LLM Authorship Impersonation Fails to Fool Verification

A new research paper published on arXiv reveals a significant finding for the digital authenticity community: despite the impressive text generation capabilities of modern large language models, using LLM prompting to impersonate another author's writing style does not successfully evade established authorship verification methods. The study offers reassurance that stylometric analysis remains a viable defense against AI-powered identity manipulation in written content.

The Growing Threat of AI-Powered Authorship Impersonation

As LLMs like GPT-4, Claude, and others have grown increasingly sophisticated, concerns have mounted about their potential use for impersonating specific individuals' writing styles. Authorship impersonation — where an adversary attempts to write text that appears to originate from a specific target author — represents a serious threat to digital authenticity. From fabricated social media posts to forged documents, the ability to convincingly mimic someone's prose could undermine trust in written communication at a fundamental level.

This form of synthetic text generation is conceptually parallel to deepfakes in the visual and audio domains. Just as face-swapping technology can make someone appear to say things they never said, LLM-powered authorship impersonation could make someone appear to have written things they never wrote. The implications span journalism, legal proceedings, academic integrity, and political discourse.

How the Study Was Conducted

The researchers investigated whether LLM prompting strategies — including instructing models to write in the style of a specific author, providing samples of the target's writing as context, and other prompt engineering techniques — could produce text that would fool authorship verification (AV) systems. Authorship verification is a well-established field in computational linguistics and natural language processing that uses statistical and machine learning methods to determine whether a given text was written by a claimed author.

These AV methods typically analyze stylometric features — quantifiable characteristics of writing style such as sentence length distributions, vocabulary richness, function word usage patterns, punctuation habits, and syntactic structures. More advanced approaches leverage deep learning models trained on large corpora to capture subtle stylistic fingerprints that are difficult for humans to consciously control or replicate.

The study tested multiple prompting strategies across different LLMs, pitting the generated impersonation texts against a range of authorship verification approaches to assess whether any prompting technique could consistently defeat these detection systems.

Key Findings: Verification Methods Remain Robust

The central finding is striking: LLM-generated impersonation texts were consistently detected by authorship verification methods. Despite the surface-level fluency and apparent stylistic similarity that LLMs can achieve when prompted to write like a specific author, the deeper statistical patterns that AV systems analyze remain distinguishable from genuine authorial output.

This suggests that while LLMs can capture high-level stylistic features — tone, vocabulary choices, topic framing — they struggle to replicate the full spectrum of subtle, often unconscious linguistic patterns that constitute an author's true stylometric fingerprint. The micro-level statistical distributions of linguistic features appear to remain uniquely human, at least when LLMs are guided purely through prompting rather than fine-tuning.

Implications for Digital Authenticity

This research carries significant implications for the broader digital authenticity ecosystem. It demonstrates that traditional computational forensic methods retain meaningful power even in the age of advanced generative AI. Just as deepfake detection researchers have found that visual artifacts and statistical anomalies persist in AI-generated video and images, this study confirms an analogous phenomenon in the text domain.

For organizations concerned about impersonation attacks — media outlets verifying sourced documents, platforms authenticating user-generated content, or legal teams establishing document provenance — the findings provide a degree of reassurance. Existing authorship verification tools can serve as a meaningful layer of defense against LLM-powered impersonation.

Caveats and Future Directions

It is important to note that this study specifically examines prompting-based impersonation. More sophisticated attacks — such as fine-tuning a model on a target author's corpus or combining prompting with post-processing techniques designed to match specific stylometric distributions — could potentially pose greater challenges to verification systems. As adversarial techniques evolve, so too must detection methods.

The arms race between synthetic content generation and authentication continues across every modality — video, audio, images, and text. This paper provides an important data point: in the text domain, the defenders currently maintain a meaningful advantage against prompting-based attacks, offering a foundation for building more robust content authentication systems.

Stay informed on AI video and digital authenticity. Follow Skrew AI News.