GPT-2’s “Ovid’s Unicorn” sample read like a news story. It followed a whimsical prompt about English-speaking unicorns in the Andes with a scientist, a university affiliation, fabricated quotes, and consistent journalistic style. Yet it also described “four-horned” unicorns and “fires burning underwater.” These details exposed the gap between convincing prose and a reliable world model. section 1.2
A hallucination is generated content that sounds plausible but lacks factual support or is false. Fictional details in a fictional prompt are not, by themselves, the problem: the warning is that fluent writing can make unsupported claims seem credible. GPT-2’s sample showed both how far language-model coherence had advanced and why fluency alone could not establish accuracy. section 1.2
