Concept questionEasy~5 minAI Fundamentals / 8 of 20

Interview question

What does hallucination mean in an AI application?

Defines hallucination as unsupported or incorrect model output and connects it to evidence, task design, evaluation, and user-visible uncertainty.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

Hallucination usually means the model produces content that is incorrect, fabricated, or unsupported by the evidence it was expected to use. The wording may still sound confident because the model is generating a plausible continuation, not checking every claim against a source of truth. The exact definition depends on the task: an unsupported citation, an invented API method, and a summary that contradicts its document are related but measurable in different ways.

Full answer below

Practice focus

A precise explanation of an overloaded term and practical controls that do not promise hallucination can be eliminated.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
Hallucination is a product risk only when it is defined against the task and available evidence rather than used as a label for every bad answer.
Hallucination usually means the model produces content that is incorrect, fabricated, or unsupported by the evidence it was expected to use. The wording may still sound confident because the model is generating a plausible continuation, not checking every claim against a source of truth. The exact definition depends on the task: an unsupported citation, an invented API method, and a summary that contradicts its document are related but measurable in different ways.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
I reduce the risk by narrowing the task, supplying authoritative context, asking for evidence where appropriate, validating structured facts, and allowing the system to say it lacks enough information. RAG can improve grounding but cannot guarantee it; retrieval may return the wrong passage and the model may ignore or misread it. For material decisions, application code verifies values against the system of record or requires human review.
Finish
Advanced answer
Optional deeper angle for production judgment and nuance.
I define a task-specific error taxonomy and evaluation set instead of tracking one vague hallucination rate. Grounded question answering can measure claim support and citation correctness; extraction can compare fields with labels; code generation can compile and test outputs. I also design the user experience around residual uncertainty. A confident unsupported answer can be more harmful than a refusal, while an overly cautious system can become useless, so the release decision balances both error types with the actual consequence.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Calling every undesirable response a hallucination.
Claiming RAG eliminates hallucination.
Asking the model to verify its own answer and treating that as proof.
Ignoring the harm difference between a draft and a state-changing decision.

Red flagsCandidateRed flagsSignals that an interviewer may treat as risky.Open details

Details

The product displays invented facts as authoritative records.
No labeled or evidence-based evaluation exists.
Users cannot distinguish sourced content from generated interpretation.

Follow-up prompts

Quick interviewer prompts inside this drill.

Follow-up questionsCandidateFollow-up questionsLikely next questions after the main answer.Open details

Details

Treat these as the next interviewer turn. Answer one out loud, then save the gap in your practice notes.

How would you measure hallucination for a document-answering feature?
Why can RAG still produce unsupported claims?
When should the system refuse?
Which facts should application code verify?

Continue the section

Next in AI Fundamentals: When not to use an LLM

Step 9 of 20

When should you not use an LLM?

Practice route

Read core answer Review checklist Follow-up prompts Track progress

Finished this drill?

Practice freely. Sign in only when you want to save completion, confidence, notes, or bookmarks.

New to Aporeon? Create account

Drill details

Human editorial direction

Aporeon is shaped by Aleksandar Tomovski, a software developer with experience on both sides of technical interviews. Content is reviewed for accuracy, natural spoken delivery, useful depth, and honest trade-offs.

About the creator Editorial standards

Was this drill useful?

Help improve the interview library.

Found this answer useful?

Say thanks with a standalone, one-time $5 contribution. No account required.

No subscription or perks.Choose another amount

Opening drill

Concept questionEasy~5 minAI Fundamentals / 8 of 20

Interview question

What does hallucination mean in an AI application?

Defines hallucination as unsupported or incorrect model output and connects it to evidence, task design, evaluation, and user-visible uncertainty.

1Say itAnswer out loud first

2CompareCompare with Core answer

3RefineRefine with Stronger and Advanced

TL;DR

Full answer below

Practice focus

A precise explanation of an overloaded term and practical controls that do not promise hallucination can be eliminated.

Answer first

Answer path

Core answer first, then checkpoints toward a stronger answer.

Start
Core answer
Start here. This is the direct spoken answer to practice first.
Overview
Hallucination is a product risk only when it is defined against the task and available evidence rather than used as a label for every bad answer.
Hallucination usually means the model produces content that is incorrect, fabricated, or unsupported by the evidence it was expected to use. The wording may still sound confident because the model is generating a plausible continuation, not checking every claim against a source of truth. The exact definition depends on the task: an unsupported citation, an invented API method, and a summary that contradicts its document are related but measurable in different ways.
Checkpoint
Stronger answer
Optional upgrade with examples, trade-offs, and sharper phrasing.
I reduce the risk by narrowing the task, supplying authoritative context, asking for evidence where appropriate, validating structured facts, and allowing the system to say it lacks enough information. RAG can improve grounding but cannot guarantee it; retrieval may return the wrong passage and the model may ignore or misread it. For material decisions, application code verifies values against the system of record or requires human review.
Finish
Advanced answer
Optional deeper angle for production judgment and nuance.
I define a task-specific error taxonomy and evaluation set instead of tracking one vague hallucination rate. Grounded question answering can measure claim support and citation correctness; extraction can compare fields with labels; code generation can compile and test outputs. I also design the user experience around residual uncertainty. A confident unsupported answer can be more harmful than a refusal, while an overly cautious system can become useless, so the release decision balances both error types with the actual consequence.

Review signals

Use these to catch weak phrasing, shallow reasoning, and risky claims.

Common mistakesCandidateCommon mistakesMistakes that weaken the answer or signal shallow understanding.Open details

Details

Calling every undesirable response a hallucination.
Claiming RAG eliminates hallucination.
Asking the model to verify its own answer and treating that as proof.
Ignoring the harm difference between a draft and a state-changing decision.

Red flagsCandidateRed flagsSignals that an interviewer may treat as risky.Open details

Details

The product displays invented facts as authoritative records.
No labeled or evidence-based evaluation exists.
Users cannot distinguish sourced content from generated interpretation.

Follow-up prompts

Quick interviewer prompts inside this drill.

Follow-up questionsCandidateFollow-up questionsLikely next questions after the main answer.Open details

Details

Treat these as the next interviewer turn. Answer one out loud, then save the gap in your practice notes.

How would you measure hallucination for a document-answering feature?
Why can RAG still produce unsupported claims?
When should the system refuse?
Which facts should application code verify?

Continue the section

Next in AI Fundamentals: When not to use an LLM

Step 9 of 20

When should you not use an LLM?

What does hallucination mean in an AI application?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue the section

What does hallucination mean in an AI application?

Practice focus

Answer path

Overview

Review signals

Follow-up prompts

Continue the section