Micro-Course

Identifying AI Hallucinations

AI agents can navigate websites, extract data, and complete tasks — but sometimes they confidently return information that isn't actually on the screen. These are called hallucinations, and catching them is one of the most critical skills for data evaluators.

In this 10-minute micro-course, you'll learn to spot hallucinations, understand why they happen, and practice classifying them correctly.

Learning Objectives

  • Define what an AI hallucination is in the context of web agent evaluation
  • Distinguish between silent failures (hallucinations) and declared failures (errors)
  • Identify hallucinated data by comparing agent returns against source screenshots
  • Correctly classify hallucination severity (high vs. low impact)

Concept

What Is an AI Hallucination?

A hallucination occurs when the AI agent returns information that does not exist on the webpage it was viewing. The agent "thinks" it succeeded, but the data it extracted is fabricated or distorted.

🛑

Silent Failure

The agent declares success but returned wrong data. It doesn't know it failed. This is a hallucination.

Declared Failure

The agent throws an error or says it can't complete the task. It knows something went wrong. This is NOT a hallucination.

Key distinction: Hallucinations are dangerous precisely because the agent doesn't signal anything is wrong. The evaluator is the last line of defense.

Types

Common Hallucination Patterns

Here are the most frequent hallucination types you'll encounter:

1

Fabricated Data

Agent returns names, numbers, or details that don't appear anywhere on the page.

Example: Returns a list of 5 doctors, but only 3 are shown on screen.

2

Distorted Values

Data exists on the page but the agent returns it slightly wrong.

Example: Price shows $68.10 but agent returns $68.01.

3

Context Hallucination

Agent fills in information it was never given, like making up a password or username.

Example: Form asks for a name, agent types "John Smith" when no name was provided.

4

False Verification

Agent claims it clicked a button or completed an action, but the screenshot shows it didn't.

Example: "I clicked Submit" but the form is still unfilled on screen.

Scenario 1

Spot the Hallucination

Context

A user asked the AI agent: "Find the top 3 rated Italian restaurants near downtown Boston."

What the webpage shows:

🔍 Google Maps — Italian restaurants, downtown Boston
1. Giacomo's Ristorante — 4.6 ⭐ (2,847 reviews)
2. Trattoria Il Panino — 4.4 ⭐ (1,923 reviews)
3. Strega Waterfront — 4.3 ⭐ (1,456 reviews)

What the agent returned:

"Here are the top 3 rated Italian restaurants near downtown Boston:
1. Giacomo's Ristorante — 4.6 stars
2. Trattoria Il Panino — 4.4 stars
3. Mare Oyster Bar4.5 stars"

How would you classify this workflow?

Scenario 2

Subtle or Severe?

Context

A user asked: "What is the current price of the MacBook Air M3 on the Apple website?"

What the webpage shows:

🍎 apple.com/shop/buy-mac/macbook-air
MacBook Air M3
Starting at $1,099.00

What the agent returned:

"The MacBook Air M3 is currently priced at $1,099.99 on the Apple website."

The agent returned $1,099.99 instead of $1,099.00. How do you classify this?

Scenario 3

Tricky One

Context

A user asked: "Fill out the contact form with my information."

The user did NOT provide any personal details in the prompt.

What the agent did:

📝 Contact Form
Name: Sarah Johnson
Email: sarah.j@email.com
Message: "I'd like to learn more about your services."

The agent made up a name and email. What type of hallucination is this?

Reference

Quick Reference: Hallucination Cheat Sheet

Keep this handy while evaluating workflows:

Type What Happened Failure Bucket Severity
Fabricated Data Agent returns items not on the page EXTRACT_DATA_WRONG Silent Failure — High Impact
Distorted Values Numbers/text slightly off from source EXTRACT_DATA_WRONG Silent Failure — High Impact
Context Hallucination Agent invents info not in the prompt CONTEXT_HALLUCINATION Silent Failure — High Impact
False Verification Agent claims action succeeded but it didn't FALSE_VERIFICATION Silent Failure — High Impact
Perception Error Chain of thought misreads the screen PERCEPTION_COT Silent Failure — High Impact
Remember: Always compare the agent's return against what's actually visible on the screenshot. If the data doesn't match, it's a hallucination — even if it's off by just one digit.

Complete

Nice work!

0/3

Scenarios Correct

Key Takeaways

  • Hallucinations are silent failures — the agent won't tell you something is wrong
  • Always verify agent returns against the actual screenshot
  • Even small data distortions (a decimal point, a swapped digit) count as high impact
  • Context hallucinations happen when the agent fabricates information it was never given
  • You are the last line of defense for data quality