Evaluation & Quality
Ablation study
Testing what changes when part of an AI system is removed or altered.
Example
A team checks whether an AI helper does worse when it cannot search supporting documents.
Why people use it
It helps teams find which parts of a system actually contribute to a result.
What you'll hear
“What happens if we remove that step?”
What this means for you
Compare controlled variants on representative tasks.
Can you control it?
Developer-only
The people building or running the AI choose this setup. An everyday user generally needs their help to change how this part works.
Common questions
- Does worse performance prove the component always helps?
- No. The result applies to the tested setup and may involve interactions.
- Does no change mean the removed part never matters?
- No. It may matter for cases that the particular test did not include.
- Can removing two parts differ from removing each alone?
- Yes. Parts can support or compensate for one another.