8.6.1.1 - What Are Hallucinations, and How Do We Review Outputs to Reduce Errors? (Difficulty: Beginner | Path: Launch)

Lesson Summary

The Liar in the Machine: Understanding AI Hallucinations

What is a Hallucination?

In the context of Artificial Intelligence, a \"hallucination\" is when an AI model generates a response that looks factual, confident, and grammatically correct, but is completely made up. Unlike a search engine that retrieves existing data, Large Language Models (LLMs) predict the next word in a sentence based on probability. Sometimes, the most \"probable\" sounding sentence is factually false. The AI does not \"know\" facts; it only knows patterns. It will happily invent a new feature for your product, a fake shipping policy, or a historical event that never happened, simply because it fit the sentence structure.

Why is this critical for E-commerce?

If you are using AI to write product descriptions, answer support tickets, or generate blog posts, hallucinations are a direct liability risk. Imagine an AI generating a description for a polyester raincoat that claims it is \"100% Organic Breathable Silk.\" The customer buys it, realizes it's plastic, and returns it. You lose the shipping cost, you get a return, and you may face legal action for false advertising. In customer support, the risk is even higher—an AI chatbot might promise a refund policy you don't offer or invent a discount code that doesn't exist.

Real-Life Scenario: The Air Canada Case

In a famous 2024 court case, Air Canada's AI chatbot hallucinated a bereavement fare policy that did not exist, promising a passenger a discount they weren't entitled to. When the passenger sued, the airline argued the chatbot was a separate legal entity responsible for its own actions. The court rejected this, and Air Canada was forced to pay. This sets a precedent: You are legally responsible for every word your AI generates.

How to Review and Mitigate Hallucinations

You cannot eliminate hallucinations entirely, but you can catch them before they reach the customer. Implement the \"Red Pen\" protocol:

1. The \"Fact-Check\" Pass: Never copy-paste AI content directly to your store. Assign a human reviewer to verify every specific claim. Look for numbers, dates, material compositions (e.g., \"stainless steel\"), and guarantees. These are the high-risk zones.
2. Use \"Grounding\" Prompts: When using ChatGPT or similar tools, provide the source material in the prompt. Instead of asking \"Write a description for this kettle,\" paste your manufacturer's spec sheet and write: \"Write a description for this kettle using ONLY the facts provided below. Do not invent features.\"
3. Temperature Control: If you are using the API (like in Shopify Flow or a custom app), set the \"Temperature\" setting to 0 or near 0. This reduces creativity and randomness, forcing the model to stick closer to the most probable (and usually more accurate) text.
4. Spot Checks for Bulk Content: If you generate 1,000 descriptions at once, you cannot read them all deeply. Randomly sample 5-10% of the outputs. If you find a hallucination in the sample (e.g., it keeps adding \"dishwasher safe\" to wood items), you must scrap the batch and refine your prompt.

Do's and Don'ts

• Do: Treat AI as a junior intern who is a pathological liar. They are talented and fast, but you would never publish their work without checking it.
• Don't: Ask AI to verify its own facts in the same chat window. It will often double down on the lie. Use a separate search engine or source document to verify.