Why do AI models use so many em-dashes? (www.seangoedecke.com)

AI writing’s signature em-dash is surprisingly persistent, and this analysis probes why large language models use so many of them. Common explanations—simple mimicry of web text, token-efficiency, or a generic “safe” continuation strategy—don’t fully fit the facts: attempts to prompt models away from em-dashes often fail, and a Nigerian English corpus shows a lower em-dash rate (0.022%) than historical averages for English (≈0.25–0.275%). The piece also rules out a single dialectal source: while RLHF work in Kenya and Nigeria likely shaped some lexical preferences (e.g., “delve”), it doesn’t explain em-dash prevalence. The most plausible technical hypothesis is a data-shift effect: between GPT-3.5 and GPT-4o, model training appears to have incorporated far more scanned print books—many from the late 19th and early 20th centuries, when em-dash use peaked—so models learned punctuation patterns that are out of step with contemporary prose. Empirical hints: GPT-4o uses ~10× more em-dashes than GPT-3.5, classic works (Moby-Dick) are densely dashed, and other labs’ models show similar behavior. Implications: punctuation can be a persistent artifact of training-era biases, complicating style-control, forensic detection, and RLHF tuning; but the explanation remains partly speculative (synthetic data, RLHF preferences, and undisclosed corpus choices could also matter), so confirmation from insiders or corpus audits would be decisive.