The text prediction still works fine locally without the cloud feature. The data gathering is probably to refine their algorithms. Which I'd expect would include raw text input. Which in a surveillance sense is a bit scary, someone seeing your unfinished thoughts or pre-self-censorship messages.
Hi, thanks for the feedback. SwiftKey Cloud is a secure, opt-in service that allows you to safely backup your language data and sync it across your devices. We take your data security seriously and you can read how we keep your data safe here: http://swiftkey.com/en/data-security/