Bias in data and AI quiz

Question 1

What is data bias in AI?

Accepted Answer

When training data does not fairly represent the real world

Answer

When an AI crashes because of a software bug

Answer

When a model uses too many parameters

Answer

When an AI learns faster than expected

Question 2

What does 'training data' mean in machine learning?

Accepted Answer

Examples that a machine learning model learns patterns from

Answer

Data stored only for final testing, never seen during training

Answer

Data the model generates automatically after being deployed

Answer

Data collected from social media platforms for analysis

Question 3

What does 'garbage in, garbage out' mean in AI?

Accepted Answer

Poor-quality input data leads to poor-quality or biased output

Answer

AI models automatically fix low-quality data during training

Answer

Garbage collection is a memory management technique in programming

Answer

Computers produce random outputs regardless of input quality

Question 4

If a face recognition system is trained only on photos of people with light skin, what problem is likely to occur?

Accepted Answer

The system will perform worse on people with darker skin tones

Answer

The system will treat all skin tones equally since the algorithm is neutral

Answer

The system will stop working entirely when it sees new images

Answer

The system will run slower when processing images of people with darker skin

Question 5

What is an imbalanced dataset?

Accepted Answer

A dataset where one category has far more examples than another

Answer

A dataset where images are stored in different file formats

Answer

A dataset that is split evenly between training and test sets

Answer

A dataset distributed across multiple storage servers

Question 6

What is a 'label' in machine learning, and how can biased labeling cause problems?

Accepted Answer

The correct answer or category attached to a training example by a human annotator

Answer

A decorative tag shown to users in the final app interface

Answer

An automatic output generated by the AI after training completes

Answer

A code comment explaining the purpose of a machine learning function

Question 7

Can simply removing columns like 'race' or 'gender' from a dataset always prevent AI bias?

Accepted Answer

No, because other variables called proxies can still encode the removed information

Answer

Yes, removing sensitive columns always makes a model completely fair

Answer

Yes, but only if the model is also retrained from scratch after removing the columns

Answer

No, because the model can access deleted columns through hidden backup files

Question 8

What is automation bias?

Accepted Answer

The tendency for humans to over-trust automated systems, even when those systems are wrong

Answer

Bias that is automatically removed from data during preprocessing steps

Answer

An AI system that deliberately gives incorrect answers

Answer

Errors that appear only when automation software receives an update

Question 9

What is the key difference between an AI making a random mistake and an AI being biased?

Accepted Answer

Bias causes systematic errors that consistently disadvantage certain groups, unlike random errors that affect everyone equally

Answer

There is no real difference; all AI errors are caused by bias

Answer

Biased errors are random, while other mistakes follow a predictable pattern

Answer

Random mistakes are more serious because they are harder to predict or prevent

Question 10

Why does having diverse training data matter for building fair AI?

Accepted Answer

Diverse training data helps AI work accurately and fairly for all the groups it will encounter

Answer

Diverse data makes AI models train faster but reduces their overall accuracy

Answer

Data diversity only matters for image-recognition tasks, not other types of AI

Answer

Diversity in data is not important as long as the algorithm is powerful enough

Question 11

What is selection bias in data collection?

Accepted Answer

When collected data systematically over- or under-represents certain groups or situations

Answer

When a programmer deliberately removes data to make test results look better

Answer

When a model randomly selects which data points to learn from during training

Answer

When users choose which AI features to activate in their personal settings

Question 12

What is historical bias in AI, and why is it hard to fix?

Accepted Answer

Bias that arises when training data reflects past societal inequalities, which the model then perpetuates

Answer

Bias caused by using outdated computer hardware or old software libraries

Answer

Bias that only affects AI models making predictions about past historical events

Answer

Bias introduced when AI developers use history books to design their systems

Question 13

What is a feedback loop in AI bias, and why can it be dangerous?

Accepted Answer

A cycle where biased AI decisions generate new biased data, making future models even more biased

Answer

A safety mechanism that stops an AI from making the same error twice in a row

Answer

A programming technique used to iterate efficiently through large training datasets

Answer

An error that appears when two AI systems exchange information in a network

Question 14

What is representation bias in AI?

Accepted Answer

When certain groups do not have enough examples in the training dataset for the AI to learn about them fairly

Answer

When an AI model is too large to represent data efficiently in computer memory

Answer

When data is stored as numbers instead of as images or text

Answer

When a politician uses AI to misrepresent survey or polling data

Question 15

What is measurement bias in AI?

Accepted Answer

When the data collection method consistently measures inaccurately for certain groups

Answer

When a ruler or scale gives wrong readings during a school science experiment

Answer

When a machine learning evaluation metric like accuracy is computed incorrectly

Answer

When a dataset is too small to produce statistically significant measurements

Question 16

What is a proxy variable, and how can it sneak bias into an AI model?

Accepted Answer

A variable correlated with a sensitive attribute that introduces indirect discrimination

Answer

A variable used to protect sensitive user data from being seen by the model

Answer

A variable used only during model testing and removed before production

Answer

A variable that only ever takes the binary values zero or one

Question 17

What is confirmation bias in AI development?

Accepted Answer

When developers unconsciously interpret results to confirm pre-existing beliefs, missing real failures

Answer

When a model only confirms user inputs without generating new suggestions

Answer

When an independent auditing body certifies a model as accurate and unbiased

Answer

When two different AI models always produce identical outputs on the same input

Question 18

What is gender bias in Natural Language Processing (NLP)?

