Why do LLMs sound so average? The training data is the culprit.
Large language models are trained to predict the most probable next word given a sequence of previous words, using enormous datasets scraped from the internet—books, articles, forums, and more. This statistical averaging means the model learns the most common patterns, not the idiosyncratic ones that make a human writer distinctive. A 2025 study analyzed lexical, grammatical, and rhetorical features in texts from Llama 3 and GPT-4o and found systematic differences between human and LLM writing styles; these differences persisted even as models got larger, and were actually more pronounced in instruction-tuned models (those fine-tuned to follow human preferences) than in base models [1]. In other words, the more the model is optimized to please a general audience, the more it sheds any hint of a unique voice.
Another study from 2023 directly measured content diversity in a controlled experiment where people wrote argumentative essays with and without model assistance. It found that using InstructGPT (a feedback-tuned model) led to a statistically significant reduction in diversity: essays from different authors became more similar to each other, and overall lexical and content diversity dropped [3]. The researchers traced this effect to the model itself contributing less diverse text, while the human-written portions remained unchanged. This confirms that the averaging effect comes from the model's output, not from users copying its style.
Not all LLMs are equally average—model size and training method matter.
The tendency to produce averaged-out writing varies by model. A 2025 study using authorship verification (a technique that compares writing styles) found that texts generated by GPT-2 had the highest stylistic similarity to human-written texts, while GPT-3 and ChatGPT outputs were very different from human writing and similar to each other [4]. This suggests that newer, larger models (like GPT-3 and ChatGPT) have become more homogenized—they converge on a narrower stylistic range. The same study found that LLaMA-generated texts fell somewhere in between, sharing some similarity with the original human snippet but also clustering with other LLaMA outputs [4].
A 2024 study on creative, witty texts (Reddit Showerthoughts) compared GPT-2, GPT-Neo, and GPT-3.5. Human evaluators rated the AI-generated texts slightly worse on creative quality, but could not reliably tell them apart from human-written ones [5]. This means the averaging effect isn't always obvious to readers—the texts can still pass as human, but they lack the spark of originality. The study also found that fine-tuned classifiers could detect AI-generated texts with high accuracy, confirming that subtle stylistic fingerprints remain even when humans are fooled.
Instruction tuning (making models more helpful) actually makes writing more uniform.
One of the clearest findings across these studies is that instruction-tuned models—those fine-tuned with human feedback to be more helpful, harmless, and honest—produce even more averaged-out writing than base models. The 2025 rhetorical-style study found that instruction-tuned Llama 3 and GPT-4o showed larger stylistic differences from humans than their base counterparts [1]. Similarly, the 2023 diversity experiment found that InstructGPT (a feedback-tuned model) reduced content diversity, while the base GPT-3 did not [3]. This suggests a trade-off: the very process that makes models better at following instructions and avoiding offensive content also squeezes out stylistic variation, pushing them toward a bland, consensus style.
This is not a minor effect—the 2023 study reported that InstructGPT increased the similarity between different authors' essays, meaning two people writing with the same model ended up sounding more alike [3]. For anyone using LLMs for creative writing, journalism, or any task where a distinctive voice matters, this is a real limitation.
About These Sources
This answer is built on 5 peer-reviewed studies — published from 2023 to 2025, 4 from 2024 or later, 1 in Q1 journals — selected as the most relevant from 5 studies that passed quality screening, drawn from 35 papers retrieved from a database of over 500 million.
Sources used in this answer
Do LLMs write like humans? Variation in grammatical and rhetorical styles
Using Douglas Biber's set of lexical, grammatical, and rhetorical features, this 2025 study found systematic differences between human and LLM writing styles (Llama 3 and GPT-4o) that persisted across model sizes and were larger for instruction-tuned models than base models.
Unveiling ChatGPT text using writing style
This 2024 study used stylometric features to detect ChatGPT-generated text, achieving 100% accuracy with an XGBoost classifier, showing that AI writing has distinct, detectable stylistic fingerprints.
Does Writing with Language Models Reduce Content Diversity?
In a 2023 controlled experiment, writing with InstructGPT (but not base GPT-3) led to a statistically significant reduction in content diversity—essays from different authors became more similar, and lexical diversity dropped—because the model contributed less diverse text.
Be Sure to Use the Same Writing Style: Applying Authorship Verification on Large-Language-Model-Generated Texts
Using authorship verification, this 2025 study found that GPT-2 texts were most stylistically similar to human writing, while GPT-3 and ChatGPT texts were very different from humans and similar to each other, indicating increasing homogenization in newer models.
Investigating Wit, Creativity, and Detectability of Large Language Models in Domain-Specific Writing Style Adaptation of Reddit's Showerthoughts
In a 2024 study of creative Showerthoughts, human evaluators rated GPT-2, GPT-Neo, and GPT-3.5 texts slightly worse on creative quality but could not reliably distinguish them from human-written texts, while fine-tuned classifiers could.