Perkembangan teknologi Artificial Intelligence (AI) telah memungkinkan sistem menghasilkan tulisan yang menyerupai tulisan tangan manusia. Hal ini menimbulkan tantangan dalam membedakan keaslian tulisan tangan, khususnya pada tugas mata kuliah yang masih banyak menggunakan tulisan manual. Penelitian ini bertujuan untuk merancang dan mengimplementasikan model Convolutional Neural Network (CNN) yang dapat mendeteksi tulisan tangan berbasis AI pada tugas akademik. Metode yang digunakan dalam penelitian ini adalah CNN dengan pendekatan berbasis YOLO yang dilatih untuk mendeteksi dua jenis tulisan, yaitu tulisan tangan manusia (handwriting) dan tulisan berbasis AI (botwriting). Proses pelatihan dilakukan dengan variasi jumlah epoch secara bertahap, yaitu 20, 40, 60, 80, dan 100 epoch. Evaluasi performa model dilakukan menggunakan metrik precision, recall, serta confusion matrix. Hasil penelitian menunjukkan bahwa peningkatan jumlah epoch berpengaruh signifikan terhadap performa model. Model pada 20 epoch masih berada pada tahap awal pembelajaran, sedangkan pada 40 epoch mulai menunjukkan kestabilan. Performa optimal dicapai pada 60 hingga 100 epoch dengan kondisi konvergen. Berdasarkan hasil evaluasi pada epoch ke 100, model memperoleh nilai accuracy sebesar 99,669%, precision sebesar 99,342%, recall sebesar 99,665%, dan F1-score sebesar 99,668%, yang menunjukkan bahwa model memiliki kemampuan klasifikasi yang sangat baik dengan tingkat kesalahan yang sangat rendah. Namun, pengujian lebih lanjut pada data baru dari luar distribusi training menunjukkan adanya kesenjangan performa (generalization gap) yang cukup signifikan, dengan rata-rata accuracy turun menjadi 80%, precision 86,44%, recall 80%, dan F1-score 81,25%, terutama disebabkan oleh rendahnya recall kelas botwriting (66%) akibat keterbatasan variasi jenis font botwriting pada data training.

Kata Kunci: Artificial Intelligence (AI), Convolutional Neural Network (CNN), Handwriting (Tulisan Tangan), Botwriting (Tulisan AI), You Only Look Once (YOLO).

The development of Artificial Intelligence (AI) technology has enabled systems to generate writing that resembles human handwriting. This poses a challenge in distinguishing the authenticity of handwriting, especially in academic assignments that still widely use manual writing. This study aims to design and implement a Convolutional Neural Network (CNN) model capable of detecting AI-based writing in academic tasks. The method used in this study is a CNN with a YOLO-based approach, trained to detect two types of writing, namely human handwriting (handwriting) and AI-generated writing (botwriting). The training process was carried out with gradual variations in the number of epochs, namely 20, 40, 60, 80, and 100 epochs. Model performance evaluation was conducted using precision, recall, mean Average Precision, and a confusion matrix. The results show that increasing the number of epochs has a significant effect on model performance. The model at 20 epochs is still in the early learning stage, while at 40 epochs it begins to show stability. Optimal performance is achieved at 60 to 100 epochs under convergent conditions. Based on the evaluation results at epoch 100, the model achieved an accuracy of 99.669%, precision of 99.342%, recall of 99.665%, and an F1-score of 99.668%, indicating that the model has very high classification capability with a very low error rate. However, further testing on new data outside the training distribution showed a fairly significant performance gap (generalization gap), with average accuracy dropping to 80%, precision 86.44%, recall 80%, and F1-score 81.25%, mainly due to the low recall of the botwriting class (66%) because of limited variation in botwriting font types in the training data.

Keywords: Artificial Intelligence (AI), Convolutional Neural Network (CNN), Handwriting, Botwriting, You Only Look Once (YOLO).