Understanding Ai Bias and Fairness

• A. Overfitting
• B. AI bias ✓
• C. Latency
• D. Model drift

Why: AI bias refers to systematic, unfair errors in outputs that favor or disadvantage particular groups. Overfitting, latency, and model drift are different concepts related to model performance rather than fairness.

• A. A faster processor
• B. Training data that underrepresents certain groups
• C. Using a larger screen for display
• D. Storing models in the cloud

• A. Bias introduced when a model is compressed for deployment
• B. Bias that exists in the real world and is reflected in data, even when the data is accurately collected
• C. Bias caused by outdated software libraries
• D. Bias from using an old programming language

• A. The camera hardware is randomly faulty
• B. The training dataset contained mostly lighter-skinned faces
• C. Darker-skinned faces are harder to store in memory
• D. The system uses too many layers in its neural network

• A. Because overall accuracy can hide large performance gaps for smaller or underrepresented groups
• B. Because demographic data makes models run faster
• C. Because it is required to reduce the file size of the model
• D. Because it eliminates the need for testing altogether

• A. Measurement bias
• B. Sampling bias ✓
• C. Confirmation bias
• D. Automation bias

Why: Sampling bias occurs when the training data is collected in a way that over- or under-represents certain groups, so the model performs poorly for the underrepresented groups when deployed on the broader population.

• A. Label noise
• B. Data leakage
• C. Representation bias ✓
• D. Overfitting

Why: Representation bias happens when certain groups are underrepresented in the training data, causing the model to learn patterns that work well for the majority group but poorly for others.

• A. Using a proxy such as arrest records to stand in for actual criminal behavior, when arrests reflect policing patterns as well as crime ✓
• B. Randomly splitting data into training and test sets
• C. Collecting survey responses from a large, diverse group of participants
• D. Normalizing numerical features before training a model

Why: Measurement bias occurs when the variable used to capture a concept is a flawed or systematically skewed stand-in for it. Arrest records reflect where police patrol as well as actual offending, so they distort what they are meant to measure.

• A. Historical bias, where the data accurately reflects past discriminatory or unequal patterns ✓
• B. Random noise in the dataset
• C. Insufficient computing power during training
• D. A bug in the model's code

Why: Historical bias arises when data faithfully records a world that was itself unequal. Even perfectly collected data can encode past discrimination, and models trained on it will tend to reproduce that pattern.

• A. Because collection methods determine who is included or excluded, and hidden gaps there can produce biased outcomes that no later modeling step can fully fix ✓
• B. Because data collection only affects the model's speed, not its accuracy
• C. Because collected data is always automatically balanced across groups
• D. Because data collection methods are only relevant for legal compliance

Why: Who gets into a dataset and how their information is recorded is decided at collection time. If certain groups are missing or measured poorly, the model inherits those gaps, and downstream fixes often cannot fully recover the missing information.

• A. Measure how quickly a model produces predictions
• B. Quantify whether a model's outcomes or errors differ across groups defined by sensitive attributes ✓
• C. Determine the accuracy of a model on its training data only
• D. Calculate the total number of features used by a model

Why: Fairness metrics quantify whether a model treats different demographic or social groups equitably, typically by comparing outcomes or error rates across groups defined by attributes such as race, gender, or age.

• A. Individual fairness
• B. Calibration
• C. Demographic parity ✓
• D. Predictive accuracy

Why: Demographic parity requires that the rate of positive outcomes be roughly equal across groups. A 40% versus 20% selection rate shows a large gap, violating demographic parity.

• A. Equal overall accuracy across all groups
• B. Equal true positive rates across groups ✓
• C. Equal number of positive predictions across groups
• D. Identical feature distributions for each group

Why: Equal opportunity requires that qualified individuals (true positives) be identified at the same rate across groups, meaning the true positive rate should be equal for each group.

• A. Calibration requires the model to ignore group membership entirely
• B. Calibration only ensures predicted probabilities match observed frequencies, not that error rates are balanced across groups ✓
• C. Calibration measures only computational efficiency
• D. Calibration is only applicable to regression models

Why: Calibration means that a predicted score of, say, 0.7 corresponds to a 70% observed rate within each group. This does not guarantee equal false positive or false negative rates across groups, so other fairness criteria can still be violated.

