Definition
The CAP theorem is a result about distributed systems stating that when a network partition occurs, a system must sacrifice either consistency (every read sees the latest write) or availability (every request gets a response). Partition tolerance is not optional — in any real network, partitions will happen — so the practical content of the theorem is the choice a system makes while a partition is in progress.
A CP system rejects requests it cannot serve safely, staying consistent at the cost of some downtime; an AP system keeps answering with possibly stale data. Most production databases are configurable along this spectrum rather than purely one or the other, and the choice can differ per operation: a banking ledger wants CP behavior, a social feed tolerates AP happily.
How it comes up in interviews
Interviewers rarely ask for the definition — they ask you to apply it: "during a partition, what does your design do?" Strong answers name the choice explicitly per data type ("the payment path fails closed, the feed serves stale") instead of claiming a system is "CA", which is not a real option in a distributed setting.