A company wants to let users verify the trustworthiness of a generative AI's answers. What is the practice of explicitly showing the references (citations) and data sources to indicate which materials an answer is based on called?
A question about choosing the practice of indicating the sources of an answer.
Watermarking
Watermarking is a technique that embeds a mark (watermark) in generated output to indicate it was AI-produced.
Its purpose is identifying generated output, and it is not the practice of showing the source materials of an answer, so this is incorrect.
Data lineage
Data lineage is a concept that internally tracks the origin and transformation history of data such as training data.
It is not the practice of showing users the references of an answer, so this is incorrect.
Documenting source citations and data attribution.
Correct. By explicitly documenting the references (citations) and data sources an answer is based on, users can verify the basis, increasing transparency and trustworthiness. With RAG, the referenced documents can be shown as sources.
Model card
A model card documents the uses, performance, and risks of the model as a whole.
It does not indicate the source materials of an individual answer, so this is incorrect.
Note the correct answer, documenting source citations and data attribution.
- The practice of explicitly documenting the references (citations) and data sources an answer is based on.
- Users can verify the basis, increasing transparency and trustworthiness (with RAG, the referenced documents can be presented as sources).