• A. All fairness metrics can be satisfied simultaneously for any dataset
• B. Fairness metrics are always identical in practice
• C. When base rates differ across groups, several common fairness criteria cannot all be satisfied at once unless the classifier is perfect ✓
• D. Fairness can only be measured after a model is deployed

Why: Results by Kleinberg, Mullainathan, and Chouldechova show that calibration, equal false positive rates, and equal false negative rates cannot all hold together when base rates differ between groups, except in trivial cases such as a perfect predictor. This forces practitioners to make value-laden trade-offs.

• A. Equalized odds
• B. Demographic parity ✓
• C. Calibration
• D. Predictive parity

Why: Demographic parity (also called statistical parity) requires that the proportion of positive predictions be equal across groups, regardless of the actual outcome rates. Equalized odds and predictive parity instead condition on the true outcome or the predicted score.

• A. Overall accuracy only
• B. Positive predictive value only
• C. True positive rates and false positive rates ✓
• D. The base rate of the outcome in each group

Why: Equalized odds requires both the true positive rate and the false positive rate to be the same across protected groups, meaning errors are distributed equally regardless of group membership.

• A. Accuracy and speed
• B. Calibration, predictive parity, and equalized odds ✓
• C. Demographic parity and privacy
• D. Transparency and interpretability

Why: Impossibility results (e.g., Kleinberg et al. and Chouldechova) show that calibration within groups, balance of positive and negative predictive values, and equal error rates cannot all hold together when base rates differ and the predictor is not perfect.

• A. Demographic parity always improves model accuracy
• B. Demographic parity ensures the model uses no features at all
• C. Demographic parity focuses on equal selection rates, which may be desired to address historical underrepresentation, even if qualification rates differ ✓
• D. Demographic parity eliminates the need to collect demographic data

Why: Demographic parity targets equal selection rates across groups, which can be chosen to correct for historical underrepresentation. Equalized odds instead ties fairness to the true qualification outcome, which may preserve existing disparities if those reflect biased historical data.

• A. Demographic parity
• B. Calibration within groups ✓
• C. Equal opportunity
• D. Individual fairness

Why: Calibration within groups requires that a given predicted score means the same probability of the outcome regardless of group. Here, the same score of 0.8 corresponds to different actual default rates across groups, so calibration is violated.

• A. A deliberate design choice by engineers to favor one group over another
• B. Systematic and repeatable errors that produce unfair outcomes for certain groups of people ✓
• C. Random errors that occur when a model is trained on too little data
• D. A hardware failure that causes incorrect predictions in deployed systems

Why: Algorithmic bias refers to systematic, repeatable errors in an AI system that create unfair outcomes, often disadvantaging particular groups. It is not necessarily intentional, and it is not the same as random error or hardware failure.

• A. The model was trained on a dataset that underrepresented darker-skinned faces ✓
• B. The camera hardware is incapable of capturing darker skin tones
• C. Darker-skinned users tend to move more during image capture
• D. The algorithm uses a random number generator that favors lighter tones

Why: Unrepresentative training data is one of the most common sources of bias. When darker-skinned faces are underrepresented in training data, the model learns patterns that generalize poorly to those faces. Documented cases, such as the Gender Shades study, showed this effect in commercial systems.

• A. Measurement bias
• B. Sampling bias ✓
• C. Confirmation bias
• D. Automation bias

Why: Sampling bias occurs when the training sample is not representative of the target population, so the model learns an incomplete picture. Confirmation bias and automation bias are human cognitive biases, and measurement bias concerns how a variable is recorded rather than who is included.

• A. The algorithm will automatically correct for past discrimination
• B. The algorithm may learn to favor male candidates by reproducing past patterns ✓
• C. The algorithm will become neutral because it has no access to gender information
• D. The algorithm will favor female candidates to balance the historical record

Why: Models trained on historical decisions learn the patterns in those decisions, including discriminatory ones. Even without an explicit gender field, proxies such as certain schools or activities can let the model reproduce the same bias. Automatic correction does not happen by default.

