AplikaceAplikace
Nastavení

AI-Brown

AI-Brown is a generated, annotated, multi-genre corpus of English texts produced by large language models (LLMs).

Version v1 v2
Positions Number of positions (tokens) 27 661 454 53 862 480
Number of positions (excl. punctuation) 23 975 982 46 376 397
Number of word forms (excl. punctuation) 125 896 181 535
Number of lemmas (excl. punctuation) 110 835 172 410
Further information Number of subcorpora 32 49
Number of models 16 27
Publication year 2025 2026

Modeled on the BE21 Corpus1) — a modern implementation of the original Brown Corpus—AI-Brown was created to replicate its structure, genre diversity, and linguistic richness, enabling systematic comparisons between human and machine-generated English texts. The corpus comprises outputs from 13 frontier LLMs developed by OpenAI, Anthropic, Meta, Alphabet, and DeepSeek. Each model was prompted using the first 500 words of BE21 text samples, with the remaining portion reserved as human-authored reference material, ensuring genre-aligned and topically consistent comparisons. Like BE21, AI-Brown spans a wide range of contemporary English genres. All generated texts are tokenized, lemmatized, and annotated morphologically and syntactically using the Universal Dependencies framework, and are provided in both plain text and CoNLL-U formats. AI-Brown is a large-scale English LLM-generated corpus explicitly designed for cross-model and human-machine linguistic analysis.

Corpus preparation

The preprocessing pipeline for the original reference BE21 corpus included several steps to prepare the data for prompt-based generation. Clean texts and metadata were extracted from the verticals, and structural tags were aligned with the Czech corpus format to ensure cross-linguistic consistency.

Each BE21 text sample was split into two parts to support controlled generation:

  • Prompt portion: The first 500 words (including punctuation) served as generation prompts
  • Reference portion: The remaining text (approximately 1,500 words) provided human-authored comparison material

This segmentation ensured that models received sufficient context for meaningful text generation while preserving a substantial portion of reference text for evaluation. A 500-word prompt fits comfortably within the input limits of older models (e.g., davinci-002, which has a 2,049-token context window), using roughly 670 tokens.

To maintain comparability with the Czech AI corpus AI-Koditex and avoid over-representation, we selected written texts only, limiting the sample to one excerpt per source text. The final AI-Brown dataset contains 500 samples, matching the original structure of BE21.

Generating corpora

For each model, we generated two versions of each corpus: one using temperature 0 (deterministic generation) and one using temperature 1 (stochastic generation). However, we encountered mode collapse with the oldest model (davinci-002) at zero temperature, resulting in constant repetition of identical sentences.

For base models operating in completion mode (davinci-002, GPT-3.5-turbo, Meta-Llama-3.1-405B), we used only the first portion of each source text as input, allowing the models to function as traditional language models for text prediction. For instruction-tuned models, we employed minimal system prompts requesting long continuation of given text. Without such prompts, models' default helpful assistant persona emerged who typically attempted to analyze, summarize, or answer questions within the source text rather than continuing it. We used the following system prompt: Please continue the text in the same manner and style, ensuring it contains at least five thousand words. The text does not need to be factually correct, but please make sure it fits stylistically.

To ensure reproducibility, we used random seed 42 for all OpenAI API calls. Unfortunately, other providers do not offer comparable deterministic generation options. Llama generations used 16-bit floating-point quantization (the highest available quality).

API responses were preserved in their entirety, including token probabilities and alternative tokens when available, to enable future analysis of generation uncertainty and model confidence.

Post-processing

Texts that were too short were removed. For instruction-tuned models, the original phrases such as „I'd be happy to continue“ were also removed.

Annotation

We used Universal Dependencies for annotation, as UDPipe represents the state of the art for multi-level linguistic processing (including tokenization, lemmatization, syntax, and morphology). The resulting annotation is in the CoNLL-U format, which is a widely adopted standard compatible with most modern NLP tools.

Version 1

AI-Brown v1 was the first proper English llm-generated corpus available. Besides being searchable through KonText, the full texts and annotated corpus can be downloaded from LINDAT.

Version 2

AI-Brown v2 contains all subcorpora of the first version, complemented with new texts generated by new models, all of them annotated by the newest version of UDPipe. This version is also available from LINDAT.

How to cite AI-Brown

Milička, J. – Marklová, A. – Cvrček, V. (2026): AI Brown and AI Koditex: LLM-Generated Corpora Comparable to Traditional Corpora of English and Czech Texts. Language Resources and Evaluation 60/77. Arxiv preprint from 2025: https://arxiv.org/abs/2509.22996

Milička, J. – Marklová, A. – Cvrček, V.: AI-Brown, version 1, 1. 7. 2025. Department of Linguistics, Faculty of Arts, Charles University, Prague 2025. Available at WWW: www.korpus.cz

Milička, J. – Marklová, A. – Cvrček, V.: AI-Brown, version 2, 6. 6. 2026. Department of Linguistics, Faculty of Arts, Charles University, Prague 2026. Available at WWW: www.korpus.cz

1)
Baker, P. (2023) A year to remember? Introducing the BE21 corpus and exploring recent part of speech tag change in British English. International Journal of Corpus Linguistics.