The Machine's Hand: A Survey of Stylometry and Forensic Attribution in LLM Text

Stylometry offers one of the most useful lenses for studying whether generated text carries measurable traces of its source. In the setting of large language models (LLMs), those traces may include lexical, syntactic, punctuation, discourse-marker, curvature, likelihood, and feature-distribution patterns. This paper is a synthesis and methodological survey. It does not report new experiments, propose a new detector, or claim new information-theoretic theorems. Instead, it collects established ideas from authorship attribution, AI-generated-text detection, LLM-generated-text attribution, classical hypothesis testing, and recent work on prompt sensitivity and paraphrase evasion, and it translates them into a common notation for forensic use. The survey's main practical recommendation is to treat LLM source attribution as a calibrated, prompt-conditioned, pairwise statistical problem rather than as a search for a universal fingerprint. Fixed word-count thresholds should be replaced by prompt-stratified error curves, confidence intervals, and empirical estimates of the hardest source-pair separation. Stylometric evidence can be strong in controlled domains, but it is domain-bound and can be weakened by prompts, short outputs, template constraints, paraphrasing, normalization, model drift, and open-set uncertainty. The defensible conclusion is therefore conditional: attribution becomes credible only to the extent that the chosen observation pipeline preserves measurable separation among the candidate sources in the relevant domain.