This may depend heavily on what your purpose is in learning the language. The pragmatics of using a language are often not covered exhaustively by learning resources, human or otherwise. So if the primary purpose of learning is to enable two-way communication with members of the target culture, too much decontextualized learning (in which the focus is placed on syntax and semantics but not pragmatics) or miscontextualized learning (in which there is a focus on pragmatics, but not in the target communicative context) could teach the learner inappropriate patterns, which would be detrimental - and possibly difficult to reverse - in real conversation.
However, the fact that content management is dependent on both interlocutors may help with this problem on the infrequent occasion that one encounters a native speaker. This paper (PDF), for example, considers the specific concerns raised using a test case of Japanese-second-language speakers in Japan. It contrasts theoretical failures to assume the appropriate voice with perceived failures in achieved communication.
It is necessary to consider that both the native and the non-native speakers are engaged in the act of impression management in conversational interactions and that through language use and behavior one's subjectivity is co-constructed. (emphasis mine)
language learner pragmatic inappropnateness in Japan is not
necessarily viewed as failure during actual conversational encounters
Native speakers interacting with non native speakers play an active role in building the assumptions on which the conversation is based, so they may compensate for theoretical failures by including in those assumptions the fact that they are speaking to a second language learner, thereby mitigating the effects of learning out of context. The lack of actual detriment in this case is found even though the lapses in appropriateness are substantial.
And if the intended purpose of learning the language is for one-way comprehension alone (e.g. reading documents for research), this effect could be totally absent, as long as one had learned sufficient abstract rules to describe the samples being consumed.