Abstract
The rapid advancement of large language models (LLMs) has blurred the line between AI-generated and human-written text. This progress brings societal risks such as misinformation, authorship ambiguity, and intellectual property concerns, highlighting the urgent need for reliable AI-generated text detection methods. However, recent advances in generative language modeling have resulted in significant overlap between the feature distributions of human-written and AI-generated text, blurring classification boundaries and making accurate detection increasingly challenging. To address the above challenges, we propose a DNA-inspired perspective, leveraging a repair-based process to directly and interpretably capture the intrinsic differences between human-written and AI-generated text. Building on this perspective, we introduce DNA-DetectLLM, a zero-shot detection method for distinguishing AI-generated and human-written text. The method constructs an ideal AI-generated sequence for each input, iteratively repairs non-optimal tokens, and quantifies the cumulative repair effort as an interpretable detection signal. Empirical evaluations demonstrate that our method achieves state-of-the-art detection performance and exhibits strong robustness against various adversarial attacks and input lengths. Specifically, DNA-DetectLLM achieves relative improvements of 5.55% in AUROC and 2.08% in F1 score across multiple public benchmark datasets.
Introduction
As Large Language Models (LLMs) evolve, the statistical gap between human-written and AI-generated text is rapidly closing. Traditional detection methods—whether relying on simple probability thresholds or classifiers trained on massive datasets—are increasingly struggling. The primary obstacle is distributional overlap: as LLMs are optimized for human-like fluency, their statistical signatures begin to merge with those of human authors, leading to poor classification performance.
To overcome this, we introduce DNA-DetectLLM, a zero-shot detection method inspired by biological mutation-repair mechanisms. By viewing AI-generated text as a stable "genetic template" and human text as a "mutated" variant, we can measure the intrinsic "repair effort" required to transform any given text back into its most probable AI-generated form.
The Mutation-Repair Paradigm
In molecular biology, DNA replication is a highly stable process, yet mutations can introduce variations. DNA-DetectLLM applies this analogy to text generation. We define an "ideal" AI sequence as a template strand—the sequence generated by an LLM when it consistently selects the most probable token at every step.
Traditional zero-shot detectors often look at static snapshots of probability (like perplexity). In contrast, DNA-DetectLLM focuses on the trajectory of a repair process. Human writing follows a logic that is fundamentally "non-optimal" from an LLM's perspective. Therefore, repairing human text to match an AI template requires significantly more computational effort than repairing slightly "mutated" (edited or noisy) AI text.
Methodology
The DNA-DetectLLM workflow consists of three primary stages following the initial input of text.
- Obtaining the Ideal Sequence: Given an input sequence $s$, we use a reference model (e.g., Falcon-7B) to construct an "ideal" AI sequence $\hat{s}$. This is done by greedily selecting the token with the highest probability at each position.
- Mutation-Repair Mechanism: The input text is treated as a "mutated" strand. The system iteratively identifies tokens that deviate from the ideal sequence and calculates the difficulty of "repairing" them.
- Repair Score-Based Detection: The cumulative repair effort is quantified into a Repair Score. A high repair score indicates that the text deviates significantly and fundamentally from the LLM's internal logic, suggesting it is human-written. Conversely, a low repair score suggests the text is AI-generated.
To ensure practical utility, the researchers derived a simplified scoring formula that approximates this iterative process, allowing for rapid detection (approx. 0.8 seconds per sample) without sacrificing accuracy.
Experimental Evaluation
We evaluated DNA-DetectLLM across diverse datasets, including custom tasks (XSum, WritingPrompts, Arxiv) and public benchmarks (M4, DetectRL, RealDet).
Performance and Robustness
DNA-DetectLLM achieves state-of-the-art results, with an average AUROC of 98.30%. It outperforms existing methods like Binoculars and Fast-DetectGPT by a relative 5.55% in AUROC and 2.08% in F1 score.
One of the method's standout features is its robustness:
- Adversarial Attacks: When subjected to paraphrasing attacks (using DIPPER), DNA-DetectLLM maintained an AUROC of 97.23%, demonstrating resilience where other detectors failed.
- Short Text Detection: Detection accuracy typically plummets for short inputs. However, DNA-DetectLLM maintains a significant lead, outperforming competitors by over 3% on samples as short as 40 tokens.
Conclusion
DNA-DetectLLM introduces a powerful new lens for AI text detection. By shifting the focus from static probability thresholds to the dynamic effort of "repairing" text toward an AI template, we capture a more intrinsic and interpretable signal of authorship. Our results demonstrate that this DNA-inspired paradigm not only improves detection accuracy but also provides the robustness necessary for real-world deployment against increasingly sophisticated generative models.
Citation: Xiaowei Zhu, Yubing Ren, et al. "DNA-DetectLLM: Unveiling AI-Generated Text via a DNA-Inspired Mutation-Repair Paradigm." 39th Conference on Neural Information Processing Systems (NeurIPS 2025). Link to Code
Summary updated