AI · Theory

AI Hallucination

Theory

191/215

AI Hallucination

Hallucination: the model outputs fluent text that is false — fake citations, wrong dates, invented APIs. LLMs predict likely words, not a truth database. Mitigate: retrieval (RAG), citations, human check, lower temperature for facts.

Classroom — ask for a paper title that does not exist — a chat model may invent one. That demo is the warning.

AI Hallucination — output — warning. LLMs predict likely words, not a library catalogue.

Exam tip

What it is + one mitigation.

Example

# Hallucination
print("Asked: cite the 2019 Sharma paper on jug algorithms.")
print("Model: invents author, title, year — looks real.")
print("Fix: retrieve docs / ask a human / lower temperature.")

AI Hallucination — output: warning. LLMs predict likely words, not a library catalogue.

Short notes

DefFluent but false.
RuleFix : RAG + human check.

# Hallucination print("Asked: cite the 2019 Sharma paper on jug algorithms.") print("Model: invents author, title, year — looks real.") print("Fix: retrieve docs / ask a human / lower temperature.")

AI Hallucination

Example

Short notes

Questions

AI Hallucination

Example

Short notes

Questions