• A. Measuring only overall accuracy across the full test set
• B. Evaluating performance separately across different demographic groups ✓
• C. Increasing the model's size so it captures more patterns
• D. Removing the test set to avoid overfitting to it

Why: Overall accuracy can hide large performance gaps between groups. Disaggregated evaluation, which compares error rates and outcomes across demographic subgroups, is a key way to surface bias before deployment. Model size and removing the test set do not address fairness.

• A. The random noise present in a training dataset
• B. Systematic errors in model outputs that unfairly disadvantage certain groups of people ✓
• C. The process of selecting the best hyperparameters for a model
• D. The difference between training and test accuracy scores

Why: In fairness discussions, bias refers to systematic, repeatable errors that produce unfair outcomes for particular groups, rather than random error or routine modeling choices.

• A. Measurement bias
• B. Historical bias ✓
• C. Overfitting bias
• D. Confirmation bias

Why: Historical bias arises when training data reflects past inequalities or practices. Here, the model learns patterns from a historically male-dominated workforce, even if the data is accurately recorded.

• A. Regularization
• B. Dropout
• C. Resampling or reweighting ✓
• D. Early stopping

Why: Resampling (oversampling underrepresented groups or undersampling overrepresented ones) and reweighting (assigning higher weights to underrepresented examples) are pre-processing methods that address imbalance in the training data.

• A. It always reduces the model's overall accuracy to zero
• B. Other features, such as zip code, can act as proxies that still encode the sensitive information ✓
• C. It makes the model impossible to train with gradient descent
• D. It requires the model to be retrained from scratch every day

Why: This approach, often called 'fairness through unawareness,' fails because correlated features like zip code, education, or purchase history can reconstruct protected attributes and reproduce the same bias.

• A. Equalized odds
• B. Demographic parity ✓
• C. Calibration
• D. Predictive parity

Why: Demographic parity (also called statistical parity) requires that the proportion of positive predictions be equal across groups, regardless of actual outcomes. Equalized odds and predictive parity involve error rates or outcome conditioning instead.

• A. To maximize the speed at which AI models are deployed to production
• B. To evaluate AI systems for potential harms, bias, and societal impact before and during deployment ✓
• C. To reduce the cost of training data by outsourcing annotation work
• D. To ensure AI models achieve the highest possible accuracy on benchmark tests

Why: An AI ethics review board examines systems for potential harms, bias, fairness concerns, and broader societal impact, providing oversight beyond technical performance metrics. Speed, cost, and benchmark accuracy may matter, but they are not the board's core function.

• A. AI systems being able to explain every decision in plain language
• B. Clearly identifying who is responsible for an AI system's outcomes and having mechanisms to address harm ✓
• C. AI systems automatically correcting their own errors without human involvement
• D. Companies publishing their model architecture publicly

Why: Accountability means that specific people or organizations are answerable for how an AI system behaves, and that there are processes for redress when it causes harm. Explainability and transparency support accountability but are not the same thing.

• A. To ban all artificial intelligence applications within the European Union
• B. To apply regulatory requirements proportionate to the level of risk an AI system poses ✓
• C. To require every AI system to be open source
• D. To give AI systems the same legal status as human workers

Why: The EU AI Act categorizes AI systems by risk level, from unacceptable (prohibited) to high, limited, and minimal risk, with stricter obligations for higher-risk uses. It does not ban all AI, require open source, or grant AI legal personhood.

• A. The algorithm will run too slowly to be useful for recruiters
• B. The algorithm may reproduce and amplify historical bias, disadvantaging qualified candidates from other backgrounds ✓
• C. The algorithm violates copyright because it uses university data
• D. The algorithm will always select candidates with the highest test scores

Why: When training data reflects past discriminatory or biased human decisions, a model learns those patterns and can perpetuate them at scale. This is a central fairness concern in AI ethics and a reason for auditing training data and outcomes.

• A. Because deployed models automatically become less accurate and must be discarded
• B. Because data distributions and user populations can shift over time, potentially causing new or unexpected biases and failures ✓
• C. Because legal regulations require audits only once, at the time of product launch
• D. Because audits are needed to prevent the model from learning new information

Why: Real-world conditions change: the people using a system, the data it sees, and the social context can all shift after launch. Ongoing monitoring and auditing help detect emerging bias, performance drops, and harms that were not visible during initial testing.