AI Bias & Fairness Vocabulary
Learn the language of AI fairness: statistical bias, demographic parity, disparate impact, and bias audit vocabulary.
0 / 26 completed
1 / 26
In AI fairness, 'demographic parity' means a model should:
Demographic parity (also called statistical parity) requires that the positive prediction rate is equal across groups — e.g., loan approval rates should be the same across racial groups.
2 / 26
'Disparate impact' in AI systems refers to:
Disparate impact is a legal and ethical concept: a system can be discriminatory in effect (disparate impact) even if it was not designed to discriminate.
3 / 26
What is 'selection bias' in a training dataset?
Selection bias occurs when training data systematically misrepresents the deployment population — for example, a medical model trained on hospital records that over-represents severe cases.
4 / 26
Equalized odds as a fairness metric requires that:
Equalized odds requires both equal TPR (catching real positives) and equal FPR (avoiding false alarms) across groups — stricter than demographic parity alone.
5 / 26
A 'red team' exercise in AI ethics involves:
AI red teaming involves actively trying to elicit harmful, biased, or unsafe outputs from a model to identify risks before deployment — borrowed from security red teaming.
6 / 26
Reviewer: 'I've flagged this recommendation engine. The model consistently suggests higher interest rates to users identifying as 'Hispanic.' This feels… concerning. Can you investigate the data and feature engineering?
(Excerpt from a Slack message during code review)This scenario highlights 'disparate impact,' which isn't merely about reflecting existing societal biases but rather when the algorithm *exacerbates* those biases. The reviewer correctly identifies a potentially harmful outcome – disproportionately higher interest rates for Hispanic users – and emphasizes the need for investigation and remediation. Simply accepting correlation as 'acceptable' ignores the ethical implications of algorithmic amplification.
7 / 26
PR Description: 'Implemented new feature for fraud detection. Model now predicts higher risk scores for transactions originating from IP addresses in countries with historically high rates of fraudulent activity. Data was sourced primarily from publicly available government records.' What is the *primary* source of potential bias in this scenario?Selection bias occurs when the data used to train or evaluate a model isn't representative of the population it will eventually serve. In this case, relying solely on government records, likely skewed by reporting biases and geographic concentration, creates a biased training set. This leads to unfairly high risk scores for transactions originating from those countries, irrespective of individual user behavior. Options A and C are irrelevant to the core issue; option D is an unfounded criticism.
8 / 26
Slack Message: '@john.doe I'm worried about our new chatbot. Initial testing shows it consistently uses more polite and helpful language when interacting with users identified as female. This could inadvertently reinforce gender stereotypes.' What is the most relevant technical concept being discussed?Confirmation bias is closely related here; the chatbot's behavior is reinforcing pre-existing stereotypes by exhibiting different language styles based on gender identification. This demonstrates how a biased system can amplify and perpetuate those biases. While algorithmic transparency and data drift are important considerations, they aren't the primary focus of this Slack conversation regarding biased output.
9 / 26
API Response: `{'metric': 'false_positive_rate', 'group': ['age', 'location'], 'value': 0.25, 'threshold': 0.1}`. This API provides metrics for a loan application model. The 'false_positive_rate' indicates the percentage of applications incorrectly flagged as high-risk. What does this response suggest about potential fairness concerns?A high false positive rate, particularly when combined with specific demographic groups (indicated by 'age' and 'location'), suggests a potential issue of *disparate impact*. This means the model is disproportionately flagging applications from certain demographics as high-risk, even if they are low-risk. Options A and C contradict this finding; option D is irrelevant to the metric itself.
10 / 26
Reviewer: 'I've flagged this recommendation engine. The model consistently suggests higher interest rates to users identifying as 'Hispanic.' This feels… concerning. Can you investigate the data and feature engineering?
