How Does Preds Detect AI-Generated Text?

Detailed technical documentation and research findings

Overview

An unprecedented volume of AI-generated content is produced every day. This rapid growth is driven by the widespread accessibility of advanced large language models to non-technical users. Consequently, robust algorithms capable of distinguishing human from machine-authored content have become essential. The rapid evolution of LLMs raises critical questions around copyright integrity, misinformation, and above all, academic evaluation. Our solution was specifically engineered to address this challenge. We began researching this domain in August 2023 (initially in English), providing us with more than two years of specialized engineering experience.

The Preds model was architected to achieve the lowest possible false positive rate, ensuring that genuine human writing is virtually never misclassified as AI-generated. Our proprietary training methodology minimizes false classifications by analyzing subtle statistical signatures, token transition dynamics, and linguistic anomalies characteristic of machine synthesis.

Methodology & Architecture

At the core of our solution lies a transformer encoder architecture fine-tuned for the Hungarian language. The language model segments input text into subword tokens and maps them into contextual vector representations. A token represents a word or subword fragment. The transformer dynamically weights these tokens relative to their contextual dependencies, evaluating semantic consistency, syntactic coherence, and stylistic nuances typical of generative AI.

These dense contextual embeddings flow through a custom neural classification head yielding a calibrated binary prediction: 0 for human, 1 for AI-generated. Standard classification heads alone cannot meet production demands—Preds therefore incorporates multiple research-grade optimization techniques to drive false positive rates down to near zero.

Model Training & Corpora

The model was trained on more than 350,000 documents drawn from diverse sources: authentic human writings (academic essays, forum discussions, journalistic articles) and synthetic texts generated by over 40 distinct large language models (including GPT, Claude, Gemini, LLaMA, and Mistral). The dataset was rigorously balanced to ensure the network internalizes genuine linguistic abstractions rather than superficial stylistic artifacts. Key training techniques include: Hard Negative Mining (HNM): The model was iteratively retrained on challenging edge cases that it previously misjudged. This sharpened its ability to discern complex, eloquent human writing from sophisticated AI output. Mirror Prompts: Portions of synthetic samples were constructed using mirror prompts (matching the genre, tone, and prompt constraints of human sources, with paraphrasing variations). This prevents reliance on specific keywords, training the network to focus on structural grammar and deep semantics. Additional proprietary loss functions and custom regularization heuristics are integrated into our fine-tuning pipeline to prevent overfitting. The result is a robust, highly sensitive detector attuned to the morphological and syntactic intricacies of Hungarian, yielding exceptionally low false positive rates in real-world deployments.

Image placeholder

Benchmark Results & Evaluation

During the evaluation phase, the model was tested on over 3,000 held-out, unseen samples. The current production version of Preds achieves: • **Accuracy:** 97.4% • **F1-Score (AI vs. Human):** 0.961 • **False Positive Rate (FPR):** 0.008 (< 1%) • **AUROC:** 0.992 These metrics represent state-of-the-art performance in both academic benchmarks and production environments. The system produces multi-tiered confidence scores, providing users with transparent reliability indicators alongside each prediction.

Image placeholder

Ready to get started?

Experience our AI detection technology today.