Artificial intelligence practice questions.

A national tax authority used a self-learning risk model to select childcare benefit claims for fraud investigation. Families holding a second nationality were selected far more often, thousands were wrongly accused of fraud, and the government resigned in January 2021. Which feature of the system best accounts for the pattern?

• nationality was among the indicators feeding the risk score
• the model produced scores that investigators never looked at
• the model was a rule-based system rather than a learning one
• too few claims were available for the model to be trained at all

Answer: nationality was among the indicators feeding the risk score

The bias entered through an input variable and a selection process chosen by the institution, not through a coding error, and it was amplified because flagged families were then treated as fraudsters by default. Naming the input is what separates an examinable answer from a claim that the algorithm was unfair.

A program trained only on the board game Go defeated a leading human player in 2016. A newspaper described the result as proof that machines now think in the way people do. The strongest objection to that description is that

• the game of Go has fewer possible positions than chess does
• the program was following rules written out by its programmers
• the outcome was decided by a panel of judges rather than by play
• its competence does not transfer to any task outside the game

Answer: its competence does not transfer to any task outside the game

Systems in use are narrow: they are optimised for one task and hold no general capability, and no system with human-like general intelligence exists. Success in a closed game with fixed rules and a clear win condition says nothing about open-ended reasoning.

A bank in the EU refuses a loan application. The decision is produced entirely by a scoring model, with no member of staff involved. Under the GDPR, the applicant is entitled to

• a copy of the model's source code and of the weights that it uses
• human intervention in the decision and the chance to contest it
• an automatic second application processed by a different model
• compensation calculated from the amount of the refused loan

Answer: human intervention in the decision and the chance to contest it

Article 22 restricts decisions taken solely by automated means where they significantly affect a person, and the safeguards are human intervention, an explanation of the logic involved and a right to contest the outcome, not disclosure of the model. Data protection rights and trade secrets sit side by side here.

In December 2023 a newspaper began legal action against the developer of a large language model, arguing that its archive had been used to train the model without permission and that the model could reproduce passages from its articles. Which question does the case put to digital society?

• Whether a language model is capable of storing text at all, which the newspaper's own filing denies is technically possible
• Whether readers are able to tell the difference between an article and a passage produced by the model
• Whether the newspaper's archive was stored securely enough to prevent unauthorised copying
• Whether training on copyrighted work without a licence is a use the rights holder can control

Answer: Whether training on copyrighted work without a licence is a use the rights holder can control

The dispute is about whether reading a work in order to build a statistical model counts as a use that requires permission, which no copyright regime was written to answer. Its outcome decides who is paid when a model is trained, not whether models may exist.

A hospital introduces a tool that ranks scans by urgency. Over several months the radiographers report fewer findings among the scans the tool ranked low, including during a period when the tool was known to be mis-calibrated. Which effect does this illustrate?

• Automation bias: a recommendation is trusted more readily than the reviewer's own reading of the evidence
• Overfitting: the tool has learned the training set instead of the condition it is supposed to detect
• Sampling bias: the scans used to build the tool came from a different hospital population entirely
• Data drift: the population being scanned has changed since the tool was built and validated

Answer: Automation bias: a recommendation is trusted more readily than the reviewer's own reading of the evidence

The failure sits in the human half of the system: an ordering that is only a suggestion becomes the order in which attention is spent, so the tool's errors are inherited rather than caught. This is why 'a clinician remains in charge' is a claim that has to be tested rather than asserted.

A Time investigation published in January 2023 reported that workers in Kenya, paid less than two US dollars an hour, had labelled violent and abusive text so that a widely used chatbot released in November 2022 could filter such content. The situation best illustrates that

• automated moderation systems remove the need for human moderators
• labelling work of this kind can only be carried out where wages are low
• the chatbot's safety filters were trained without any human involvement
• systems presented as automated rest on hidden, low-paid human labour

Answer: systems presented as automated rest on hidden, low-paid human labour

Data labelling, content moderation and rating model outputs are human jobs sitting inside the 'automated' pipeline — researchers call this ghost work. The distractors invert the finding: the labour exists precisely because filters cannot be trained without it, and locating it in Kenya was a cost decision, not a technical necessity.