(Excerpt from a Slack message during code review)This scenario highlights 'disparate impact,' which isn't merely about reflecting existing societal biases but rather when the algorithm *exacerbates* those biases. The reviewer correctly identifies a potentially harmful outcome – disproportionately higher interest rates for Hispanic users – and emphasizes the need for investigation and remediation. Simply accepting correlation as 'acceptable' ignores the ethical implications of algorithmic amplification.
11 / 26
PR Description: 'Implemented new feature for fraud detection. Model now predicts higher risk scores for transactions originating from IP addresses in countries with historically high rates of fraudulent activity. Data was sourced primarily from publicly available government records.' What is the *primary* source of potential bias in this scenario?Selection bias occurs when the data used to train or evaluate a model isn't representative of the population it will eventually serve. In this case, relying solely on government records, likely skewed by reporting biases and geographic concentration, creates a biased training set. This leads to unfairly high risk scores for transactions originating from those countries, irrespective of individual user behavior. Options A and C are irrelevant to the core issue; option D is an unfounded criticism.
12 / 26
Slack Message: '@john.doe I'm worried about our new chatbot. Initial testing shows it consistently uses more polite and helpful language when interacting with users identified as female. This could inadvertently reinforce gender stereotypes.' What is the most relevant technical concept being discussed?Confirmation bias is closely related here; the chatbot's behavior is reinforcing pre-existing stereotypes by exhibiting different language styles based on gender identification. This demonstrates how a biased system can amplify and perpetuate those biases. While algorithmic transparency and data drift are important considerations, they aren't the primary focus of this Slack conversation regarding biased output.
13 / 26
API Response: `{'metric': 'false_positive_rate', 'group': ['age', 'location'], 'value': 0.25, 'threshold': 0.1}`. This API provides metrics for a loan application model. The 'false_positive_rate' indicates the percentage of applications incorrectly flagged as high-risk. What does this response suggest about potential fairness concerns?A high false positive rate, particularly when combined with specific demographic groups (indicated by 'age' and 'location'), suggests a potential issue of *disparate impact*. This means the model is disproportionately flagging applications from certain demographics as high-risk, even if they are low-risk. Options A and C contradict this finding; option D is irrelevant to the metric itself.
14 / 26
Reviewer: 'I've flagged this recommendation engine. The model consistently suggests higher interest rates to users identifying as 'Hispanic.' This feels… concerning. Can you investigate the data and feature engineering?
(Excerpt from a Slack message during code review)This scenario highlights 'disparate impact,' which isn't merely about reflecting existing societal biases but rather when the algorithm *exacerbates* those biases. The reviewer correctly identifies a potentially harmful outcome – disproportionately higher interest rates for Hispanic users – and emphasizes the need for investigation and remediation. Simply accepting correlation as 'acceptable' ignores the ethical implications of algorithmic amplification.
15 / 26
PR Description: 'Implemented new feature for fraud detection. Model now predicts higher risk scores for transactions originating from IP addresses in countries with historically high rates of fraudulent activity. Data was sourced primarily from publicly available government records.' What is the *primary* source of potential bias in this scenario?Selection bias occurs when the data used to train or evaluate a model isn't representative of the population it will eventually serve. In this case, relying solely on government records, likely skewed by reporting biases and geographic concentration, creates a biased training set. This leads to unfairly high risk scores for transactions originating from those countries, irrespective of individual user behavior. Options A and C are irrelevant to the core issue; option D is an unfounded criticism.
16 / 26
Slack Message: '@john.doe I'm worried about our new chatbot. Initial testing shows it consistently uses more polite and helpful language when interacting with users identified as female. This could inadvertently reinforce gender stereotypes.' What is the most relevant technical concept being discussed?Confirmation bias is closely related here; the chatbot's behavior is reinforcing pre-existing stereotypes by exhibiting different language styles based on gender identification. This demonstrates how a biased system can amplify and perpetuate those biases. While algorithmic transparency and data drift are important considerations, they aren't the primary focus of this Slack conversation regarding biased output.
