---
title: 'AI-Brown: LLM-Generated English Corpus'
url: https://www.emergentmind.com/topics/ai-brown
type: topic
---

# AI-Brown: LLM-Generated English Corpus

Searching arXiv for the specified paper to ground the article in the published record.
AI-Brown is a multi-genre, LLM-generated English corpus built to mirror the modern BE21 Brown-family reference corpus, with 500 samples and 15–17 genres, and was designed to enable direct quantitative and qualitative comparisons between human-written texts and machine-generated continuations [2509.22996]. It forms part of a paired resource with AI Koditex for Czech, and its stated purposes are to provide a stable archive of frontier LLM behaviors, including models that are later modified or discontinued, and to serve as a resource for corpus-linguistic studies of syntax, morphology, lexicon, and style in AI-produced language. The corpus is annotated according to the Universal Dependencies standard and is freely distributed for research use.

## 1. Conceptual definition and relation to Brown-family corpora

AI-Brown is explicitly positioned as a generated corpus that replicates a reference human corpus rather than as an unconstrained collection of model outputs. The reference corpus is BE21 by Paul Baker, described as a modern version of the original Brown Corpus. In this design, AI-Brown inherits the Brown-family emphasis on balanced sampling across multiple genres and text types while replacing human-authored continuations with LLM-generated continuations [2509.22996].

The corpus is intended for comparison rather than substitution. Its function is to support analyses in which human-written texts in BE21 are compared directly with machine continuations under matched genre and topic conditions. This suggests a methodological commitment to comparability at the level of sampling frame, not merely at the level of aggregate corpus size. A plausible implication is that AI-Brown is useful where researchers require controlled contrasts between human and model-produced language across syntax, morphology, lexicon, and style.

The project also defines AI-Brown as an archive of model behavior. Because it includes outputs from models that may later be modified or discontinued, the corpus preserves historically situated LLM realizations. In that respect, it operates both as a corpus-linguistic resource and as a record of a specific generation environment.

## 2. Corpus construction and sampling protocol

The source dataset is BE21, described as 500 modern English texts in 15–17 genres such as News, Editorial, Fiction, and Academic. Each BE21 text was split so that the first 500 words served as the prompt and the remainder was reserved for human reference. Sampling mirrors BE21 proportions exactly: one continuation per text, preserving the original genre-by-text distribution, approximately 33k words per genre [2509.22996].

The paper states that this procedure ensures that each subcorpus has the same genre mix and comparable topical coverage as BE21. The preservation of genre proportion is formalized as

$$
\text{proportion}_{\text{genre}} = \frac{\#\text{texts in genre}}{500}.
$$

This construction is central to the corpus’s comparability claims. Because genre proportions are fixed by the reference corpus, differences observed in downstream analysis can be interpreted against a stable compositional baseline. This suggests that deviations in frequency distributions, lexical diversity, or stylistic dimensions are less likely to be artifacts of corpus imbalance.

The workflow is also deliberately text-aligned. There is one machine continuation for each BE21 text, which means that comparisons can be made at the level of genre, topic, or prompt-conditioned continuation behavior. A plausible implication is that AI-Brown supports both corpus-wide and item-level analysis.

## 3. Model inventory, generation regimes, and scale

AI-Brown was generated with thirteen “frontier” or widely used LLMs in completion and chat modes, across two sampling temperatures, $T=0$ and $T=1$. The models listed are:

| Provider | Models and modes |
|---|---|
| OpenAI | davinci-002 (GPT-3 base, completion-mode); gpt-3.5-turbo (completion-mode; chat-mode); gpt-3.5-turbo-instruct (completion-mode, instruction-tuned); gpt-4 (chat-mode); gpt-4-turbo (chat-mode); gpt-4o (multimodal chat-mode); gpt-4.5 (chat-mode) |
| Anthropic | claude-3-haiku; claude-3-opus; claude-3.5-sonnet |
| Meta | llama-3.1-405B (base); llama-3.1-405B-instruct |
| Alphabet/Google | gemini-1.5-pro-002; gemini-2.0-flash-exp; gemini-2.0-pro-exp |
| DeepSeek | deepseek-v3 |

For the English corpus, the subcorpus size is on average 864 000 tokens per model under UDPipe tokenization, with 27 000 000 tokens in total across 31 model×temperature combinations. Illustrative figures reported for chat-mode at $T=1$ are 688 280 tokens for GPT-4.5, 1 496 285 tokens for Claude-3-haiku, and 960 888 tokens for Llama-3.1-405B. The averages were computed over all accepted, non-refusal samples [2509.22996].

The corpus records refusal behavior rather than excluding it from methodological consideration. Refusal rates are tracked per subcorpus; an example given is that GPT-3.5-turbo chat at $T=1$ had approximately 7% refusals. This indicates that generation success is itself treated as a measurable property of the model-condition pair.

## 4. Annotation architecture and representational formats

AI-Brown is tagged according to the Universal Dependencies standard. Tokenization, sentence segmentation, UPOS/XPOS, lemmas, morphological features, and dependency relations were added by UDPipe using the `english-ewt-ud-2.15` model. The resulting outputs are distributed in standard CoNLL-U and in a verticalized format for Manatee/CWB, and are searchable via KonText/CQPweb [2509.22996].

