Why "Humanize My Writing" Tools Don't Work

In December 2024, Florida State University linguists Tom Juzek and Zina Ward sought to answer a common joke among AI output readers: why does ChatGPT use the word "delve" frequently? Rather than speculating, they developed a study to eliminate incorrect explanations systematically. First, they ruled out the possibility of these words being common in the training data.

Next, they dismissed the idea that it might be related to the model architecture or the mechanics of selecting the next word. After comparing a base version of Llama 2 with the same model fine-tuned using human feedback, the only remaining explanation was the fine-tuning stage itself. This revelation wasn't a fringe result. A separate team led by Dmitry Kobak at the University of Tübingen analyzed over 15 million PubMed abstracts published between 2010 and 2024, identifying that "delve" wasn't the only word that spiked in usage.

Words such as "meticulously" (137% increase year over year), "intricate" (117%), and "commendable" (83%) also saw significant rises. The FSU team's finding, that at least 13.5% of 2024 biomedical abstracts show signs of LLM involvement, is still unreplicated. However, it is already becoming outdated as people who frequently read AI text are adopting its vocabulary, including words like "delve," "boast," and "meticulous," into their own writing.

This shift is already bleeding into mainstream language use. The issue with many "humanize my writing" tools lies in their reliance on a list of words. After examining several GitHub repositories, the researcher found a spectrum of effort applied to the same underlying mechanism. One of the most popular tools, with over 36,000 stars, is a simple markdown file that protects ordinary formal vocabulary, avoids inventing facts during rewriting, and matches a user's writing style.

However, this doesn't change the fact that it's still checking a passage against roughly three dozen fixed patterns, which become outdated as models or detectors evolve. Another tool acknowledges previous false-positive research on AI detectors but still provides a single rule and claims its signals should be taken seriously. The most extreme tool instructs its model to flag text as machine-written if even one of the listed words appears.

These banned words include phrases like "robust," "scalable," "integrated," and "proactive," which frequently appear in ordinary technical writing. Despite the care taken by some tools, they all share the same vulnerability - they are cheap to build, fast to become outdated, and target the weakest part of the signal. While many of these tools aim to make writing appear human, peer-reviewed computational linguistics research suggests that structural analysis of sentences and stories is a more reliable indicator of AI-generated content.

A 2024 study out of the Universidade da Coruña compared six sets of LLM-generated news text against real human-written articles and found that models consistently produced sentences with similar lengths and structures, regardless of the specific model used. Similarly, a PNAS study discovered that LLM-generated narratives follow a smaller set of recurring plot patterns compared to human-written narratives, even across different prompts.

These findings highlight the limitations of attempting to humanize writing through word manipulation alone, as structural features provide a stronger signal of AI generation.