Computing Bias MCQ Practice

Question 1 of 10

A resume screener learns from past hires who mostly listed unpaid internships. Applicants who worked paid jobs to support families are filtered out more often.

Which bias source fits best?

  1. A. Bias that requires no training data at all.
  2. B. Paid-job applicants being the majority of past hires in the training set.
  3. C. Training examples that privilege one work-history pattern over others.
  4. D. Equal representation of unpaid and paid work histories in training.

Question 2 of 10

A speech-to-text tool is tested mostly on speakers without accents used in the training region. Speakers with other accents see higher error rates in production.

Who is most clearly mis-served?

  1. A. Only speakers who matched the test accents perfectly and never err.
  2. B. No speakers, because accent never interacts with speech models.
  3. C. Only the microphones, which cannot relate to dataset bias.
  4. D. Speakers whose accents were scarce in training and testing.

Question 3 of 10

A bike-share demand model trains on summer ridership. Winter riders find stations empty in cold months.

Which explanation is best?

  1. A. Seasonal training data left out winter patterns, mis-serving winter riders.
  2. B. Summer data automatically includes accurate winter demand.
  3. C. Empty winter stations prove summer riders were excluded from training.
  4. D. Demand models ignore which seasons appear in datasets.

Question 4 of 10

A content filter is trained on flagged posts from one language. Harmful posts in another language pass more often.

Training flagged posts Language coverage
95 percent Language X
5 percent Language Y

Which conclusion fits the table?

  1. A. Filters cannot reflect language imbalance in training data.
  2. B. Language Y content may be mis-served by weaker filtering because it was scarce in training.
  3. C. Language Y is guaranteed better filtering than Language X.
  4. D. Equal training coverage is shown for both languages.

Question 5 of 10

A university admission model trains on applicants who could afford multiple AP exams. Applicants from schools offering few AP courses score as less prepared.

Which claim is NOT supported?

  1. A. Training features can encode opportunity differences.
  2. B. Who appears in historical admits can shape biased readiness scores.
  3. C. AP-exam counts in training cannot bias against students with fewer course offerings.
  4. D. Students from low-AP schools may be scored as less prepared.

Question 6 of 10

A pedestrian-detection car system trains on daytime city footage. Night rural walkers are detected less reliably.

Which group is poorly represented?

  1. A. Daytime city pedestrians who never appear in training.
  2. B. All lighting conditions equally by the description given.
  3. C. Only parked cars, which detection ignores entirely.
  4. D. Night rural pedestrians underrepresented in training footage.

Question 7 of 10

A mental-health keyword model trains on clinical notes written by specialists. Casual teen forum posts about stress are scored inaccurately.

Who is most clearly mis-served?

  1. A. Teen forum writers whose language style was missing from clinical training notes.
  2. B. Specialists who wrote the notes and always get the worst scores.
  3. C. No group, because keyword models ignore training text style.
  4. D. Only printed textbooks, which the model ranks first always.

Question 8 of 10

A delivery ETA model trains on truck routes in flat cities. Mountain-town customers see chronically late predictions.

Which statement is best supported?

  1. A. Flat-city training guarantees mountain accuracy.
  2. B. Customers in terrains left out of training can be mis-served by bad ETAs.
  3. C. Late mountain ETAs prove mountain towns dominated training.
  4. D. ETA models cannot encode geographic training gaps.

Question 9 of 10

An image search for "nurse" returns mostly one gender because training web images overrepresent that gender in nurse photos.

Which bias pathway fits?

  1. A. Search results ignore training image composition.
  2. B. Gender skew in results proves training had equal gender counts.
  3. C. Skewed training images lead to skewed search results that misrepresent who nurses are.
  4. D. Balanced nurse images in training produced the skew.

Question 10 of 10

A classroom cheating detector trains on laptop keystroke data. Students who write on paper and scan work are flagged as anomalous more often.

Which evaluation is most accurate?

  1. A. Keystroke-only training guarantees fair treatment of paper workflows.
  2. B. Anomaly flags prove paper students dominated the training set.
  3. C. Cheating detectors cannot reflect modality gaps in training data.
  4. D. Paper-based students can be mis-served when training assumes laptop keystrokes.