AI evidence extraction

Authorship Attribution in the Era of LLMs: Problems, Methodologies, and Challenges

Baixiang Huang, Canyu Chen, Kai Shu · ACM SIGKDD Explorations Newsletter · 2025

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

46.40

FWCI

Top 10%

Impact

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.1145/3715073.3715076

Methodology & findings

Study design

Narrative literature review systematically categorizing four representative problems in authorship attribution: (1) Human-written Text Attribution; (2) LLM-generated Text Detection; (3) LLM-generated Text Attribution; and (4) Human-LLM Co-authored Text Attribution.

Main result

The study found that "the rapid advancements of Large Language Models (LLMs) have blurred the lines between human and machine authorship, posing significant challenges for traditional methods." The review systematically categorizes authorship attribution into four representative problems: (1) Human-written Text Attribution; (2) LLM-generated Text Detection; (3) LLM-generated Text Attribution; and (4) Human-LLM Co-authored Text Attribution. Additionally, "Neural network-based detectors, generally outperform metric-based methods in both human authorship attribution and LLM-generated text detection problems" though "these neural network approaches often offer less explainability compared to their metric-based counterparts."

Research paradigm

Positivist/interpretive synthesis of computational and empirical literature

Author conclusions

The authors conclude: "The field of authorship attribution is experiencing both unprecedented challenges and remarkable opportunities with the advent of LLMs. Whether the objective is to identify human authors, differentiate between human and machine-generated texts, attribute texts to specific LLMs, or manage the complexities of human-LLM co-authored texts, ongoing innovation is imperative. Effectively addressing these multifaceted issues requires interdisciplinary approaches and collaborative efforts among researchers." They emphasize that "By integrating robustness, explainability, and interdisciplinary perspectives, the insights gained are not only accurate but also socially relevant and trustworthy."

Risk of bias

Selection bias in included literature based on availability and accessibility of published papers; Publication bias toward successful methods and novel approaches; Language bias toward English-language papers and English-centric datasets; Potential under-representation of smaller-scale or negative result studies; Selection bias: The review's scope is limited to published literature, potentially excluding unpublished or grey literature on authorship attribution; Publication bias: Only papers that were published and indexed are included; Language bias: The focus on English-language authorship attribution may exclude non-English research; Methodological bias: The review acknowledges that "detectors could be biased against non-native English writers"; Selection bias in literature reviewed (publication bias toward successful methods); Potential geographic/language bias toward English-language and Western research; Funding source bias (government and corporate funding acknowledged)

Limitations

The review identifies several critical limitations: "These detectors require retraining when encountering text from new LLMs to ensure reliable detection," and "LLM-generated text detectors often struggle to generalize to unseen domains encountered during training." Additionally, "existing detectors also lack robustness to various factors, such as alternative decoding strategies, input sequence length, different prompts, repetition penalties, and human edits." The paper also notes that "Traditional stylometric methods rely on human expertise and manually crafted features, whereas deep learning methods demand significant computational resources and extensive labeled data, with the risk of catastrophic forgetting."

Open questions raised

Need for robust methods that generalize across domains, genres, and languages
Improved explainability and transparency in attribution model decisions for legal and forensic applications
Development of standardized benchmarks encompassing diverse text types and sources
Better handling of adversarial attacks and paraphrasing techniques
Methods robust to out-of-distribution LLMs and languages not encountered during training
Detection and prevention of malicious activities including misinformation, plagiarism, and propaganda

Data: OpenGPTText dataset (Chen et al., 2023a); MIXSET dataset (Zhang et al., 2024); MIXSET dataset (Zhang et al., 2024) - includes text refined by LLMs through polishing, completion, and rewriting operations; Curated list of papers regularly updated at https://llm-authorship.github.io/; Resources available at https://llm-authorship.github.io/ (curated list of papers and resources, regularly updated)Code: https://llm-authorship.github.io/ (curated list of papers and resources, regularly updated); https://llm-authorship.github.io/ (curated resource collection)Extracted from: pdfAgreement 58%

Explore related topics

Authorship & Attribution Research Integrity Literature Review