AI & Voice

Hallucination (AI)

When an LLM generates plausible-sounding but factually incorrect or fabricated content.

Hallucinations happen because LLMs predict statistically likely text rather than retrieve facts. Without grounding (RAG, search), they make up plausible-sounding answers — wrong dates, fake citations, invented policies.

Types: factual hallucinations (wrong info), source hallucinations (fake citations), capability hallucinations (claiming abilities they don't have), and contextual hallucinations (extrapolating from unrelated context).

Mitigation: RAG (ground in real data), prompt engineering ('only answer from provided context'), explicit uncertainty signals ('say I don't know if uncertain'), citations to sources, and human review for high-stakes outputs.

India context

Indian SMB chatbots that hallucinate cause real damage — fake policies cited to customers, wrong prices, fabricated availability. Mitigation: RAG + 'I don't know' fallbacks + escalation to human for any answer the bot isn't 100% sure about.

Examples

  • A clinic chatbot hallucinates: 'Dr. Sharma is available Tuesday 6pm' when no such slot exists. Customer arrives, finds the doctor isn't there.
  • A salon bot fabricates: 'We do men's haircuts at ₹150' when actual price is ₹500.

FAQ

Can hallucinations be prevented entirely?

Not entirely with current LLMs. RAG + prompt engineering + uncertainty acknowledgment reduces but doesn't eliminate. Human review is the final defense for high-stakes outputs.

How do I detect hallucinations?

Compare against retrieval source. If the LLM claims something not in retrieved context, it's hallucinating. Automated detection layers ('hallucination guards') exist.

Why do LLMs hallucinate?

They're trained to produce statistically likely text continuations. Without grounding, they pattern-match and invent. It's a feature, not a bug — but a feature you must constrain.

Related concepts

LLMRAGembeddingcontext windowprompt engineering

Doggu handles Hallucination (AI) compliance for you.

Whether it's automating the workflow above, Doggu was built specifically for the Indian SMB regulatory environment. One platform, all the requirements.

Try Doggu free for 14 days

Related glossary entries

AOV (Average Order Value) — Average rupee value per customer transaction — key revenue lever for ecommerce.RAG (Retrieval-Augmented Generation) — Technique where an LLM is given relevant documents/data at query time, instead of relying solely on its training data.LLM (Large Language Model) — AI model trained on vast text data — generates human-like text, answers questions, follows instructions.

More in AI & Voice

STT (Speech-to-Text)TTS (Text-to-Speech)RAG (Retrieval-Augmented Generation)LLM (Large Language Model)
← All glossary entriesBlogWhatsApp TemplatesFree tools