An approachable 2023 explanation of fluent but unsupported model outputs, with examples of contradiction, fabrication, and irrelevance. Its prompting advice is useful as risk reduction, but the correct habit is still to verify consequential claims against reliable sources.
Plan for this page
Watch actively with preparation, signposts, and a comprehension check.
You will leave with
What practitioners call hallucination, several forms it can take, contributing conditions, and a few ways users try to reduce it.
Time
10 min
Before you begin
Everyday users who need a first explanation before learning a stronger verification workflow.
Do this now
Read the terms and attention points before loading the video.
Why Large Language Models HallucinateWhat practitioners call hallucination, several forms it can take, contributing conditions, and a few ways users try to reduce it.
00:00 · Three plausible falsehoods
01:43 · Forms of hallucination
03:32 · Why it can happen
04:56 · Generation trade-offs
05:29 · Input context
What it teaches
What practitioners call hallucination, several forms it can take, contributing conditions, and a few ways users try to reduce it.
Everyday users who need a first explanation before learning a stronger verification workflow.
Pay attention to
Compare fluent wording with whether a claim is actually supported.
Separate possible contributing factors from a complete mechanical explanation.
Treat better prompts as mitigation, not a guarantee of truth.
A confident, coherent answer can still be unsupported or false.
Prompt clarity can reduce some failures but cannot turn a generator into a truth source.
For consequential use, require sources, inspect them, and verify the claim outside the model.
Active check
A chatbot gives a polished answer with dates and citations. What should you conclude?
Caveats
This 2023 explainer gives a simplified list of causes; hallucination mechanisms and evaluation methods remain active research areas and vary by model and task.
Clear prompts, lower temperature, or examples may improve consistency, but none proves a claim correct.
The video underplays retrieval, grounded tool use, calibrated abstention, source inspection, and formal evaluation as system-level mitigations.
Accessibility
Manually supplied English captions and an English transcript were available. No manually supplied French caption track was confirmed. On-screen labels and examples reinforce the narration but are mostly stated aloud.
inAi