Interview: What is Artificial Intelligence?

Interview Questions - What Is Artificial Intelligence?

These answers model precision. Adapt the depth to the role rather than reciting the same definition at every level.

Beginner

1. What is an AI system?

An AI system is a machine-based system that receives inputs and infers how to produce outputs such as predictions, content, recommendations, or decisions for objectives in a context. The definition does not require consciousness or human-like thought.

2. Is Machine Learning the same as AI?

No. AI is the broad field. Machine Learning is one family of methods that fits behavior or parameters from data. Symbolic rules, search, planning, constraint solving, and optimization can also contribute to AI systems.

3. Must every AI system learn?

No. A symbolic expert system can infer from manually authored knowledge and rules. A search-based planner can select actions without a learned model. Even a Machine Learning model may remain frozen after deployment rather than adapting online.

4. What is the difference between an LLM and a chat application?

An LLM is a model component. A chat application adds identity, conversation state, context retrieval, prompts, tools, authorization, validation, user experience, evaluation, monitoring, and fallback behavior.

5. Is Generative AI a model family?

Generative AI is better treated as a capability: producing a new artifact such as text, code, an image, audio, or structured data. Several model families can provide that capability.

Engineer

6. How would you classify an AI system without using one hierarchy?

I would classify it by field, method, model family, capability, application, and authority. For example, a policy assistant may belong to AI and NLP, combine retrieval with Deep Learning inference, use an LLM, retrieve and generate text, serve an employee workflow, and have authority only to inform or recommend.

7. Where does the model boundary end?

The model boundary ends at inference input and output. Authentication, context authorization, source selection, workflow state, output validation, side-effect authorization, evaluation, observability, and rollback are application or platform responsibilities even when they react to model output.

8. Why is a confident answer not an authorized decision?

Confidence, when available, describes a model signal under assumptions. Authority is granted by policy and application controls. Impact, reversibility, regulation, human oversight, and ownership determine whether an output may affect a user.

9. Give an example of AI that is not Generative AI.

A gradient-boosted fraud model predicts a risk score. It uses Machine Learning and belongs to an AI application, but it does not construct a new content artifact. The application may use the score to prioritize review.

10. When would you choose deterministic software over AI?

I would choose deterministic software when an authoritative formula, complete policy rule, or direct database query solves the problem; exact reproducibility is required; or the organization cannot define evaluation and failure ownership.

Senior Engineer

11. A stakeholder says, "We need an LLM for churn prediction." How do you respond?

I would return to the output and evidence contract. Churn prediction is usually a bounded supervised prediction problem over structured and behavioral data. I would compare a deterministic baseline and classical ML before an LLM, define the decision using the score, select metrics that reflect business cost, and design for drift and intervention measurement.

12. How do you evaluate a retrieval-assisted policy answer?

I separate retrieval and generation. Retrieval evaluation measures whether authorized, current, relevant evidence was selected. Generation evaluation measures whether material claims are supported by that evidence, the answer follows policy, and the system abstains when evidence is absent or conflicting. I also test access control, latency, cost, and failure recovery.

13. Why is online adaptation an architectural decision?

Online adaptation changes behavior between controlled releases. That creates new requirements for data validation, poisoning resistance, version lineage, evaluation windows, monitoring, approval, and rollback. A frozen model with periodic governed retraining is often the safer first design.

14. What belongs in an AI failure contract?

For each material failure: the detection signal, bounded immediate behavior, user communication, owner, recovery mechanism, and evidence that recovery succeeded. Typical behaviors include abstention, deterministic fallback, human escalation, retry with limits, or reversal of a side effect.

Architect / Principal Engineer

15. Design the control boundary for an internal policy assistant.

I would authenticate the employee before retrieval, enforce source permissions before context assembly, attach source version and owner metadata, route model calls through a governed gateway, validate citation support, and limit authority to answer or draft. Missing, conflicting, stale, or unauthorized evidence causes abstention or policy-owner escalation. Traces link user intent, access decision, retrieved evidence, model and prompt versions, validation, latency, cost, and feedback without exposing sensitive content unnecessarily.

16. How do you decide how much authority an AI application receives?

I consider impact, reversibility, evidence quality, uncertainty, legal and policy constraints, user expectations, detection time, and recovery capability. I begin with the lowest authority that creates value, measure behavior, and require an explicit architecture and governance decision before moving from inform to recommend, draft, decide, or execute.

17. What is the difference between model selection and AI architecture?

Model selection chooses a component against capability, quality, latency, cost, deployment, and governance needs. AI architecture defines the complete system: workflow, data and trust boundaries, orchestration, controls, evaluation, observability, ownership, failure containment, and change management.

18. What evidence would you require before approving production release?

I would require a versioned evaluation set representing normal, boundary, adversarial, access-control, abstention, and failure-recovery cases; thresholds linked to user impact; traceable model, prompt, data, and policy versions; load and cost evidence; threat-model results; rollback rehearsal; named operational owners; and an explicit residual-risk acceptance.

19. How would you challenge the claim that a product "understands" its domain?

I would ask for an operational claim. Which tasks, populations, environments, and failure modes were evaluated? Can the system distinguish absent from conflicting evidence, maintain performance under distribution shift, explain or cite support, and abstain appropriately? Broad cognitive language should be replaced by bounded, measurable capability claims.

Principal Discussion Scenario

A company proposes allowing a policy assistant to approve low-value exceptions because its offline answer accuracy is 96 percent.

A strong response challenges the metric before the threshold. Which cases make up the four percent? Are errors concentrated in protected or high-impact groups? Does answer accuracy test authorization, policy precedence, stale sources, and side effects? What is the cost and reversibility of a wrong approval? The safer path may retain human approval while automating evidence collection and draft recommendations, then revisit authority only after prospective operational evidence and governance approval.