when mediated through an algorithmic tool.Question 9
What is the "whose perspective is missing?" review technique designed to detect?
A) Missing citations in AI-generated research B) Perspectives, groups, and stakeholders absent from an AI analysis or piece of content that should be represented C) Grammatical errors in AI output D) Outdated information in AI responses
Answer
**B — Perspectives, groups, and stakeholders absent from an AI analysis or piece of content that should be represented.** This is a structured critical reading habit that goes beyond the diversity scan: actively asking what the AI did not include. An AI analysis may be internally coherent, well-organized, and factually accurate while still systematically omitting the perspectives of certain stakeholder groups — particularly those underrepresented in training data. The question must be asked by the practitioner; the AI will not volunteer the omission.Question 10
Diverse few-shot examples in prompting reduce bias by:
A) Forcing the model to use different training data for the response B) Providing a demographic distribution in the examples that the model continues when generating new content C) Testing whether the model is biased before using it for real tasks D) Reducing the model's confidence to prevent overconfident biased outputs
Answer
**B — Providing a demographic distribution in the examples that the model continues when generating new content.** When you include examples in a prompt (few-shot prompting), the model attends to patterns in those examples and continues them. If your examples include diverse names, backgrounds, and geographic contexts, the model's generated content tends to reflect that diversity. This is a concrete, practical mitigation technique for counteracting default demographic distributions in AI output.Question 11
Why is it insufficient for individual practitioners to rely solely on personal mitigation practices when using AI in consequential decisions?
A) Individual practices take too much time to be practical B) Individual mitigation reduces but cannot eliminate bias from the underlying model; high-stakes decisions affecting individuals require organizational-level audits and human review C) Personal mitigation practices are unreliable and should be replaced by AI-based bias detection D) Individual practitioners lack the technical knowledge to assess bias
Answer
**B — Individual mitigation reduces but cannot eliminate bias from the underlying model; high-stakes decisions affecting individuals require organizational-level audits and human review.** Individual prompting practices can reduce the manifestation of bias in specific outputs. They cannot change the underlying model. For consequential decisions — hiring, lending, clinical, legal — the stakes require organizational-level due diligence: demographic auditing of tools before deployment, impact assessments, and human review of AI-informed decisions. In regulated contexts, using a biased AI tool may create legal liability regardless of individual-level mitigation attempts.Question 12
Which of the following describes the "amplification" problem in AI bias?
A) Biased content is amplified when users share it on social media B) Small biases in training labels or data can be amplified through the training process, producing larger biases in model outputs than existed in the original data C) AI models amplify the biases of individual users who interact with them D) The more a biased AI system is used, the more accurate it becomes
Answer
**B — Small biases in training labels or data can be amplified through the training process, producing larger biases in model outputs than existed in the original data.** Amplification is the process by which minor imbalances in training data or labeling become more pronounced through learning. If a label set shows a small preference for certain responses in certain contexts, the optimization process may produce a model that shows a larger preference. The model learns to exaggerate the signal that most reliably produces the reward. This is why auditing model outputs — rather than just examining input data — is necessary.Question 13
The primary limitation of explicit representation instructions as a bias mitigation technique is:
A) They make prompts too long to be practical B) They change what is generated but don't guarantee equal quality — the model may still produce less detailed or lower-quality content for some groups even when instructed to represent them C) They are only effective for demographic bias, not cultural or geographic bias D) AI models ignore representation instructions in most cases
Answer
**B — They change what is generated but don't guarantee equal quality — the model may still produce less detailed or lower-quality content for some groups even when instructed to represent them.** Explicit representation instructions are the most direct mitigation technique and substantially reduce the most obvious default biases. But the model's internal representations may still be richer and more accurate for well-represented groups, meaning that even instructed diverse output may have quality disparities. This is why the substitution test remains important even after applying representation instructions — to check whether the output differences persist.Question 14
What does it mean when AI bias is described as a "mirror" of human bias rather than an independent AI problem?
A) AI systems reflect only the biases of the user asking the question B) AI bias originated in human choices — what was written, what was labeled as good or bad, what defaults were selected — and cannot be fully separated from the human systems that produced it C) Humans and AI systems are equally biased on all topics D) AI bias will be automatically corrected as human society becomes less biased
Answer
**B — AI bias originated in human choices — what was written, what was labeled as good or bad, what defaults were selected — and cannot be fully separated from the human systems that produced it.** AI models learn from human-produced data, are labeled by human raters with human preferences, and are shaped by human training choices. The biases they exhibit are not invented by the AI — they are learned from the humans whose outputs, choices, and preferences shaped the training process. This framing matters because it clarifies responsibility: the bias reflects human systems, and addressing it requires changes to those systems as well as to model development practices.Question 15
A colleague argues: "AI bias is a problem for researchers and big tech companies, not for individual professionals using these tools." What is the most complete response?
A) They're right — individual users have no power to change AI bias and should focus only on using available tools B) Individual practitioners face real professional, ethical, and legal exposure from biased AI outputs in consequential contexts, and have available mitigation practices that reduce but cannot eliminate risk — while also having a stake in advocating for better tools C) All AI tools are equally biased, so the choice of tool doesn't matter for individuals D) Bias only matters in AI research contexts, not practical professional applications