In the Afro-Colombian village of San Basilio de Palenque, there are ongoing efforts to revitalize the traditional Spanish-lexified creole language Palenquero. Currently, some instruction in Palenquero is part of the school curriculum, but with few exceptions, there is no grammatical description, no contrasting with Spanish, no emphasis on actual communication and no corrective feedback loop to guide the desired transition to fluency. Nor do traditional adult speakers routinely speak Palenquero to young learners. Cultural factors (e.g. reluctance to criticize community members’ speech) as well as pedagogical limitations (e.g. lack of second-language teaching experience) are obstacles to the scrutiny of learners’ production. This linguistic environment has produced outcomes that would be unexpected in other language revitalization scenarios, constituting a “perfect storm” for unchecked miscues and unanticipated divergences from the target language. The result is two hived-off versions of the Palenquero language within the same community. The reasons for the divergence are clear, but the precise nature of L2 Palenquero remains elusive. The present study describes the results of a series of deep-learning models using recurrent neural networks trained on Palenquero corpora, to determine the extent to which L1 vs. L2 Palenquero can be distinguished without human intervention, and to model the points of divergence between the two, including a classification model that takes a novel utterance as input and predicts the likelihood that it was produced by an L2 speaker. The discussion offers ideas on how the L2 innovations can arise and propagate unnoticed. Even with a corpus that is miniscule in comparison with the input to large-language models, the results are quite encouraging for the use of deep learning techniques to model seemingly intractable sociolinguistic environments.