17 / 26
API Response: `{'metric': 'false_positive_rate', 'group': ['age', 'location'], 'value': 0.25, 'threshold': 0.1}`. This API provides metrics for a loan application model. The 'false_positive_rate' indicates the percentage of applications incorrectly flagged as high-risk. What does this response suggest about potential fairness concerns?A high false positive rate, particularly when combined with specific demographic groups (indicated by 'age' and 'location'), suggests a potential issue of *disparate impact*. This means the model is disproportionately flagging applications from certain demographics as high-risk, even if they are low-risk. Options A and C contradict this finding; option D is irrelevant to the metric itself.
18 / 26
Reviewer: 'I've flagged this recommendation engine. The model consistently suggests higher interest rates to users identifying as 'Hispanic.' This feels… concerning. Can you investigate the data and feature engineering?
(Excerpt from a Slack message during code review)This scenario highlights 'disparate impact,' which isn't merely about reflecting existing societal biases but rather when the algorithm *exacerbates* those biases. The reviewer correctly identifies a potentially harmful outcome – disproportionately higher interest rates for Hispanic users – and emphasizes the need for investigation and remediation. Simply accepting correlation as 'acceptable' ignores the ethical implications of algorithmic amplification.
19 / 26
PR Description: 'Implemented new feature for fraud detection. Model now predicts higher risk scores for transactions originating from IP addresses in countries with historically high rates of fraudulent activity. Data was sourced primarily from publicly available government records.' What is the *primary* source of potential bias in this scenario?Selection bias occurs when the data used to train or evaluate a model isn't representative of the population it will eventually serve. In this case, relying solely on government records, likely skewed by reporting biases and geographic concentration, creates a biased training set. This leads to unfairly high risk scores for transactions originating from those countries, irrespective of individual user behavior. Options A and C are irrelevant to the core issue; option D is an unfounded criticism.
20 / 26
Slack Message: '@john.doe I'm worried about our new chatbot. Initial testing shows it consistently uses more polite and helpful language when interacting with users identified as female. This could inadvertently reinforce gender stereotypes.' What is the most relevant technical concept being discussed?Confirmation bias is closely related here; the chatbot's behavior is reinforcing pre-existing stereotypes by exhibiting different language styles based on gender identification. This demonstrates how a biased system can amplify and perpetuate those biases. While algorithmic transparency and data drift are important considerations, they aren't the primary focus of this Slack conversation regarding biased output.
21 / 26
API Response: `{'metric': 'false_positive_rate', 'group': ['age', 'location'], 'value': 0.25, 'threshold': 0.1}`. This API provides metrics for a loan application model. The 'false_positive_rate' indicates the percentage of applications incorrectly flagged as high-risk. What does this response suggest about potential fairness concerns?A high false positive rate, particularly when combined with specific demographic groups (indicated by 'age' and 'location'), suggests a potential issue of *disparate impact*. This means the model is disproportionately flagging applications from certain demographics as high-risk, even if they are low-risk. Options A and C contradict this finding; option D is irrelevant to the metric itself.
22 / 26
Reviewer: 'I've flagged this risk assessment model. The API consistently returns a higher false positive rate for loan applications submitted by users with names containing traditionally Black surnames. This suggests potential bias in the feature engineering related to name recognition. Can you investigate and ensure fairness?'
The reviewer correctly identifies a key concern: disparate outcomes based on protected characteristics. A simple reduction of the overall error rate doesn't guarantee fairness. Focusing solely on the false positive rate overlooks potential disparities in other metrics like approval rates. Investigating the name recognition feature is critical to understanding and mitigating bias.
23 / 26
@alice.smith: 'I'm seeing a pattern in our credit scoring model's predictions. Users with zip codes predominantly associated with lower-income neighborhoods consistently receive lower credit scores than similar users in higher-income areas – even after accounting for income level. This could be driven by biased data reflecting historical socioeconomic disparities. I need to explore whether we are inadvertently perpetuating these inequalities.'
