Detecting AI-Generated Essays in Indonesian Education
BACKGROUND
This research analyzes semantic embeddings, similarities and linguistic differences between Indonesian-language essays written by high school students and essays generated by AI.
The main focus is not plagiarism in the traditional sense, but how AI-generated texts differ semantically and linguistically from human writing, even when the content appears original.
SCOPE OF WORK
This work was conducted as an undergraduate thesis.
The published journal represents a selected subset of the thesis, focusing on the core methodology and key experimental results, while the thesis includes broader analysis, additional experiments, and system implementation.
DATA & PROBLEM SETUP
Essays written by Indonesian high school students (SMA/SMK/MA)
Essays generated by AI using controlled prompts
All texts are written in Bahasa Indonesia
Essays are segmented into paragraphs and sentences for fine-grained analysis
The core problem addressed is whether semantic representations and writing style can reliably differentiate student-written and AI-generated texts.
METHODOLOGY
The research applies a semantic similarity and classification pipeline consisting of:
IndoBERT embeddings to capture contextual meaning
Triplet Loss to learn semantic distance between student and AI essays
Cosine similarity to measure semantic closeness
Linguistic features such as lexical diversity, sentence length, modal usage, and stopword ratio
These features are combined into a neural network classifier.
DETAILS
Training & Evaluation
Dataset split: 75% training, 5% validation, 20% testing
Experiments conducted at paragraph and sentence levels
Evaluation metrics include:
Area Under Curve (AUC)
Accuracy
Precision
Recall
F1-Score
This setup ensures balanced evaluation and generalization testing.
KEY FINDINGS
Student essays and AI-generated essays form distinct semantic clusters with some outliers
Semantic similarity between same-class essays is consistently high
Cross-class similarity is significantly lower, indicating clear separation
Student essays show higher linguistic variation, while AI essays are more uniform and consistent
These results indicate that semantic modeling and linguistic analysis are effective for distinguishing AI-generated text in Indonesian.
0.99
AUC (Paper)
90.55%
Accuracy (Paper)
0.98
AUC (Thesis)
98%
Accuracy (Thesis)
OUTCOMES
The trained model was successfully implemented into a web-based application (prototype) that allows users to:
Input an essay text
Analyze semantic similarity
Classify the text as student-written or AI-generated
This demonstrates real-world applicability in educational contexts.
RESEARCH CONTRIBUTIONS
Provides empirical evidence for AI-text detection in Bahasa Indonesia
Combines semantic similarity and linguistic analysis in one framework
Extends AI-detection research beyond English-language datasets
Serves as a foundation for future academic integrity tools
NEW PROTOTYPE
Latest Prototype Deployed – Built using AI Tools from Google (Stitch) for rapid UI iteration
Agentic Development – Powered by IDE Agent Antigravity to streamline structured development
Containerized Deployment – Packaged with Docker for consistent and reproducible environments
Public Demo – Published via Hugging Face Spaces for easy access and testing
VIEW MORE
Let's Build Something Meaningful
I’m always open to meaningful conversations, from new ideas, collaborations, to exchanging thoughts about building better digital experiences.