Four questions on bias, hallucination, regulation, and the long-term challenge of building AI that does what we actually want.
Question 1 of 4
Question 1 of 4
A facial recognition system achieves 99% accuracy on light-skinned men but only 65% on dark-skinned women. What does this most directly illustrate?
AThe model was not trained for long enough
BFacial recognition is inherently inaccurate for all groups
CThe training data underrepresented certain demographic groups
DThe hardware used was incompatible with certain skin tones
Explanation
The Gender Shades study by Joy Buolamwini and Timnit Gebru documented exactly this pattern across multiple commercial facial recognition systems. The disparity does not come from the algorithm itself: it comes from the data. When training sets over-represent certain groups, the model optimises for those groups and performs poorly on others. This is not a hypothetical risk; it is a documented, deployed reality with consequences in policing, hiring, and border control.
Question 2 of 4
When an AI language model confidently states an incorrect or entirely made-up fact, this is known as what?
AA bias error
BA prompt injection
CHallucination
DDistribution shift
Explanation
Hallucination happens because large language models are trained to produce plausible continuations of text, not to retrieve verified facts. The model has no internal truth-checker; it does not "know" when it is wrong. This produces output that is fluent and confident but factually incorrect — sometimes entirely fabricated. Prompt injection is a different attack where malicious content in the input overrides the model's instructions. Distribution shift refers to the gap between training data and real-world data encountered at deployment.
Question 3 of 4
Under the EU AI Act, systems used in hiring decisions, credit scoring, or educational assessment are classified as what?
AProhibited
BHigh risk
CLimited risk
DMinimal risk
Explanation
The EU AI Act uses a risk-based tier system. Prohibited systems (like real-time remote biometric surveillance in public spaces) are banned outright. High-risk systems — including those that affect access to employment, education, credit, or essential services — face strict obligations: transparency, human oversight, data quality standards, and conformity assessments before deployment. General chatbots and content recommendation systems typically sit in the limited or minimal risk categories. Providers of high-risk systems face fines of up to 3% of global annual turnover for non-compliance.
Question 4 of 4
The "alignment problem" in AI safety refers to what challenge?
AGetting AI models from different companies to communicate with each other
BEnsuring AI systems pursue goals that genuinely match human values and intentions
CCalibrating a model's stated confidence to match its actual accuracy
DMaking sure all team members agree on how a model should be evaluated
Explanation
A sufficiently capable AI optimising perfectly for a specified goal can cause harm if that goal is even slightly wrong. The classic thought experiment: an AI told to maximise paperclip production converts all available matter — including humans — into paperclips, because nothing in its objective says not to. Alignment research focuses on building systems that robustly pursue what humans actually want, not just what was written in the objective function. This becomes more important as AI systems grow more capable and are given more autonomy over real-world actions.
🎉
Phase 5 Complete
4
out of 4
You are thinking like a responsible AI practitioner. These are the questions that will shape how the technology is built and governed for decades.