MPPSC 2022 Prelims Paper-I · Question 50 of 100

Correct — D, Ability. The widely-recognized sources of bias in AI systems are: Data (biased or unrepresentative training data), Algorithm (biased model design or optimization choices), and People (human bias introduced by developers, labelers, or users). 'Ability' is not part of this standard taxonomy of AI-bias sources.

• (a)Data — Biased, incomplete, or unrepresentative training data is one of the most common sources of AI bias — e.g., under-representing certain groups.
• (b)Algorithm — The way a model is designed, weighted, or optimized can itself introduce systematic bias, independent of the data used.
• (c)People — The humans who collect, label, and interpret data — and who use the AI's output — can introduce their own conscious or unconscious biases.

AI bias refers to a systematic, unfair skew in an AI system's outputs. It is commonly traced to three sources: biased training Data, biased Algorithm design, and biased People (the developers, labelers, or users involved in building and operating the system). Recognizing all three is central to designing fairer, more accountable AI.

The exam trap tests whether the candidate can distinguish the genuine, commonly-cited sources of AI bias (Data / Algorithm / People) from an unrelated but plausible-sounding term ('Ability') that is not part of the standard framework.

• Three commonly-cited sources of AI bias: Data, Algorithm, People
• Data bias — training data that under- or over-represents certain groups or scenarios
• Algorithmic bias — bias introduced through model design or optimization choices
• People/human bias — bias introduced by developers, labelers, or end users

• Confusing the statistical 'bias-variance tradeoff' (a technical ML term) with 'AI bias' (a fairness/ethics term) — they are different concepts
• Picking a plausible-sounding but non-standard term like 'Ability' as if it were a recognized bias source

An odd-one-out on AI-ethics fundamentals — an increasingly common 'current affairs meets tech' question in MPPSC and other state PCS papers.

No directly related past PYQ was found.

• practice — not a real PYQ

  Bias in an AI system arising from training data that under-represents certain groups is called:

  • (a)Algorithmic bias
  • (b)Data bias
  • (c)Confirmation bias
  • (d)Sampling variance

  Answer(b) Data bias — arises from unrepresentative or skewed training data.

• practice — not a real PYQ

  Which of the following is a recognized measure to reduce AI bias?

  • (a)Using diverse and representative training datasets
  • (b)Reducing the size of the model
  • (c)Increasing processing speed
  • (d)Removing all human oversight

  Answer(a) Using diverse and representative training datasets — directly addresses data-source bias.