Frankentext: Stitching random text fragments into long-form narratives

Chau Minh Pham, Jenna Russell, Dzung Pham, Mohit Iyyer

Introduction

Recent advancements in Large Language Models (LLMs) have shifted the focus from simple text generation to more complex forms of creative synthesis. Researchers have introduced Frankentexts, a novel generation paradigm where LLMs act as "composers" rather than "authors." In this framework, the model is tasked with stitching together thousands of random human-written snippets into coherent, long-form narratives. Even under extreme constraints—where up to 90% of the output must be copied verbatim from source material—these "stitched" stories have been shown to outperform standard AI writing in quality, creativity, and the ability to evade detection.

The Challenge of "AI Slop"

Standard LLM generation often results in what critics call "AI slop"—prose that is formulaic, predictable, and lacks the nuanced creative spark of human authorship. While techniques like Retrieval-Augmented Generation (RAG) provide factual grounding, they are rarely used to drive the creative "voice" of a story.

Furthermore, the proliferation of AI-generated content has sparked an arms race between generators and detectors. Frankentext explores a middle ground of mixed-authorship. By forcing an LLM to construct a narrative almost entirely from existing human-written fragments, the researchers test whether AI can navigate a massive combinatorial search space to create something novel that feels authentically human, while simultaneously challenging traditional definitions of authorship and copyright.

Methodology: The Composer Pipeline

The creation of a Frankentext involves a multi-stage pipeline designed to balance verbatim constraints with narrative flow:

  1. Snippet Sampling: The model is provided with a specific writing prompt and a pool of thousands of randomly sampled paragraphs (snippets) from a large corpus, such as Books3.
  2. Drafting: The LLM is instructed to "stitch" these snippets into a story of approximately 500 words. The primary constraint is that the vast majority of tokens must be verbatim copies of the provided snippets, with the LLM providing only the necessary connective tissue.
  3. Iterative Polishing: To mitigate "tonal whiplash" or grammatical inconsistencies caused by jumping between random snippets, the model performs up to three rounds of self-editing. This process improves coherence while strictly maintaining the verbatim copy constraint.
  4. Agentic Retrieval: In advanced iterations, the authors utilized a Model Context Protocol (MCP) tool, allowing the LLM to actively search a semantic index of 160 million snippets to find the ideal "building blocks" for the narrative.
The Frankentext generation pipeline: from random sampling to iterative refinement.

Key Results and Evaluation

The performance of Frankentexts was evaluated against "vanilla" (standard) AI generations using both human judges and LLM-based metrics.

Superior Quality and Diversity

In human evaluations, Frankentexts were consistently preferred over standard AI outputs across several dimensions, including plot development, creativity, and language use. On a 7-point Likert scale, an LLM judge rated Frankentexts at 4.21, significantly higher than the 3.18 score assigned to vanilla outputs. Additionally, Frankentexts exhibited higher diversity, producing more varied content clusters (2.74) compared to vanilla generations (1.76) for identical prompts.

Comparison of average scores between Vanilla AI generation and Frankentext across qualitative metrics.

Detection Evasion

Because the bulk of the text is composed of human-authored prose, current state-of-the-art AI detectors struggle to identify Frankentexts. In testing against the Pangram detector, Frankentexts produced by Gemini 2.5 Pro were misclassified as human-written 72% of the time. This highlights a significant blindspot in current detection methodologies that rely on identifying the statistical signatures of LLM-generated transitions.

Discussion and Limitations

Despite their success, Frankentexts face several hurdles:

  • Coherence: Human annotators noted that stories can occasionally suffer from abrupt tonal shifts or uneven grammar at the junctions where snippets meet.
  • Computational Cost: The process is currently 100–200x more expensive than standard decoding due to the massive context windows required to process thousands of snippets simultaneously.
  • Legal and Ethical Concerns: As the method involves verbatim copying of copyrighted works, it occupies a precarious legal position regarding derivative works and fair use. This raises the "Ship of Theseus" authorship paradox: if the AI orchestrates the arrangement of human materials, who is the true author?

Conclusion

Frankentext demonstrates that LLMs can achieve higher levels of creative quality by acting as curators of human expression rather than independent authors. Future work will likely focus on improving retrieval efficiency and developing fine-grained detectors capable of attributing authorship at the token level, moving beyond binary "AI vs. Human" classifications.

Summary updated

Abstract

We introduce Frankentexts, a long-form narrative generation paradigm that treats a large language model (LLM) as a composer of existing texts rather than as an author. Given a writing prompt and thousands of randomly sampled human-written snippets, the model is asked to produce a narrative under the extreme constraint that most tokens (e.g., 90%) must be copied verbatim from the provided paragraphs. This task is effectively intractable for humans: selecting and ordering snippets yields a combinatorial search space that an LLM implicitly explores, before minimally editing and stitching together selected fragments into a coherent long-form story. Despite the extreme challenge, we observe through extensive automatic and human evaluation that Frankentexts significantly improve over vanilla LLM generations in terms of writing quality, diversity, and originality, while remaining coherent and relevant to the prompt. Furthermore, Frankentexts pose a fundamental challenge to detectors of AI-generated text: 72% of Frankentexts produced by our best Gemini 2.5 Pro configuration are misclassified as human-written by Pangram, a state-of-the-art detector. Human annotators praise Frankentexts for their inventive premises, vivid descriptions, and dry humor; on the other hand, they identify issues with abrupt tonal shifts and uneven grammar across segments, particularly in longer pieces. The emergence of high-quality Frankentexts raises serious questions about authorship and copyright: when humans provide the raw materials and LLMs orchestrate them into new narratives, who truly owns the result?
BibTeX
@inproceedings{pham2026frankentext,
  author = {Chau Minh Pham and Jenna Russell and Dzung Pham and Mohit Iyyer},
  title = {{Frankentext: Stitching random text fragments into long-form narratives}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.1457},
}
Figure