AI Hallucination

Definition

In the context of artificial intelligence and large language models (LLMs), hallucination refers to the phenomenon where an AI model generates output that is fluent and confident-sounding but factually incorrect, fabricated, or unsupported by the input data. The term is borrowed loosely from psychology — the model 'perceives' something that isn't there. Hallucinations occur because LLMs are trained to predict statistically likely next tokens (words or word-pieces) given a context, rather than to retrieve verified facts from a database. When a model encounters a gap in its training data or is asked about something ambiguous, it may 'fill in' plausible-sounding but false information rather than admitting uncertainty. Hallucinations can be subtle (slightly wrong dates, misquoted names) or dramatic (entirely invented events or quotes). They are more likely to occur when: - The model is given too little context or a vague prompt - The topic falls outside the model's training distribution - Temperature or top-p sampling settings allow higher randomness - The model is asked to summarize or paraphrase rather than quote directly **Hallucination in video workflows** is particularly relevant for speech-to-text transcription and AI-generated captions. A transcription model that hallucinates may insert words a speaker never said, or subtly alter quotes — which is especially problematic for factual or journalistic content. AI clip detection systems that rely on transcript analysis can also be misled if the underlying transcript contains hallucinated text. OpenClip uses AI for viral moment detection and scoring, grounding its analysis in the actual video transcript rather than generating free-form content — which significantly reduces (though does not eliminate) hallucination risk. Transcription is handled by dedicated speech-to-text models optimized for accuracy, and caption output reflects the spoken audio rather than model-generated text.