You know how you text your friends in lowercase with no punctuation, but the moment you draft a LinkedIn post, you instinctively capitalize and add semicolons? That's code-switching — adjusting your register to the perceived audience and formality of the channel. Arabic speakers have been doing a more extreme version of this for centuries, toggling between Standard Arabic (the prestige 'High' variety used in news, religion, and official contexts) and Colloquial Arabic (the everyday 'Low' varieties that differ by region). This paper asks: does that ancient toggle survive on social media, or does the internet flatten it? The committed claim: digital platforms don't erode Arabic diglossia — they reconfigure it. X (formerly Twitter) attracts significantly more Standard Arabic use than Facebook across most discourse categories, and the effect isn't uniform across topics. The study collected 16,754 public posts via Python scraping, retained 10,000 (5,000 per platform), and manually classified them into seven discourse domains: politics, technology, science, business, culture, fun, and sports. Posts were then coded as Standard or Colloquial Arabic. The statistical machinery is classical: Chi-square tests, Cramér's V for effect size, and binary logistic regression with platform-by-domain interaction terms. The numbers are striking in their asymmetry. On X, the association between discourse category and variety choice is moderate (Cramér's V = .347). On Facebook, it's large (V = .500). The platform-level association is smaller but significant (V = .162). The logistic regression is where the real story lives: the odds of Standard Arabic on X versus Facebook are 1.31× higher for politics — a modest tilt. But for sports, the odds ratio explodes to 26.71×, meaning sports posts on X are overwhelmingly Standard Arabic while Facebook sports discourse is almost entirely Colloquial. Fun content shows a similar gap (OR = 6.34). Science and technology flip direction — Facebook actually shows slightly higher odds of Standard Arabic in those domains (OR = 0.41 and 0.71 respectively). The architecture here is straightforward corpus sociolinguistics: manual annotation of a stratified sample, bivariate association tests, and regression with interactions. There is no NLP classifier in the pipeline — the coding was done by the researchers. This is both a strength (no model error to worry about) and a constraint (scalability ceiling, potential coder bias). The statistical approach is appropriate for the question, though the absence of multilevel modeling (accounting for user-level clustering) is a notable gap. If prolific users skew toward one variety, the independence assumption underlying Chi-square and logistic regression is violated. Integrity is mixed. The dataset is reasonably large for manual sociolinguistics, but the sampling methodology — keyword-based Python collection of public posts — introduces selection bias that the paper doesn't deeply interrogate. Which keywords? In which variety? Were the keywords themselves biased toward Standard Arabic, creating a floor effect on Colloquial representation? The coding scheme (binary SA/CA) flattens a rich continuum; many Arabic social media posts blend registers within a single sentence, and the paper doesn't address how mixed-register posts were handled. No inter-annotator reliability statistics are reported. The benchmarks are internal — there's no comparison to prior corpus studies of Arabic digital diglossia. The broader contribution is confirming what sociolinguists suspected but hadn't quantified at this scale: Ferguson's 1959 diglossia framework doesn't collapse online; it gets remapped. The platform's affordances — character limits, audience composition, perceived formality — become the new contextual triggers for register selection, replacing the mosque-versus-marketplace distinction of pre-digital Arabic sociolinguistics. The interaction effects (sports wildly different from science) suggest it's not just platform prestige but platform-by-domain combinations that drive variety choice, which is a genuinely useful finding for computational sociolinguistics and Arabic NLP pipelines that need to handle register variation. What's missing is any engagement with the NLP side of this question. Modern Arabic NLP struggles precisely because Standard and Colloquial Arabic are linguistically distant enough to break models trained on one variety when applied to the other. This paper generates evidence that could directly inform dialect-aware model training — which platforms and domains to sample for which variety — but doesn't make that connection. The next obvious experiment is running a dialect classifier on a much larger sample and validating these manual findings at scale.
