Over the past decade, artificial writing has become increasingly indistinguishable from human writing. AI detectors have focused predominantly on prose writing [1], leaving a critical gap in research regarding poetry. Additionally, current detectors (e.g, GPTZero, GrammarlyAI) are trained with primarily English datasets that may reach 1,250,000 individual entries [1]. As this restricts the feasibility of detection in non-English speaking and high-latency areas, we resolved to create an efficient, lightweight detector with pre-built, alterable multilingual datasets. We present three models that allow us to 1. detect, with a lightweight detector, whether a poem is created by humans or AI, 2. understand the effects hyperparameters have on accuracy, and 3. determine the feasibility of nonromanized poem detection for multiple languages. Via our curated datasets, our lightweight models achieved accuracy rates 35% higher than previous research with outputs accelerated by at least two orders of magnitude, paving the way for authenticity verification in highlatency, non-English-speaking areas.