A hallucination is generated output that makes something up or presents a fabricated or nonsensical answer as if it were sound. An LLM is trained to produce coherent, context-aware text, not necessarily factually accurate answers. In an enterprise chatbot, that distinction matters most when customers may rely on the response. section 1.4 section 1.5
Inaccurate but task-relevant context can prompt a hallucination. A larger context window is not automatically safer: it lets an application supply more information, but may also increase the complexity and uncertainty of the answer. Hallucinations can misinform users, disrupt business decisions, weaken customer trust, and cause financial or reputational harm. section 1.4 section 8.1.1 section 13.1.2
A plausible answer with a false claim
At Google’s Bard launch, a response about discoveries from the James Webb Space Telescope incorrectly said it took the first pictures of exoplanets. The answer’s conversational form did not make that claim reliable. section 13.1.2
Reduce exposure rather than expecting to eliminate hallucinations:
- Provide accurate, current information. Ground answers in supplied data through retrieval-augmented generation (RAG), or frame a suitable task as summarization rather than open-ended question answering. section 13.1.2 section 6.8
- Use prompt context and examples, and instruct a chatbot to say it does not know rather than invent an answer. Add pre- and post-generation checks where correctness matters. section 13.1.2 section 6.4.1
- Tune generation settings for the task. Very high temperature can produce more hallucinations and nonsensical text; a large context window alone does not replace grounding. section 3.2.4 section 8.1.1
- Keep a human responsible for the final decision in life-and-death situations, with AI acting as a copilot. section 1.5
Listing 3.11 Completions output with the temperature at 1.8
{
"choices": [
{
"finish_reason": "stop",
"index": 0, #1
"logprobs": null,
"text": "\n\n1. ComfortGroom Pet Furnishing \n2. Pampered TreaBankant Carers \n3. Toptech Sunny Haven Promotion.",
"content_filter_results"={...}
},
{
"finish_reason": "stop",
"index": 1, #2
"logprobs": null,
"text": "\n\n1: Naturalistov ClearlywowGroomingz
Pet Luxusia \n2: VipalMinderers Pet
Starencatines grooming \n3: Brisasia
Crownsnus Take Care Buddsroshesipalising",
"content_filter_results"={...}
},
{
"finish_reason": "length",
"index": 2,
"logprobs": null,
"text": "\n\n1. TrustowStar Pet Salon\n
2. Hartipad TailTagz Grooming & Styles\n
3. LittleLoft Millonista Cosmania DipSavez
Hubopolis ShineBright Princessly
Prosnoiffarianistics Kensoph Cowlosophy
Expressionala Navixfordti Mundulante Effority
DivineSponn BordloveDV EnityzBFA Prestageinato
SuperGold Cloutoilyna Critinarillies
Prochromomumphance Toud",
"content_filter_results"={...}
}
],
...
}
Compare the generated responses at temperature 1.8 with the lower-temperature examples: several names become nonsensical, illustrating why more randomness is not always useful. section 3.2.4
