What Is an AI Hallucination?
An AI hallucination occurs when a large language model generates information that is factually incorrect, fabricated, or contradicts verifiable reality — and presents it as if it were true. The term "hallucination" is borrowed from psychology to describe the model "perceiving" information that doesn't exist in reality.
For businesses, AI hallucinations can take many forms:
- Incorrect business hours, address, or phone number
- Fabricated services or products the business doesn't offer
- False claims about credentials, certifications, or awards
- Attribution of a competitor's characteristics to your business
- Invented quotes from business owners or employees
- Outdated information presented as current
Why AI Models Hallucinate
AI hallucinations occur due to fundamental properties of how language models work:
Probabilistic generation: LLMs generate text by predicting the most statistically likely next token (word/phrase), based on patterns in training data. When specific facts are uncertain, the model may generate plausible-sounding content rather than explicitly acknowledging uncertainty.
Conflicting training data: If multiple sources contradict each other about a business (e.g., old vs. current address), the model may average between them or pick the wrong one.
Sparse training signal: When a business has minimal training data coverage, the model fills in gaps with statistically likely characteristics of similar businesses — which may not be accurate.
Knowledge cutoff: Information that changed after the model's training cutoff date will cause the model to cite outdated information as current.
Hallucinations vs. Outdated Information
It's useful to distinguish between:
- True hallucination: The model generates information that was never true (invented services, fabricated reviews)
- Stale information: The model cites information that was once true but is now outdated (old address, discontinued services)
Both cause the same practical harm but require different remediation approaches.
Reducing Hallucinations About Your Business
The most effective hallucination prevention strategies:
- Consistent, accurate NAP data across all directories and platforms
- Comprehensive schema markup with explicit, machine-readable facts
- Current, detailed FAQ content addressing specific facts about your business
- Strong citation signal so AI has high-confidence, consistent data to draw from
- Regular AI audits to detect and correct hallucinations before they scale
Q: Can I sue an AI company for hallucinations about my business? A: The legal landscape is evolving. Some jurisdictions have begun establishing AI-specific liability frameworks, and traditional defamation law may apply to materially false AI outputs that cause measurable harm. Consult a communications attorney if you believe AI hallucinations have caused significant business damage.
Q: How long does it take for hallucinations to disappear after I fix the source data? A: For retrieval-based AI (Perplexity, Bing Copilot), corrections can propagate within days to weeks as the corrected source data is crawled. For training-based models (base ChatGPT, Claude), corrections only propagate at the next model training cycle — potentially 6-12 months later.