1. What is the difference between narrow AI and general AI?○ Narrow AI runs on small devices; general AI runs in data centres○ Narrow AI performs one task or a set of related tasks and is what exists today; general AI would have broad, flexible human-level competence across arbitrary tasks and does not exist○ Narrow AI uses classical methods; general AI uses deep learning○ They are marketing terms with no technical meaning
2. Why does the narrow-versus-general distinction matter practically?○ Because general AI is regulated differently○ Because most public discussion of AI risk concerns general AI, while the real organisational risks come from narrow AI — biased filters, silently degrading models, confidently wrong assistants○ Because narrow AI cannot be biased○ Because general AI is expected on a known timeline
3. Which of these routes allow bias to enter a machine learning system? Select all that apply.☐ Historical decisions in the training data being skewed☐ A group being thinly represented in the training data☐ The choice of what to predict encoding a judgement☐ Deploying a model on a population different from the one it was built for
4. Amazon's experimental recruiting tool penalised résumés containing the word "women's" despite never being given applicants' gender. What does this demonstrate?○ That the engineers deliberately encoded a preference○ That models find proxies — signals correlated with a protected characteristic — so removing the sensitive field does not remove the bias○ That recruitment cannot be assisted by machine learning○ That the model needed more training data
5. The Gender Shades study found error rates up to 34.7% for darker-skinned women against a maximum of 0.8% for lighter-skinned men. What is the transferable lesson?○ Facial analysis should never be deployed commercially○ An aggregate accuracy figure can conceal near-total failure for a subgroup, so accuracy must be reported broken down by group○ Commercial systems are always less accurate than research systems○ Error rates always improve with newer model versions
6. What makes a bias feedback loop particularly dangerous?○ It causes the model to crash under load○ When a model's decisions determine the data it later learns from, small asymmetries compound with each retraining — while appearing as improving accuracy on the narrowed population it now serves○ It only affects models trained with deep learning○ It makes models slower over time
7. Which practical steps can a non-technical manager take to reduce bias risk? Select all that apply.☐ Ask for accuracy broken down by group rather than only overall☐ Ask what the label actually encodes and whose past decisions it reflects☐ Insist on a route for a person to challenge an automated decision☐ Remove all protected attributes from the data and consider the issue resolved
8. What is an adversarial example?○ A training example that was labelled incorrectly○ An input deliberately crafted to make a model produce a wrong output, typically while looking unremarkable to a human○ A test case designed to measure model accuracy○ A competing model built by a rival organisation
9. Which systems most warrant concern about adversarial attack?○ All machine learning systems equally○ Systems standing between someone and money or access — fraud detection, credit scoring, identity verification, content moderation — because those face motivated adversaries○ Only systems that process images○ Internal demand forecasting and maintenance prediction
10. What is prompt injection?○ Supplying a model with too many examples in a single prompt○ Hiding instructions inside content an AI assistant reads, hijacking its behaviour, because such systems cannot reliably separate content being read from instructions being followed○ Injecting training data into a deployed model○ Automatically generating prompts from a template
11. What is the most durable organisational defence against voice-cloned executive fraud?○ Training staff to detect synthetic audio○ Redesigning the control so it does not depend on recognising a voice — call back on a number from your own records, require two authorisers above a threshold, and treat urgency plus secrecy as red flags○ Prohibiting phone-based approvals entirely and using email instead○ Recording all calls for forensic comparison
12. What is the most accurate framing of AI's effect on employment?○ AI eliminates whole occupations, so the question is which job titles disappear○ AI automates tasks within jobs, so the question is what proportion of a role is repetitive, fast judgement with recorded outcomes○ AI has no measurable effect on employment○ AI affects only manual and low-skilled work