If anybody using them in Python 3.3 is already depending upon them returning strings, changing it to return bytes will break their code... so that ship as unfortunately sailed.  Changing that only in 3.4 would just cause churn and make code using this more difficult to be compatible across both.

Updating the documentation to state the utf-8 decoding behavior and the caveats of that is important.  That documentation note should also mention what people should use instead if they want binary data instead of text.