interesting idea! it is possible, but with a number of limitations:
1. it would be style specific (each style would be hard-coded), unless there are scripts to recognize a CSL style out there.
2. it might not work 100% of the time, especially if formatting is taken into account. For example your reference @luciano seems like it's from Journal of Evolutionary Biology, but the volume is not bolded?
3. the BibTeX key creation algorithm would not check for duplicates (unless the workflow has access to the whole library).
In conclusion, one could use an existing library to recognize references from text (but I am not aware of any), or this can be done efficiently only if you expect to convert a few CSL styles that can be hard coded.