The rapid and widespread use of large language models (LLMs) has ushered in the rapid increase in the quantity of AI-produced text everywhere on the Internet, which has necessitated significant concerns about the field of digital forensics, cybersecurity, and academic integrity. The correct distinction between text produced by human authors and text produced by artificial intelligence can be described as a difficult task, which is largely conditioned by the growing fluency of the generative models as well as by the little generalizability of the existing detection tools. The model in question uses a combination of TF-IDF in unigrams and bigrams with 21 linguistically interpretable features. The features are created to match the text and capture it in the stylistic, syntactic, semantic, and discourse coherence levels. They were conducted on the large-scale Human vs. LLM dataset that contains 788,000 text samples of various sources of humans and several advanced and modern language models generated texts. It was tested on the performance of several classification algorithms, such as logistic regression, linear support vector machine, random forest, and LightGBM, but not limited to these products. The experimental outcomes proved that retaining natural linguistic patterns and applying criminal linguistic characteristics played an important role in enhancing detection performance. The LightGBM classifier achieved the best results, with an accuracy rate of 93.4% and an F1 score of 0.93 % for human texts and 0.94 % for AI-generated texts. These findings, moving away from hidden deep learning models, help forensic experts because the new method makes digital investigations more reliable and makes AI more responsible. This transparent tool ensures that textual evidence is ready for court; consequently, study contributes to better safety in the digital world.