Interview: What is Artificial Intelligence?

Interview Questions - What Is Artificial Intelligence?

"What is AI?" sounds like a warm-up question, and it quietly separates candidates. Weak answers recite a hierarchy of buzzwords; strong answers name whose definition they are using, describe the system along separate axes, and get to authority quickly. The answers below show the strong version, each followed by what it demonstrates.

Beginner

1. How would you define an AI system?

I would use the operational definition the OECD adopted in November 2023 and the EU AI Act copied into Article 3(1): a machine-based system that, for explicit or implicit objectives, infers from its input how to generate outputs such as predictions, content, recommendations or decisions that can influence physical or virtual environments. Autonomy can vary and adaptiveness after deployment is optional. The key word is "infers": the Act's Recital 12 excludes systems that only execute rules defined entirely by people.

What the interviewer is checking: a precise, sourced definition instead of "software that thinks".

2. Is machine learning the same thing as AI?

No. Machine learning is a method, fitting a model's parameters to data. AI as a research field also includes search, logic, planning and optimisation; a chess program built on minimax search is classic AI without any learning. Going the other way, the European Commission's February 2025 guidelines suggest that some very simple statistical predictors fall outside the legal definition even though they technically learn from data.

What the interviewer is checking: that you keep method and field apart, and know the definitions differ at both edges.

3. Is "generative AI" a type of model?

It is a capability: producing new text, code, images, audio or structured data. Large language models provide it, and so do image diffusion models and others. Calling a system "generative AI" tells you what it outputs, not how it works or what it may do.

What the interviewer is checking: that you place the term on the capability axis.

Working Developer

4. Describe GitHub Copilot's inline suggestions on six axes.

Field: AI, code generation. Method: machine learning. Model family: a large language model. Capability: generate. Application: a code editor. Authority: draft. The suggestion appears as ghost text and becomes code only when the developer presses Tab, which is level 5 on the Sheridan–Verplank scale: the machine executes the suggestion if the human approves. The same model inside an agent that commits code by itself would have a different authority, so the model does not determine the risk.

What the interviewer is checking: fluent use of all six axes, including the one most people skip.

5. What went wrong with Air Canada's chatbot, in engineering terms?

It had the wrong authority for its design. It was built to inform, but the tribunal treated its answers as the airline's own statements: in February 2024 it rejected the argument that the chatbot was a separate legal entity and ordered C$812.02 in damages, interest and fees. The chatbot told the customer he could claim a bereavement fare within 90 days, while the policy page said the claim had to come before travel, and nothing kept the two consistent. I would generate policy answers from the canonical policy text with a link, route anything involving money or exceptions to a person or a rules engine, and log answers for disputes.

What the interviewer is checking: that you diagnose an authority failure rather than a model-accuracy failure.

6. When would you not use AI at all?

When a formula, a ledger or a query produces the authoritative answer, such as an account balance; when the rules are complete and change through a controlled process; when nobody can say how quality would be measured; or when nobody owns the outcome of a wrong answer. If nothing needs to be inferred, an AI component adds cost and uncertainty without removing anything.

What the interviewer is checking: judgment about when the simpler tool is the right one.

Senior Engineer

7. How do you decide how much authority an AI output should have?

By reversibility, detectability and cost per mistake, not by how capable the model is. Gmail's spam filter can move mail on its own because a misfiled message is visible and undone in one click. A refund is effectively irreversible and noticed days later at reconciliation, so it starts at draft authority, where the system prepares and a person approves, and moves to execute only for cases where a measured error rate justifies it. On the Sheridan–Verplank scale, that is the difference between level 5 and level 7 or above.

What the interviewer is checking: a principled, reversibility-first rule with examples.

8. Should a person review every answer our support assistant gives?

It depends on rates, so I would compute it. With 100,000 conversations a month, 10% of wrong answers acted on, C$812 per relied-upon error, reviews at C$1.00 each and a 90% catch rate, full review beats no review only above about a 1.37% error rate. Routing only refunds and policy exceptions, a tenth of conversations holding 80% of the costly errors, beats no review above 0.17% and beats full review up to about 6.2%. Volume cancels out, so the decision rests on measuring the error rate, and showing it is below 0.17% at 95% confidence takes about 1,750 clean answers.

What the interviewer is checking: that you turn a policy argument into a small cost model with flip points, and know what must be measured.

9. A vendor says its product is "AI-powered". What do you ask?

Which axis they mean. What is inferred, from which inputs, by which method, and what the outputs may do. I would ask for evidence, because the label itself is cheap: in March 2024 the SEC fined two investment advisers, Delphia and Global Predictions, $225,000 and $175,000 for AI claims they could not substantiate. If the "AI" is a fixed rule set, that may be fine, but it changes how I test it and whether regulations such as the EU AI Act apply.

What the interviewer is checking: scepticism about labels, backed by a real enforcement case.

Architect

10. How does the choice of definition affect a compliance programme?

It decides whether the rules apply at all. Under the EU AI Act, only systems that meet the Article 3(1) definition are in scope; the Commission's February 2025 guidelines suggest classical heuristics such as minimax search, basic data processing and simple prediction may fall outside. For systems in scope, dates matter: prohibitions since 2 February 2025, general-purpose model obligations since 2 August 2025, transparency duties since 2 August 2026, and, after the Digital Omnibus on AI took effect on 27 July 2026, high-risk Annex III obligations from 2 December 2027. I would record, for each system in the inventory, which definition was applied and why, so the scoping decision can be audited.

What the interviewer is checking: that you treat scope as a documented decision with current dates.

11. Klarna automated two-thirds of its customer chats in 2024 and then brought humans back. What would you have done differently?

Klarna reported 2.3 million conversations in the first month, the work of about 700 agents, and resolution in under 2 minutes instead of 11. In May 2025 its chief executive said cost had been "a too predominant evaluation factor" and quality had suffered. I would have set authority by category from the start: execute for routine, reversible requests; draft for refunds, disputes and anything a customer might rely on; and a guaranteed route to a person. I would have tracked a quality measure alongside cost, such as repeat contacts and complaint rate by category, with a rule for moving categories between draft and execute as the measured error rates changed.

What the interviewer is checking: that you design authority per category with measurements, rather than one switch for the whole assistant.

12. Where should authority be enforced in the architecture?

Outside the model, in the application: the model proposes, and deterministic code decides what a proposal may do. That means tool permissions scoped to the task, approval steps for irreversible effects, and policy answers generated from the canonical source. A model cannot be trusted to limit its own authority, for the reason Chapter 1 gives: you cannot verify every behaviour of a component that can express arbitrary output, but you can withhold the capability to do harm.

What the interviewer is checking: that enforcement lives in code the organisation controls, not in the model's instructions.