AIGP Study Guide
Module 2: AI Impacts & Responsible AI · BoK II.A

Individual harms and the anatomy of bias

Individual harms hit civil liberties, safety or economic opportunity, and bias is the engine. Know Implicit bias, Sampling bias and Temporal bias on sight, plus the failure modes Overfitting, Underfitting and Edge cases & outliers - and remember not all bias is harmful or illegal.

Individual harms hit a person's civil liberties, rights, physical or psychological safety, or economic opportunity. Bias is the engine behind most of them. Humans configure AI systems, so the systems mirror human biases, morals and values.

Three bias types to name on sight:

  • 🫥 Implicit biasunconscious discrimination stemming from biased data or developer assumptions, harming fairness and equity.
  • 🧪 Sampling bias → training data skews toward a subset, favouring specific groups → inaccurate or unfair outcomes.
  • Temporal bias → models trained on current data fail to adapt to future changes → accuracy decays over time.

Three model failure modes:

  • 🎯 Overfitting → learns the training examples too specifically → weak on new, unseen data.
  • 📉 Underfitting → model is too simple, misses important patterns → poor performance.
  • 🦄 Edge cases & outliers → rare points outside the normal range cause mistakes the model was never trained to handle.
Exam flash: legal vs illegal bias

Not all bias is harmful or illegal. Lending less based on income may be an allowed bias. Lending less based on race is not. The skill is spotting whether a bias is illegal or undesirable, then managing it.

Real-world bias cases the course names
CategoryExampleImpact
Employment biasAmazon's AI hiring toolDiscriminated against women due to biased training data
Facial recognitionLondon police AI system81% inaccuracy rate caused biased policing
Economic discriminationAI in financial lendingLoans denied based on economic background
Privacy violationsAI using social media dataData used without proper consent for training
Education accessAI school selection systemsStudents excluded by biased algorithms

Civil rights & privacy concerns:

  • 🧬 Reidentification → massive multi-source datasets make it easier to recombine personal data and compromise privacy, with weak safeguards against unauthorised access.
  • 📥 Appropriation → people consented to data use for specific purposes, not for training AI systems → an ethics-of-consent problem.
  • 🔮 Faulty inference → AI predicts and identifies, but accuracy is unreliable → data can be attributed to the wrong individual, causing harm.
  • 🪟 Opacity & accuracy → people should be notified when AI is used (e.g. chatbots); models are only as good as their training data → data accuracy is vital.

Key terms - quick answers

What is “Implicit bias”?
Unconscious discrimination stemming from biased data or developer assumptions.
What is “Sampling bias”?
Training data skews toward a subset, favouring specific groups and giving unfair outcomes.
What is “Temporal bias”?
Models trained on current data fail to adapt to future changes, so accuracy decays over time.
What is “Overfitting”?
Model learns the training examples too specifically and performs weakly on new, unseen data.