AI Ethics & Safety - MCQ Practice Questions

Explainability (also called interpretability or transparency) is the principle that AI systems must be able to provide understandable reasons for their outputs and decisions. This is crucial for building trust, enabling oversight, and ensuring accountability, especially in high-stakes domains like healthcare and criminal justice.

AI alignment is the research problem of ensuring that AI systems pursue goals and exhibit behaviors that are truly aligned with human values, intentions, and well-being. A misaligned AI might achieve its programmed objective in ways that are harmful or unintended, which is considered a major long-term safety risk.

Algorithmic bias refers to systematic and unjust outcomes produced by AI systems, typically arising from biased training data, flawed model design, or skewed feedback loops. A well-known example is facial recognition systems that perform poorly on darker-skinned individuals due to underrepresentation in training datasets.

Under the EU AI Act, real-time remote biometric identification (such as live facial recognition) of individuals in publicly accessible spaces by law enforcement is classified as an unacceptable risk and is generally prohibited, with very narrow exceptions. This is due to serious concerns about mass surveillance and violations of fundamental rights.

Deepfakes use generative AI (particularly GANs) to create hyper-realistic fabricated videos, images, or audio. The primary ethical concern is their misuse for spreading misinformation (e.g., fake political speeches), non-consensual intimate imagery, identity fraud, and undermining public trust in authentic media.

The trolley problem is a classic moral dilemma used in AI ethics to illustrate the challenge of programming autonomous vehicles to make life-and-death decisions when a collision is unavoidable. It forces engineers to consider whose safety to prioritize (e.g., passengers vs. pedestrians), raising deep ethical questions about embedding moral frameworks into AI systems.

Differential privacy is a rigorous mathematical definition of privacy that ensures the output of a data analysis is statistically indistinguishable whether or not any single individual's data is included. It allows organizations (like Google and Apple) to learn aggregate patterns from user data while providing provable privacy guarantees to individuals, making it a key tool in ethical AI data handling.

Human-in-the-loop (HITL) refers to a design paradigm where a human is involved in the AI's decision-making cycle, allowing for review, correction, or override of AI outputs. This is considered essential for safety in high-stakes applications such as medical diagnosis, autonomous weapons, and judicial decisions, where AI errors can have severe consequences.

The OECD Principles on Artificial Intelligence, adopted in May 2019, were the first intergovernmental standard on AI ethics. They were endorsed by all 36 OECD member countries plus several partner economies. The principles cover inclusive growth, human-centred values, transparency, security, and accountability, forming a foundational international framework for responsible AI governance.

Model collapse is a phenomenon identified in generative AI research where models trained on data that itself was generated by AI (rather than real human data) tend to lose the diversity and statistical richness of the original data distribution over successive generations. This creates a feedback loop of degradation, raising concerns about the long-term quality and safety of AI systems as the internet becomes flooded with AI-generated content.