Bani Chaudhuri starts by narrowing the term. Hallucination is not an AI problem in general, it is specific to large language models, the text tools you type into and get text back from. What those models do is predict what words usually come next, based on patterns learned from an enormous amount of training text. They are not looking anything up.

That explains why a made-up answer is so convincing. The model has learned the shape of how facts are presented, how a citation is formatted, how a confident answer sounds, so invented information comes out looking identical to real information. She names the conditions that make it worse, recent events past the training cutoff, obscure topics, requests for specific numbers and dates, and fields like medicine and law where being wrong actually costs something. The closing thought is the uncomfortable one, that as more generated text gets published, models train on it.

What this video covers

• Why hallucination is specific to language models rather than to AI as a category
• What a model is doing when it answers, and why prediction is not the same as looking something up
• Why invented answers arrive perfectly formatted, citations included
• The four situations where the risk climbs, recent events, obscure topics, specific figures, and high stakes fields
• Prompts and habits that catch it, including giving the model permission to say it does not know
• What happens as future models train on text that earlier models generated

Chapters

• 0:00 Which kinds of AI can hallucinate
• 0:58 What a language model actually is
• 1:58 Where hallucinations come from
• 3:31 When hallucinations get worse
• 4:15 Whether this can be fixed
• 4:39 How to work around it
• 6:31 Choosing the right tool for the job
• 7:21 What happens as models train on generated text