Accepted Answer

When NLP models associate words and roles with particular genders in stereotyped ways, learned from training text

Answer

A programming bug that causes NLP models to only accept text written by one gender

Answer

A deliberate design choice to make AI voice assistants sound more masculine

Answer

When an NLP system refuses to process text that mentions gender-related topics

Question 19

How can a biased AI hiring tool harm job seekers from certain groups?

Accepted Answer

A biased hiring AI may systematically rank qualified candidates from certain groups lower, preventing them from getting interviews

Answer

A biased hiring AI will always crash when processing resumes from certain demographic groups

Answer

A biased hiring tool is immediately spotted and corrected by experienced human recruiters

Answer

A biased hiring AI only affects candidates who apply through online portals, not those who apply in person

Question 20

What is the difference between a model being accurate and a model being fair?

Accepted Answer

A model can be highly accurate overall but still perform poorly and unfairly for certain subgroups

Answer

Accuracy and fairness mean exactly the same thing in machine learning

Answer

A fair model is always less accurate than an unfair model due to the constraints imposed

Answer

Accuracy is a technical measure while fairness is philosophical and cannot be measured at all

Question 21

How can you use Python to quickly check whether a dataset is imbalanced?

Accepted Answer

Use df["label"].value_counts() to count how many examples belong to each class

Answer

Use df.head() to display the first five rows and visually spot class imbalances

Answer

Use df.describe() to get statistical summaries that reveal class distribution

Answer

Use df.isnull().sum() to count missing values as a proxy for imbalance

Question 22

What is societal bias and how does it end up inside an AI system?

Accepted Answer

Bias that enters AI when training data reflects existing human prejudices and social inequalities

Answer

Bias deliberately inserted into AI systems by social media companies for profit

Answer

Bias that only exists in social networking applications and chatbots

Answer

The tendency of AI systems to prefer social media data over data from other sources

Question 23

Why does having a diverse team of AI developers help reduce bias in AI systems?

Accepted Answer

Diverse teams bring different lived experiences that help identify edge cases a less diverse team might overlook

Answer

Diverse teams code faster, which reduces the time for bias to creep into the project

Answer

Diverse teams use different programming languages, making the model more robust

Answer

Diverse teams only matter for marketing and user-interface design, not technical model quality

Question 24

What is the 'filter bubble' or 'echo chamber' effect caused by recommendation algorithms?

Accepted Answer

When recommendation algorithms show users mainly content matching their existing views, filtering out diverse perspectives

Answer

When a recommendation system crashes and shows the same content in an endless loop

Answer

When noise in training data causes the algorithm to make completely random recommendations

Answer

When the algorithm runs out of new content and starts repeating old recommendations

Question 25

What is the difference between statistical bias and ethical/social bias in AI?

Accepted Answer

Statistical bias is a mathematical deviation from true values; ethical bias is unfair treatment of people — and the two are distinct concepts

Answer

Statistical bias is always more harmful than ethical bias because it affects prediction accuracy

Answer

Statistical bias and ethical bias are the same concept described in different academic fields

Answer

Statistical bias only appears in large datasets while ethical bias only appears in small datasets

Question 26

What is disparate impact, and how is it used to measure algorithmic fairness?

Accepted Answer

A legal and statistical concept where a neutral-seeming system produces outcomes that disproportionately harm a protected group, measured by comparing outcome rates across groups

Answer

A measure of how differently an AI model performs when run on different hardware configurations

Answer

The gap between a model's accuracy on the training set versus its accuracy on the test set

Answer

A concept that applies only to AI systems used in criminal justice or law enforcement

Question 27

What is a confusion matrix, and how can it reveal hidden bias in a classification model?

Accepted Answer

A table of true/false positives and negatives that reveals when error rates differ significantly across demographic groups

Answer

A diagram that maps all the different types of bias found in a dataset

Answer

A matrix used to calculate statistical correlations between input features

Answer

A tool that shows which input features confused the model the most during training

Question 28

What is data augmentation, and how can it help reduce bias caused by underrepresentation?

Accepted Answer

Creating modified versions of existing training examples to increase variety and balance underrepresented groups

Answer

Buying additional labeled data from third-party vendors to increase the dataset size

Answer

Splitting a dataset into training and test portions to evaluate and prevent overfitting

Answer

Removing duplicate rows from a dataset to improve the quality of training data

Question 29

What is word embedding bias, and how was it famously demonstrated?

Accepted Answer

Stereotyped associations encoded in word vectors learned from text, such as linking 'doctor' with 'man' and 'nurse' with 'woman'

Answer

A bias that stops the model from representing words from minority or low-resource languages

Answer

Bias caused by spelling errors and typos present in the training text corpus

Answer

Bias that only appears when the model's vocabulary exceeds one million words

Question 30

Why is it mathematically impossible to satisfy all common fairness definitions at the same time when a model is imperfect?

Accepted Answer

Mathematical proofs show that key fairness metrics such as calibration and equal error rates cannot all hold simultaneously unless predictions are perfect or base rates are equal

Answer

It is only impossible if the model uses deep learning; simpler models can satisfy all fairness definitions at once

Answer

All fairness conditions are automatically satisfied when both training and test accuracy exceed 95%

Answer

Fairness definitions cannot all be satisfied because AI can never be completely objective about any topic

Test yourself