AI Hallucinations: Why LLMs Make Up Facts and How to Detect Them

⏱ 6 min read · AI & Machine Learning

ChatGPT confidently tells you that the Eiffel Tower was built in 1923. You check — it was 1889. The AI hallucinated. It generated plausible-sounding but false information. This isn't rare. All LLMs hallucinate. Understanding why and how to detect hallucinations is crucial for using AI reliably.

LLMs don't "know" facts. They predict plausible text based on patterns in training data. Sometimes plausible text is false. The AI can't tell the difference.

What Are Hallucinations?

Hallucinations are confident but incorrect outputs from LLMs. They come in several forms:

**Factual errors:** Wrong dates, numbers, names, or events
**Fabricated sources:** Citing papers or books that don't exist
**Invented details:** Adding specifics that weren't in the input
**Logical inconsistencies:** Contradicting earlier statements
**Nonsensical outputs:** Grammatically correct but meaningless text

The dangerous part: hallucinations are delivered with the same confidence as accurate information.

LLMs don't have a "confidence score" for facts. They generate plausible text, not verified truth.

Why Hallucinations Happen

**1. Pattern matching, not knowledge:** LLMs learn statistical patterns, not facts. If training data had errors, the model learns those errors.

**2. Pressure to generate:** LLMs are trained to always produce output. When they don't know something, they generate plausible-sounding text rather than saying "I don't know."

**3. Context limitations:** With limited context window, the model might "forget" earlier facts and generate inconsistent information.

**4. Training data gaps:** If a topic wasn't well-represented in training data, the model fills gaps with plausible guesses.

**5. Ambiguous prompts:** Vague questions lead to guessed answers rather than precise responses.

Common Hallucination Patterns

**Fake citations:** "According to a 2023 study by Smith et al. published in Nature..." (study doesn't exist)

**Plausible but wrong numbers:** "The population of Tokyo is 15 million" (actually 14 million, but 15 is plausible)

**Invented people:** "Dr. Sarah Johnson, a leading expert in..." (person doesn't exist)

**Merged facts:** Combining real information from different sources incorrectly

**Outdated information:** Stating facts that were true in training data but have changed

Detection Techniques

**1. Verify specific claims:** Check dates, numbers, names, and citations independently. Don't trust AI for factual accuracy without verification.

**2. Ask for sources:** Request citations or sources. If the AI can't provide verifiable sources, be skeptical.

**3. Cross-check with multiple queries:** Ask the same question differently. Inconsistent answers indicate hallucination.

**4. Look for hedging language:** Accurate responses often include "likely," "typically," "according to." Hallucinations are often stated as absolute facts.

**5. Check for internal consistency:** Does the response contradict itself or earlier statements?

High-Risk Domains

Some domains have higher hallucination rates:

**Medical information:** Dangerous if wrong. Always verify medical advice.
**Legal information:** Laws vary by jurisdiction. AI might mix jurisdictions.
**Recent events:** Training data has a cutoff date. AI doesn't know recent events.
**Niche topics:** Less training data = more hallucination.
**Mathematical calculations:** LLMs aren't calculators. They approximate.

For these domains, treat AI output as a starting point, not final answer.

Reducing Hallucinations

**1. Be specific in prompts:** "What is the population of Tokyo according to 2020 census data?" vs "What's Tokyo's population?"

**2. Ask for uncertainty:** "If you're not certain, say so" or "Rate your confidence in this answer"

**3. Provide context:** Give the AI relevant information rather than asking it to recall from training.

**4. Use retrieval-augmented generation:** Systems that retrieve facts from databases before generating text hallucinate less.

**5. Request step-by-step reasoning:** "Explain your reasoning" forces the AI to show its work, making errors more visible.

The Citation Problem

LLMs often fabricate citations that sound real:

"According to Johnson & Smith (2022) in the Journal of Applied Psychology..."

This citation might be completely invented. The journal exists, but the paper doesn't.

**Always verify citations:**
- Search for the paper on Google Scholar
- Check if the authors exist and work in that field
- Verify the journal published that volume/issue
- Read the actual paper if possible

Code Hallucinations

LLMs hallucinate in code too:

**Invented functions:** Using API methods that don't exist
**Wrong syntax:** Mixing syntax from different languages
**Deprecated methods:** Using old APIs that no longer work
**Logical errors:** Code that runs but produces wrong results

Always test AI-generated code. Don't assume it works because it looks right.

The Confidence Paradox

Hallucinations are often delivered with high confidence. The AI doesn't "know" it's wrong.

**Confident hallucination:** "The capital of Australia is Sydney." (It's Canberra)
**Hedged accuracy:** "The capital of Australia is Canberra, though Sydney is the largest city."

Ironically, hedged language often indicates more accurate information because the AI is acknowledging complexity.

Tools and Techniques

**Fact-checking prompts:** "Before answering, consider if you're certain about this information. If not, say so."

**Multiple model verification:** Ask the same question to different models (GPT-4, Claude, Gemini). If they disagree, investigate.

**Retrieval systems:** Use AI with access to search engines or databases for factual queries.

**Human review:** For critical applications, always have human verification of AI output.

When to Trust AI

**More reliable for:**
- General knowledge (widely documented topics)
- Creative tasks (writing, brainstorming)
- Code structure (not specific APIs)
- Explanations of concepts
- Summarization of provided text

**Less reliable for:**
- Specific facts (dates, numbers, names)
- Recent events (post-training cutoff)
- Niche topics
- Medical/legal advice
- Citations and sources

The Future of Hallucinations

Newer models hallucinate less than older ones, but no model is hallucination-free. Improvements come from:

- Better training data
- Retrieval-augmented generation
- Uncertainty quantification
- Fact-checking layers
- Human feedback

But fundamentally, LLMs generate plausible text, not verified truth. This limitation is inherent to how they work.

Best Practices

**1. Verify critical information:** Don't trust AI for facts without checking.
**2. Use AI as assistant, not authority:** Let AI draft, but you verify and approve.
**3. Be skeptical of specifics:** Dates, numbers, names, citations need verification.
**4. Test code thoroughly:** AI-generated code might look right but be wrong.
**5. Document sources:** When using AI output, note that it's AI-generated and unverified.