AI can write fast, but sounding human is harder

TL;DR: A new Claude Code plugin claims to “humanize” AI text, marking yet another attempt to get AI writing to sound less like AI. Why is that so hard to actually achieve? We break it down.

What happened: For years, the human volunteers of Wikipedia have been taking notes on how to spot a Trojan horse in their midst: AI writing. Some big tells, according to their detailed article laying out the signs, include overly generic phrasing—like “the move underscores a broader shift”—exaggerated claims of importance, an absence of concrete details and nuance, and a strangely promotional tone. AI loves words like “delve” and “pivotal.”

All that hard work is now being weaponized to make AI sound, well, less like AI. A new plugin for Claude Code uses the Wiki page as an instruction manual on what the chatbot should avoid, Ars Technica reports. Aptly named the Humanizer, the plugin claims to edit a piece of text to avoid the common AI hallmarks noted by Wikipedia.

Does it actually work?: In short, no. Before I installed the plugin, I told Claude to generate a roughly 50-word snippet on a nonfiction topic of its choice. It (just as many humans likely would) chose to write about itself:

Artificial intelligence continues reshaping industries worldwide. Companies are racing to integrate AI tools into everything from customer service to scientific research. Meanwhile, governments grapple with regulation—balancing innovation against concerns about job displacement, privacy, and misinformation. The technology's rapid advancement has sparked both excitement and anxiety about what comes next.

You can see the tells like a giant neon sign—a lot of vague, sweeping statements about how AI is changing everything without actually saying much.

After installing the plugin and running the above text through the Humanizer:

AI is changing how companies work. Businesses are adding it to customer service chatbots, research pipelines, and pretty much everything in between. Governments are still figuring out the rules—how do you encourage new tech while protecting jobs and stopping misuse? Nobody really knows where this is heading, and people are split between optimism and genuine worry.

The changes it flagged based on the Wiki guide include: using more concrete language, noting that “racing to integrate” sounded “promotional,” replacing em dashes (although it didn’t actually do this), dropping “vague filler” (like “rapid advancement”) and adding voice (“pretty much everything”).

But, in this human writer’s humble opinion, it still doesn’t sound natural. And the bad news for the Humanizer and its fans is that this excerpt also gets flagged as 100% AI-written in detectors like GPTZero.

Why it matters: There’s a growing market for tools like the Humanizer, including ones from Grammarly, Quillbot, Undetectable AI, and more. But, as the Humanizer plugin shows, these tools remain unreliable and have limited usefulness—after all, LLMs are already trained on human writing.

On the other side of the coin, AI detectors are popular too, and many academic institutions use them to sniff out LLM-generated writing submissions, with some colleges spending tens of thousands of dollars a year on faulty detectors. But testing shows that they sometimes miss AI text and misdiagnose human-written text. Wikipedia’s own section on signs of human writing is extremely short—one of the biggest ways to tell, it says, is whether the purported writer can explain their editorial choices.

The bigger picture: All this shows that while AI companies tout that their tools provide gains in efficiency and optimization, authenticity matters too. People may want to use AI to write, but they don’t want it to sound like the thing it actually is. The tension between those two goals—writing quickly using AI or writing like a human—remains hard to square.

And if AI chatbots are uniformly instructed to avoid certain words and sentence constructions, that could get taken to the other extreme and become a new tell of AI writing, with many humans adjusting the way they write in response (cough, the em dash, cough).

Researchers have found that humans are increasingly speaking and writing more like AI, too—common AI-style phrasings are coming out of the mouths of lawmakers and academics. The result is a strange ouroboros: Machines imitate humans, humans imitate machines, and everyone accuses everyone else of sounding like a bot. —WK