Kamus AI
Hallucination
Model output that sounds convincing but is false or has no basis.
A model produces the most likely text, not checked text. So it can invent names, figures, quotes, article numbers or sources that do not exist, in the same confident tone as a correct answer. Hallucination cannot be removed entirely; what you can do is give sources, ask for references, and check results before use.
Also called: Hallucination, confabulation, made-up answer
Related terms
Appears in 58 lessons
- Where AI helps, and where it does not
- When AI goes wrong in hurried hands
- AI as a highly trained text predictor
- Why models make things up
- Research with AI: sources first, summary second
- Checking the output before you use it
- Our review routine
- From request to finished document
- Checking figures and spreadsheets
- Clear and fair job posts
- Interview scheduling and candidate messages
- Interview notes and offer letters
- Prompts for interviews and candidate communication
- Prompts for offers, onboarding and HR policy
- Not legal advice: AI assists, the advocate answers
- Invented provisions: never cite without opening the text
- Legal research with sources: official text first, summary second
- Tracking deadlines: computed from the rule, checked by two people
- Prompt library: legal research and summaries
- AI does not diagnose
- Writing and tidying SOPs
- Prompt library: SOPs and quality
- Safety decisions stay with people
- Check that the rule is still current
- Handling delivery exceptions
- Shift handover notes
- Following up with vendors and carriers
- Prompts for fleet plans and exceptions
- Prompts for vendors and operations data
- Late route case, part 1: reading the delay data
- Ready-to-use logistics templates
- Figures the model makes up
- Monthly reports and their commentary
- Spreadsheet formulas that get checked
- Invoices and tax documents
- Prompts for reports and commentary
- Month-end close, part 2: report and approval
- Product claims you can back up
- Working out price, margin and fees
- Simple, honest stock notes
- Prompts for catalogues and listings
- Prompts for pricing, stock and promotions
- AI does not decide grades or a student's fate
- Plans and material for each level
- Question banks and rubrics
- Messages to parents and school letters
- Prompts for material and plans
- Prompts for assessment and feedback
- Prompts for communication and admin
- The limits of AI in public service: it assists, it does not decide
- Official letters and announcements in language citizens understand
- Summarising meetings and public input faithfully
- Staff analyses and reports with checked sources and figures
- Prompt library: meetings, minutes and public input
- Code from AI is a draft, and you own it
- Prompts for testing and debugging
- Reading code from AI
- Learning sources that last