[Paper Review] AuthentiGPT: Detecting Machine-Generated Text via Black-Box Language Models Denoising

Zhen Guo, Shangdi Yu|arXiv (Cornell University)|Nov 13, 2023

Topic Modeling4 citations

TL;DR

AuthentiGPT is a novel, efficient classifier that detects machine-generated text by leveraging a black-box language model to denoise input text with synthetic noise, then comparing the semantic similarity between original and denoised versions using a single-trainable-parameter classifier. It achieves a 0.918 AUROC score on a biomedical QA dataset, outperforming zero-shot GPT-3.5, GPT-4, GPTZero, and Originality.AI without requiring training data, watermarking, or log-likelihood computation.

ABSTRACT

Large language models (LLMs) have opened up enormous opportunities while simultaneously posing ethical dilemmas. One of the major concerns is their ability to create text that closely mimics human writing, which can lead to potential misuse, such as academic misconduct, disinformation, and fraud. To address this problem, we present AuthentiGPT, an efficient classifier that distinguishes between machine-generated and human-written texts. Under the assumption that human-written text resides outside the distribution of machine-generated text, AuthentiGPT leverages a black-box LLM to denoise input text with artificially added noise, and then semantically compares the denoised text with the original to determine if the content is machine-generated. With only one trainable parameter, AuthentiGPT eliminates the need for a large training dataset, watermarking the LLM's output, or computing the log-likelihood. Importantly, the detection capability of AuthentiGPT can be easily adapted to any generative language model. With a 0.918 AUROC score on a domain-specific dataset, AuthentiGPT demonstrates its effectiveness over other commercial algorithms, highlighting its potential for detecting machine-generated text in academic settings.

Motivation & Objective

• To address the growing challenge of detecting machine-generated text that closely mimics human writing, especially in academic and ethical contexts.
• To develop a detection method that does not rely on large labeled datasets, watermarking, or log-likelihood computation.
• To create a lightweight, adaptable classifier that can be easily applied to any generative language model.
• To improve detection accuracy in domain-specific settings, such as biomedical question-answering, where subtle stylistic differences matter.

Proposed method

• AuthentiGPT introduces synthetic noise (via token masking) to input text at a controlled masking ratio to create a denoising task.
• It uses a black-box language model (e.g., GPT-3.5-turbo) to denoise the corrupted input, generating a semantic reconstruction.
• The algorithm computes the cosine similarity between the original and denoised texts to measure semantic deviation.
• A Gaussian Mixture Model (GMM) is used to cluster the similarity scores, with only one trainable parameter (λ) to separate machine-generated from human-written text.
• The method relies on the assumption that human-written text lies outside the distribution of machine-generated text, making it inherently less predictable under noise.
• The classifier is trained on a small set of 20 examples, making it highly data-efficient and adaptable to new models.

Figure 1: The AUROC scores of AuthentiGPT using different training samples and masking ratios. The x-axis has the number of averaging samples $\beta$ . Each plot shows the scores using different masking ratios $\alpha$ .

Experimental results

Research questions

• RQ1 Can a black-box language model be used to denoise text in a way that reveals distributional differences between human and machine-generated content?
• RQ2 Can semantic similarity between original and denoised text serve as a reliable signal for distinguishing machine-generated from human-written text?
• RQ3 Does a lightweight classifier with only one trainable parameter achieve high detection performance without requiring large-scale training data?
• RQ4 How does AuthentiGPT compare to zero-shot LLMs and commercial detectors like GPTZero and Originality.AI in detecting machine-generated text across diverse datasets?

Key findings

• AuthentiGPT achieves an AUROC score of 0.918 on the aggregated PubMedQA and GPT-generated QA dataset, outperforming zero-shot GPT-3.5 (0.721) and GPT-4 (0.577).
• It surpasses GPTZero (0.797 AUROC) and Originality.AI (0.906 AUROC), demonstrating state-of-the-art performance on the tested dataset.
• The method shows consistent performance across different data types, with particular strength in detecting machine-generated content in new QA generation tasks.
• Performance is sensitive to the masking ratio, with optimal results at α = 0.08, while increasing the number of averaging samples (β) improves statistical reliability.
• The algorithm exhibits limitations in detecting GPT-4 rewrites, suggesting challenges in preserving stylistic cues in rephrased content.
• False positives remain a concern, particularly when human-written text closely resembles model outputs, highlighting a need for further refinement.

Figure 2: The figure on the left is the histogram plots for the cosine similarity of the test datasets before the Box-Cox transformation. The figure on the right is the histogram plots for the cosine similarity of the test datasets after the Box-Cox transformation.