I tested popular AI humanizers — and they made my writing much worse

In a recent experiment, the author tried various AI humanizers to determine if they could improve the readability of AI-generated text. The results, however, were far from satisfactory. The writing produced by these tools came across as unnatural and stilted, with bizarre synonyms replacing clear language and the rhythm becoming choppy and unnatural. In some instances, the tools even altered facts or removed crucial context.

AI detectors focus on patterns commonly found in machine-generated text, such as predictable word choices, repetitive sentence structures, and a consistent rhythm. Many humanizers seem to target these patterns by swapping common words with less predictable alternatives and rearranging clauses. Unfortunately, the outcomes were so poor that it's better to write the content yourself to avoid embarrassment.

Although these tools do make the writing statistically less predictable, that doesn't necessarily mean it's better writing. The results often resemble bad machine translations, making it difficult to read and understand.

One experiment involved using ChatGPT to write an essay about the domestication of dogs. The original text, which read, "Dogs were domesticated thousands of years before humans began keeping livestock or growing crops," was altered by Superhuman from Grammarly, a tool the author highly recommends for catching spelling and grammar errors.

The altered version replaced straightforward words with longer, strange alternatives, creating awkward phrasing that seemed to strip away any natural voice. Even the original AI draft sounded more human-like than the rewritten version.

Another issue with AI humanizers is their potential to change facts or remove important context. For instance, when using Rehumanize.io, the author found that "dogs came from an ancient wolf group centuries ago" replaced the accurate statement that dogs descended from an ancient population of wolves at least 15,000 years ago. This alteration made the explanation less natural and accurate.

Moreover, humanizers do not guarantee a successful bypass of AI detection. Detection tools evolve and may become better at recognizing the strange synonym swaps and mangled sentence patterns produced by these tools. In some cases, using a humanizer could even increase the likelihood of being flagged as AI-generated. Additionally, some tools, like Ace.ai, require downloading an entire browser and may produce results that still feel off.

Ultimately, while AI humanizers can make the writing statistically less predictable, the results are often worse than the original AI-generated text. It's better to avoid these tools altogether and write the content yourself to ensure accuracy and readability.