TextGPT 2026: AI vs Human Text Classification

Background

As artificial intelligence becomes increasingly adept at generating human-like text, distinguishing between AI-generated and human-written content has real-world importance in academic integrity and plagiarism detection, online content moderation, fake news and misinformation detection, authorship attribution, and AI safety and accountability.

Core Objective

Participants must build models that classify sentences as human-written or AI-generated using a labeled dataset.

Label	Meaning
1	Human-written
0	AI-generated

Dataset Description

Using the Human vs AI Sentences dataset from Hugging Face. This dataset includes thousands of short text samples labeled according to whether they were written by a human or generated by an AI.

Balanced classes, short to medium length text, mixed topics, diverse writing styles.

Challenge Tasks

1. Binary Text Classification

Build a model that predicts 1 (Human-written) or 0 (AI-generated).

Traditional NLP

• TF-IDF + Logistic Regression
• n-grams + SVM
• Bag-of-Words
• Feature engineering

Deep Learning

• LSTM / GRU

2. Explainability (Bonus)

In addition to classification accuracy, teams may be evaluated on how well they explain their model decisions.

• Feature importance
• Attention visualization
• Model introspection

Transformer-based Approaches

• BERT / RoBERTa / DistilBERT / others
• Fine-tuned LLMs

Evaluation Metrics

Primary Metrics

• F1-score
• Accuracy

Secondary Metrics

• Precision
• Recall
• AUC-ROC

Scoring Breakdown

Classification F1-score

40%

Accuracy

30%

Validation Methodology

10%

Rules & Constraints

Data Split

Split	Usage
Train	70%
Validation	15%
Test	15%

Key Rules

• No future data usage or leakage
• Models must be reproducible
• All preprocessing steps must be documented
• External pretrained models allowed but disclosed
• Use of proprietary datasets prohibited