What Causes AI Hallucinations — and How to Fix The Problem in Three Simple Steps

by Nils Osmar

Imagine asking a colleague a question, and instead of saying “I don’t know,” they spend five minutes creating a beautiful, deeply detailed story which you later learned was completely made up.

In the world of Artificial Intelligence, this is an “hallucination.” It is the moment when an AI like ChatGPT, Claude, or Gemini stops being an assistant and starts being a storyteller, confusing fiction with fact. The more false the information, the more confident the AI may seem in its presentation of it.

For a long time, we thought hallucinations were “bugs” that could be patched out. But new research suggests that AI hallucinations are built into how LLM’s function. Understanding why it happens gives us a way of almost entirely fixing the problem.

Stopping AI Hallucinations in Three Steps

The main cause of hallucinations is that people are asking AI to be both a researcher and writer at the same time.
To get accurate results, you need to separate the creation of original material from the verification.
You then need to verify (or ask an AI to verify) all of the facts in your article.
Hallucinations almost always occur in the generation of new material, not one fact checking what another AI has written.

Step 1: The “Drafting” Phase (Creation)

When creating your prompt, do not just ask for a final answer. Ask the AI to show its work.

The “Escape Hatch” Prompt: End every request with: “If you are unsure of a fact or lack specific data, please say ‘I don’t know’ rather than guessing.”
The “Chain of Thought” Request: Ask the AI to “Explain your reasoning step-by-step before giving the final answer.” This forces the model to slow down and reduces “lazy” pattern-matching.
Models such as Claude also have an extended thinking mode in which the reasoning process can be shown clearly step-by-step. Make sure it is switched on.
Tell the AI clearly in your initial prompt that it is acceptable to say “I don’t know.”

Step 2: The “Audit” Phase (Verification)

After the AI gives you an initial answer, do not take it at face value. Instead, use ask the same AI (or a different one) to critique the previous output.

Copy the AI’s previous answer, paste it in as a new prompt, and say: “I am an expert fact-checker. Please review the text below and identify any specific claims (names, dates, statistics, or citations) that might be incorrect or hallucinated. Highlight the uncertainties.”
If you are using ChatGPT, try pasting the answer into Claude or Gemini and asking: “Does this information contradict any known facts in your database?”

Step 3: The “Lateral Reading” Check (Human Verification)

Never publish or rely on high-stakes AI data without a 60-second “Lateral Reading” check.

Search the citations: If an AI gives you a link or a book title, Google it.
AI is notorious for inventing “plausible-sounding” URLs and titles that do not exist.
It can also be effective to ask a different AI to verify the citations. Your first AI has created. But double checking them yourself is an important step.

Quick Checklist for Reliable Outputs:

Did I give it an “out”? (Did I tell it it’s okay to say “I don’t know”?)
Is it grounded? (Did I provide it with a source document to read first, or did I ask it to use its internal “memory”?)
Have I run an Audit? (Did I ask the AI to find its own mistakes?)

Learn More

How Large Language Models Work — and Why AI Consciousness Can’t Be Ruled Out

Reducing Hallucinations in AI

Working With ChatGPT, Claude and Other AIs More Effectively

What’s a Thread? And Is It Ethical to Delete Them?

Nils Osmar

Nils Osmar is the Director of Rekindle School, an independent educational program in Seattle, Washington, established in 2002. He has several websites, exploring questions related to:

• Consciousness in today's Large Language Models

• Life extension and anti-aging

• News, politics, culture and the arts

Unless otherwise indicated, Nils is the author of all of the articles on this site.

What Causes AI Hallucinations — and How to Fix The Problem in Three Simple Steps