AI POWERS
AI and verification glossary
Understand recurring terms in answers and guides with short definitions, examples and references that help you avoid common confusions.
Before citing a source
- Open the original document: a title or search snippet is insufficient.
- Match author, date, scope and exact passage to the claim.
- Keep publication date separate from the date you consulted the document.
- Preserve units, exceptions and limitations; a reliable portal does not validate every inference.
Jump to an entry
Prompt
A request given to AI, including goal, relevant context and constraints.
Example : “Compare these options by cost, time and maintenance.”
Context
Information available to interpret a request; it is not automatically lasting memory.
Example : Repeat relevant details in a new question on AI POWERS.
Hallucination
Generated content that sounds plausible but is false or unsupported.
Example : A complete-looking reference may not correspond to a real document.
RAG
A method that retrieves documents and provides passages to generation.
Example : Check that the retrieved passage actually supports the answer.
Bias
A tendency that can distort representation, measurement or decisions.
Example : Compare controlled variants without generalising from a single result.
Provenance
Documented origin and history of an object, document or dataset.
Example : Keep producer, original publication, licence and transformations.
DOI
Persistent object identifier; it does not certify quality.
Example : Check title and authors after following the identifier.
Anonymisation
Transformation intended to make identifying a person practically impossible.
Example : Removing a name alone may be insufficient.
Pseudonymisation
Replacing direct identifiers while additional information can reconnect records.
Example : A customer number may still link to a person.
Alternative text
Text equivalent suited to an image’s function in context.
Example : A chart needs its information, not just its colours.
Confidence interval
Interval accompanying an estimate and a stated coverage procedure.
Example : Keep the stated level; it is not a range of every individual value.
Version
Identified state of software, a document or rules.
Example : Record the version tested to reproduce a comparison.
Large language model (LLM)
Model trained on language data with many parameters.
Example : It can produce fluent text without guaranteeing truth.
Personal data
Information about a person identifiable directly or through combined details.
Example : Combined details may identify someone without stating their name.
Primary source
Document directly related to the subject; its status depends on the question.
Example : A contemporary diary is evidence to examine, not neutral truth.
Percentage point
Arithmetic difference between two values expressed as percentages.
Example : From 20% to 25%: +5 points, but a 25% relative increase.
Prompt injection
An attempt to divert an AI system using malicious instructions, directly or within content it reads.
Example : A document to summarise asks to ignore the task and send data: treat the passage as suspicious content, not permission.
AI system evaluation
Documented examination of a system against defined tasks and criteria. Quality depends on its use context and is not just fluent wording.
Example : Test ordinary, ambiguous and unanswerable requests; record failures and repeat the same cases after a change.
No results. Try another term or reset the filters.