Artificial intelligence (AI) may not excel in every domain, but its capability to produce written content is impressive. For anyone who struggled with the verbal section of standardized tests or finds longer writing tasks daunting, the ability to craft emails, essays, or cover letters can seem almost enchanting. Generating text without the stress of confronting a blank page can be a real game-changer, which explains why approximately 20% of adults in the U.S. have utilized AI for writing purposes.

However, once your text is polished—after thorough revision and fact-checking, of course—there’s still a significant challenge ahead: AI detectors. These range from humans discerning the subtleties of AI-generated prose to various online platforms that assert they can determine whether a piece was penned by a human or an AI. The accuracy of these detectors is often debated, yet they are widely used, posing a concern for anyone attempting to submit AI-crafted content as their own original work.

This leads to the emergence of AI “humanizers,” designed to refine generated text into a more human-sounding narrative by eliminating robotic phrases and common AI patterns. This enticing concept allows users to produce an AI-generated essay, then refine it through the humanizer for a final product that resembles authentic human writing (presumably, yours). The question remains: Do these tools actually deliver?

The Experiment

An exploration was conducted to evaluate these humanizing tools. While this might not qualify as comprehensive research, it certainly provides valuable insights for those who often depend on AI for drafting emails, assignments, or heartfelt messages.

To kick things off, an essay on enhancing the human-like quality of AI-generated writing was produced using ChatGPT. It generated a coherent text within seconds. This piece wasn’t fact-checked or edited, focusing solely on the raw output for testing.

The essay underwent scrutiny with several AI detection tools to confirm its status as a typical example of AI writing. The findings revealed what was expected: QuillBot detected it as 94% AI, ZeroGPT marked it at 97%, and Copyleaks flagged it at a staggering 100% AI-generated. Unanimously, the AI detection community concurred: this text was unmistakably composed by ChatGPT.

Evaluating the Outcomes

The ambitious quest to see if AI humanizer tools could adjust these scores begot interesting results. As the demand for effective humanization tools surged, a variety of popular options were selected for testing.

Initially, to calibrate the findings, the original text was fed back into ChatGPT with a request for humanized rendering. Since all these tools operate on AI, it seemed logical to request ChatGPT to alter its original style.

Subsequently, the original text was processed through a selection of four notable humanizer platforms: Paraphraser.io, StealthWriter, Grammarly, and GPTHuman.

This yielded five “humanized” versions of the essay, which had previously been flagged as AI-generated by three detectors. The anticipation was whether their scores would improve. The conclusions, however, were largely disappointing, with only one platform exhibiting a hint of effectiveness:

• Paraphraser.io: This platform didn’t fare well, scoring 83% AI-generated on QuillBot, 100% on Copyleaks, and a suspiciously precise 99.94% on ZeroGPT.

• ChatGPT: As anticipated, this tool failed, achieving scores of 100% AI-generated on QuillBot and Copyleaks, while ZeroGPT rated it at 87.77%.

• Grammarly: This humanizer showed familiar results, with scores of 99%, 97.1%, and 99.97% across the platforms.

• GPTHuman: Displayed variable results; QuillBot indicated 0% AI, while ZeroGPT presented a 60.96% rating, yet Copyleaks remained adamant, scoring it 100% AI.

• StealthWriter: A standout among those tested, with ZeroGPT rating it at a skeptical 64.89% AI-generated, yet Copyleaks detected just 3%, and QuillBot fell for it completely with a 0% score.

Notably, StealthWriter offered repetitive processing capabilities, which enhanced its effectiveness. Its initial evaluation yielded a 65% humanization score. After multiple iterations, this increased significantly to 95%, maintaining that score thereafter.

All tools advocate for manual adjustments and reviews. The focus here was merely on whether they could evade AI detection, and the resolution is mostly negative, though StealthWriter might provide some assistance.

Lastly, recognize that a plethora of AI detection tools exists, leading to variability in results (even with StealthWriter). Without knowing the specific tool employed, an AI-generated text could still face setbacks under alternative detectors. Therefore, if there’s concern about detection, the most reliable strategy remains crafting the text personally or meticulously revising AI-sourced material.