The paper provides an example CoNLL-U annotation for the sentence “The cat sat on the mat.”:

```text
1   The the DET  DT  Definite=Def|PronType=Art                    2 det
2   cat cat NOUN NN  Number=Sing                                 3 nsubj
3   sat sit VERB VBD Mood=Ind|Tense=Past|VerbForm=Fin            0 root
4   on  on  ADP  IN  _                                           6 case
5   the the DET  DT  Definite=Def|PronType=Art                    6 det
6   mat mat NOUN NN  Number=Sing                                 3 obl
```

The same structure is rendered as a UD tree in LaTeX using `forest`:

```latex
\begin{forest}
  [sat\scriptstyle{root}
    [cat\scriptstyle{nsubj}
      [The\scriptstyle{det}]
    ]
    [mat\scriptstyle{obl}
      [on\scriptstyle{case}]
      [the\scriptstyle{det}]
    ]
  ]
\end{forest}
```

The annotation design places AI-Brown within standard corpus-processing pipelines used in dependency-based corpus linguistics. Because the corpus is lemmatized and morphosyntactically annotated, it can support studies of lexical choice, inflectional patterning, dependency structure, and register-sensitive grammatical variation without requiring reannotation.

## 5. Quantitative comparability and linguistic measurements

The corpus is designed for direct comparison with BE21 through normalized frequency and diversity measures. Token frequency distributions are normalized to instances per million, defined as

$$
\text{i.p.m.} = \left(\frac{f_n}{N_{\text{tokens}}}\right)\times 10^6.
$$

Lexical diversity is measured by MATTR with window size $w=100$:

$$
\text{MATTR} = \frac{1}{(W-w+1)} \sum_{i=1}^{W-w+1} \left(\frac{\text{types}_i}{w}\right).
$$

These measures are used alongside refusal tracking and stylistic comparison. The paper reports a marked pattern of formulaicity: the bigram “testament to” occurs 222 i.p.m. in AI-Brown versus 7 i.p.m. in BE21, and the 4-gram “a testament to the” reaches 163 i.p.m. in AI-Brown and 0 in BE21 [2509.22996].

Lexical diversity under $T=1$ exceeds $T=0$ for all models, and some Anthropic models approach human BE21 diversity at $T=1$. Stylistic benchmarks using Biber-style multidimensional analysis show that base models mimic register-specific styles better than instruction-tuned models, that models vary by genre, and that no single “AI style” dominates. The paper also reports that English continuations are stylistically closer to human texts than Czech ones.

These findings are analytically important because they simultaneously support and qualify the corpus’s comparability. On one hand, syntactic and stylistic behavior can approach human baselines under some conditions; on the other hand, repeated formulaic sequences and instruction-tuning effects remain detectable. This counters a common simplification that LLM output is either uniformly human-like or uniformly machine-like.

## 6. Empirical findings, interpretive significance, and limitations

The paper reports that syntactic complexity, measured by average dependency length and clause complexity, lies within $\pm 10\%$ of BE21 for high-capacity models such as GPT-4 and Claude-3-opus. It also reports that vocabulary richness correlates with sampling temperature and that GPT-4.5 at $T=1$ outperforms GPT-4 at $T=0$. In addition, instruction tuning propagates formulaic expressions across providers [2509.22996].

These results support a differentiated interpretation of generated language. The corpus does not support a view in which stylistic behavior collapses into a single homogeneous machine register; the paper explicitly states that no single “AI style” dominates. Instead, model family, tuning regime, genre, and temperature jointly affect output characteristics. A plausible implication is that comparative claims about “LLM language” require explicit conditioning on generation settings.

The paper also identifies failure modes and asymmetries. Some older or non-English-optimized models, including `davinci-002` and `Llama-3.1-405B` in Czech, suffer from looping or incoherence, and those outputs are stored in a “Failed” folder. Czech generation is described as weaker due to underrepresentation in training data. These observations delimit the generality of cross-linguistic conclusions and indicate that performance heterogeneity is part of the resource’s evidential value rather than a preprocessing nuisance.

## 7. Availability, licensing, and research uses

AI-Brown is distributed with distinct licensing regimes for raw and annotated material. Raw JSON and plain texts are released under CC BY 4.0, while annotated CoNLL-U and vertical formats are released under CC BY-NC-SA 4.0. The corpus is available from LINDAT/CLARIAH-CZ at `http://hdl.handle.net/11234/1-5993`, searchable in KonText at `https://www.korpus.cz/kontext/query?corpname=ai_brown_v1`, and accompanied by generation scripts, metadata, and visualization code on OSF at `https://osf.io/sfk6w/` [2509.22996].

The paper identifies several research uses: corpus-linguistic studies of AI-language morphology, syntax, discourse, and semantics; benchmarking LLM outputs against human reference for stylistic, lexical, or syntactic evaluation; training and evaluating human/AI text classifiers with multi-genre, annotated data; and calibration studies of model uncertainty using recorded token log-probs.

Future releases are stated to add GPT-4.1, GPT-5, Claude 4 Opus, new open-weight models, and extended languages. Users may also fork the corpus to add domain-specific or semantic annotations, including PropBank and frame semantics. By design, AI-Brown is both a mirror of human reference data and a living archive of LLM behavior, a formulation that captures its dual role as a controlled comparative corpus and an evolving observational record.

Source: https://www.emergentmind.com/topics/ai-brown