The Slack message highlights a crucial concern: a biased prediction based on geographic location (zip code). While zip codes are correlated with socioeconomic factors, simply removing the feature isn't sufficient; the underlying bias in the data needs investigation. The core issue is the *correlation* between zip code and credit scores, potentially reflecting historical inequalities.
24 / 26
{'metric': 'disparate_impact', 'group': ['gender', 'ethnicity'], 'value': 0.75, 'threshold': 0.8}. This API provides metrics for a recruitment tool. The 'disparate impact' measures the ratio of selected candidates from different groups. A value of 0.75 indicates… A 'disparate impact' ratio of 0.75 suggests a slight bias towards one group (the higher number). A threshold of 0.8 indicates that if the ratio exceeds this value, it would be considered statistically significant evidence of discrimination. This warrants further investigation to ensure equitable outcomes.
25 / 26
'Implemented a new recommendation engine for e-commerce products. The model uses collaborative filtering based on user purchase history to suggest items. Initial analysis shows that users purchasing high-end electronics are consistently recommended luxury goods and accessories, while users buying budget appliances receive recommendations for basic tools. This could be unintentionally reinforcing stereotypes about consumer preferences. I've added a fairness constraint to the recommendation algorithm…'
The PR description highlights a potentially problematic bias: recommending luxury goods to high-end electronics purchasers and basic tools to budget appliance buyers. This could reinforce stereotypes about consumer preferences. Adding a fairness constraint demonstrates an awareness of this potential issue and is a necessary step towards mitigating bias in the recommendation algorithm.
26 / 26
'Just finished testing our fraud detection model. It's achieving a 92% accuracy rate – that's a significant improvement! However, I noticed it flags transactions from users in Southeast Asia at a much higher rate than those from North America, even when accounting for transaction volume. This could indicate a biased training dataset. Need to investigate the data sources and feature engineering.'
Even with a high accuracy rate (92%), disproportionate flagging of transactions from Southeast Asia raises serious concerns. This indicates potential bias in the training data or feature engineering – it's crucial to investigate regardless of overall performance metrics. Ignoring this issue would lead to unfair outcomes for users in that region.
Frequently Asked Questions
What will I practice in "AI Bias & Fairness Vocabulary"?
This is an AI Ethics exercise set. It walks through 26 scenario-based multiple-choice questions built around real usage of AI Ethics terminology that IT professionals encounter on the job.
Is this exercise free to use?
Yes. Every exercise on CoderSlingo, including this one, is free to complete with no account, sign-up, or paywall.
How many questions are in this exercise?
This set contains 26 questions. Each one shows immediate feedback and a detailed explanation after you answer, so you learn the correct usage right away rather than waiting for a final score.
Do I need prior experience to complete this exercise?
No prior experience is required. Each question includes a full explanation covering the reasoning behind the correct answer, so the exercise itself teaches the AI Ethics vocabulary as you go.
Can I retry the exercise if I get questions wrong?
Yes — use the "Try again" button on the results screen to reset your answers and go through all the questions again. There is no limit on attempts.
Is my progress saved?
Your answers and score for the current session are tracked in the browser as you go. No account or login is needed, and there is nothing to install.
What if I don't understand a term used in a question?
Read the explanation shown after you answer each question — it breaks down the correct term in plain English with a real-world example. You can also check the site Glossary for quick definitions.
How is this different from reading a blog article on the topic?
Exercises like this one are interactive drills that test and reinforce specific vocabulary through multiple-choice questions, while blog articles explain concepts in prose. Practising here after reading builds active recall, not just passive recognition.
Where can I find more AI Ethics exercises?
See the AI Ethics exercises hub for the full set of related pages, or browse all exercise categories from the main Exercises index.
Can I use this exercise to prepare for a technical interview?
Yes — AI Ethics vocabulary comes up often in technical discussions and interviews. Pair this exercise with our dedicated Interview Preparation section for role-specific practice.