Why Your AI Hallucinates and How to Measure It

Your AI is lying to you. Confidently. Fluently. And most of the time, you have no idea.

AI hallucination — where models generate false information presented as fact — is one of the biggest unsolved problems in production AI. Here's what's actually causing it and how to measure it.

What Actually Causes Hallucination

1. Training data gaps

When a model hasn't seen enough data about a topic, it fills in the gaps by generating plausible-sounding content based on patterns. It doesn't know what it doesn't know.

2. Overconfident generation

LLMs are trained to produce fluent, confident text. This same property makes them dangerous — they'll state false information with the same confidence as true information.

3. Context window limits

When a model's context window is too small to hold all relevant information, it starts generating from memory — which may be wrong, outdated, or simply fabricated.

"The model doesn't know it's wrong. It's not lying — it's pattern-matching toward plausibility." — AI Safety Researcher, 2025

How to Measure Hallucination Risk

Method 1: Hedge word density

Models that are uncertain often hedge. Count uncertainty signals relative to confident assertions:

import re

def hedge_score(text: str) -> float:
    hedges = re.findall(
        r'\b(probably|likely|perhaps|around|approximately|roughly|I think|I believe)\b',
        text, re.IGNORECASE
    )
    assertions = re.findall(
        r'\b(always|never|proven|confirmed|established|guaranteed)\b',
        text, re.IGNORECASE
    )
    # High assertions + low hedges = higher hallucination risk
    return min(100, len(hedges) * 5 + len(assertions) * 8)

Method 2: Source verification

Does the AI cite sources? Do those sources actually exist and support the claim?

import httpx

async def verify_sources(text: str) -> dict:
    urls = re.findall(r'https?://[^\s]+', text)
    results = []
    async with httpx.AsyncClient() as client:
        for url in urls:
            try:
                r = await client.head(url, timeout=5)
                results.append({"url": url, "exists": r.status_code == 200})
            except:
                results.append({"url": url, "exists": False})
    return results

Method 3: Self-consistency checking

Ask the model the same question multiple times with slightly different phrasing. If the answers contradict each other, it's likely hallucinating:

def consistency_check(question: str, model, n=3) -> float:
    answers = [model.generate(question) for _ in range(n)]
    # Check semantic similarity between answers
    # Low similarity = high hallucination risk
    similarities = [cosine_similarity(answers[i], answers[j])
                   for i in range(n) for j in range(i+1, n)]
    return 1 - (sum(similarities) / len(similarities))

Production Checklist

Before shipping any AI feature, verify these:

✓ Are all factual claims grounded in retrieved documents (RAG)?

✓ Does your system prompt instruct the model to say "I don't know"?

✓ Are cited URLs verified to exist?

✓ Is hallucination rate measured and monitored in production?