AI Text Detection Using Embeddings for Mixed Authorship Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The rapid advancement of generative AI models has exacerbated the challenges of detecting AI-generated text, as they can produce hallucinations and replicate human-generated text attributes, making it difficult to accurately determine the authorship of written content.
Innovation Solution
A deep learning architecture that utilizes a combination of embedding models and prediction models to perform trinary classification on documents and sentences, distinguishing between AI-generated, human-generated, and mixed content, by generating token embeddings and aggregating them to obtain document and sentence embeddings for classification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If AI models are used to generate text, then productivity and efficiency are improved, but the ability to detect AI-generated content deteriorates
Solution Approach 1:
The system segments the text into sentences and further into tokens, processing each segment through embedding models to capture local patterns. This segmentation allows the detector to identify AI-generated content at multiple levels (document, sentence, token) by analyzing statistical patterns and linguistic characteristics that differ between human and AI writing styles.
Solution Approach 2:
The patent introduces embedding models as intermediaries that convert text into numerical representations (embeddings). These embeddings serve as a mediator between the raw text and the classification model, capturing semantic meaning and stylistic patterns that enable accurate detection of AI-generated content without requiring direct comparison with reference texts.
2Measurement precision
If AI models replicate human-generated text attributes, then the quality of generated text is improved, but the reliability of detection deteriorates
Solution Approach 1:
The system applies local quality analysis by processing text at multiple granularities (tokens, sentences, documents) and assigning different weights to different regions. The detector identifies local patterns in wording, sentence structure, and paragraph transitions that are characteristic of AI generation, even when the overall text quality is high and resembles human writing.
Solution Approach 2:
The patent changes parameters by transforming text into embedding space where statistical properties can be measured. By converting text to numerical embeddings and analyzing the distribution, statistical patterns, and relationships between embeddings, the system can detect AI-generated content even when the text surface form closely mimics human writing, as AI models tend to produce certain statistical patterns that differ from human writing.
3Measurement precision
If a comprehensive detection system is implemented, then detection accuracy is improved, but device complexity increases
Solution Approach 1:
The system divides the detection task into segments: token-level embedding generation, sentence-level classification, and document-level classification. Each segment is handled by a specialized component (embedding model, sentence prediction model, document prediction model), making the overall complex task manageable through modular architecture where each component focuses on a specific aspect of the detection problem.
Data Source
AI summary
Artificial intelligence generated text detector includes performing operations that include obtaining an input document, parsing the input document into tokens, generating, by an embedding model, token embeddings for the tokens, and aggregating the token embeddings across the input document to obtain a document embedding. The operations further include performing, by a document prediction model processing the document embedding, a document level classification to obtain a set of document class probabilities including an AI generated probability, a mixed probability, and a human generated probability. The operations further include presenting the set of document class probabilities.