In 2024, the word “delves” turned up in medical research abstracts 28 times more often than its earlier trend predicted. Biologists didn’t suddenly fall in love with delving. A team at the University of Tübingen traced the spike to one thing: researchers running their writing through chatbots (Kobak et al., Science Advances, 2025).
That one word sums up the problem. Almost every AI chatbot can write. Far fewer write like a person, and readers are getting better at sensing the difference, even when they can’t explain it.
So if you’re choosing a chatbot for emails, posts, or reports, “how smart is it?” isn’t the first question. The better one is simpler. Will this still sound like me?
Why So Much AI Writing Sounds the Same
Chatbots are trained to pick a likely next word. Likely words pile up. You get the same polished cadence, the same three-part lists, the same tidy wrap-up line at the end of every paragraph.
The Tübingen team scanned more than 15 million PubMed abstracts. During Covid, the words that spiked were nouns tied to the news, like “lockdown.” In 2024, the jump was almost entirely style words, mostly verbs and adjectives such as “showcasing” and “underscores.” The topics hadn’t changed. The voice had.
Excess frequency ratio of selected words in PubMed abstracts. Source: Kobak et al., Science Advances, 2025.
It isn’t only vocabulary. In a Science Advances experiment, stories written with ChatGPT-generated ideas were rated more creative one by one, yet they came out about 10.7% more similar to each other than stories written without AI (Doshi and Hauser, 2024). Each writer improved a little. Together, they blurred.
What This Means for the People Reading Your Work
Pew Research Center’s 2025 survey of 5,023 U.S. adults found that 76% say it’s extremely or very important to know whether text, pictures, and video were made by AI or by people. Yet 53% aren’t confident they could tell. That’s a wary audience. They can’t always prove a draft came from a machine, but they’re paying attention, and one stock phrase can lose them.
This matters because writing is the main thing people bring to chatbots at work. OpenAI’s own usage study, published as an NBER working paper in 2025, found writing makes up about 40% of work-related ChatGPT messages, and roughly two-thirds of those ask the model to edit or improve text the user already wrote. Most people don’t want a ghostwriter. They want their own words, cleaned up, still sounding like them.
What to Look For in a Natural-Writing Chatbot
Leaderboards won’t help much here. They measure reasoning and coding. The test that counts is what happens to your draft.
It keeps your voice when it edits
Paste in a paragraph you wrote and ask for a tighter version. A good chatbot trims and fixes. A weak one swaps your plain words for grand ones and tacks on a summary sentence you never asked for. If the result reads like a press release, move on. Teams editing at scale have the same worry, and this guide to AI tools that cut editing time without losing brand voice shows how to lock key claims and names before any rewrite.
It takes follow-up instructions without starting over
Natural writing comes from revision. “Shorter.” “Less formal.” “Drop the second point.” The chatbot should remember the whole conversation and change only what you asked, not regenerate everything in its default style.
Its sentence rhythm varies on its own
Read the output aloud. Human writing has short bursts. Then a longer sentence that wanders for a moment before it lands somewhere useful. When every sentence is roughly the same length, your ear catches it long before any detector does.
It’s honest about sources
Easy to overlook, this one. A chatbot that sounds human but invents references is worse than one that sounds stiff, because the polish makes the fake look real.
It lets you check and fix in one place
Scanning a draft for robotic patterns and rewriting the flagged lines where you wrote them saves a lot of copying between tabs.
Red Flags That Should Make You Close the Tab
Most problems show up in the first five minutes of testing.
· Stock vocabulary. Watch for “delve,” “tapestry,” “seamless,” and “in today’s fast-paced world.” One is a habit. Five on a page is a pattern.
· Everything in threes. Three benefits, three tips, three adjectives in a row. Real writers stop at two, or ramble on to five.
· Confident fake citations. A Lancet audit of 2.5 million biomedical papers found the rate of fabricated references grew more than 12-fold from 2023 to early 2026, with the sharpest rise starting in mid-2024 as AI writing tools spread (Topaz et al., 2026). At the time of the audit, 98.4% of affected papers had seen no publisher action.
· Flattery and filler. “Great question!” up top, and a closing line that restates the paragraph you just read.
· A tone dial that doesn’t turn. If “more casual” gives you the same text plus two contractions, you’ve hit the model’s ceiling.
How to Test a Chatbot Before You Trust It
You don’t need a lab. You need 20 minutes and a piece of writing you know well.
1. Use your own sample. Pick a recent email or paragraph you wrote. Ask the chatbot to cut it by about 20%. Did it keep your phrasing, or replace it?
2. Run a revision chain. Ask for a draft, then give three follow-ups in a row: shorter, warmer, cut the last point. Check that each change sticks.
3. Ask for sources on something specific. Then open every link. One invented citation tells you plenty.
4. Scan the result. Run it through an AI detector to see which sentences read as machine-made. Treat the score as a pointer to weak spots, not a verdict. Detectors disagree with each other, and none can prove who wrote something.
5. Read it out loud. Still the best test there is.
If you’ve tried patching a general chatbot with clever prompts, know the limits. Phrasly tested two popular humanizer GPTs on one ChatGPT draft, and both kept the facts but changed so little that GPTZero still rated the rewrites fully AI. Prompts help. They rarely change the underlying style.
| Look for | Avoid |
| Keeps your wording when editing | Swaps plain words for grand ones |
| Remembers follow-up instructions | Regenerates everything from scratch |
| Mixed short and long sentences | Every sentence the same length |
| Real, checkable sources | Polished but invented references |
| Tone that actually shifts on request | Same voice no matter what you ask |
Where Phrasly Neo Fits In
If you want an AI chatbot that writes naturally from the first draft, Phrasly Neo is built around that goal. It’s Phrasly’s everyday chat model, tuned for human-sounding writing and conversation rather than leaderboard scores. You can draft an email, then nudge it (“shorter, and mention the Thursday deadline”), and it builds on the last reply instead of starting over.
A few practical details:
· Neo chat is free on every plan, with daily limits. Web search and image creation are paid features.
· It reads PDFs, scans, images, and audio, so your notes can become a first draft.
· Ask for sources and it searches for real ones, cites them, and builds a reference list in APA, MLA, or Chicago. Phrasly says it writes around a point when it can’t find support. Check every reference anyway.
· When a draft needs a heavier rewrite, Neo can call Phrasly Ultra, the specialist humanizer, inside the same chat.
One honest caveat, straight from Phrasly’s own FAQ: a Neo reply isn’t automatically run through the humanizer, and nothing guarantees a particular detector result. Treat it like a good writing partner. It gets you most of the way. Your edits finish the job.
Choose the Chatbot That Sounds Like You, Not Like Everyone Else
The “delves” spike is funny until you see what it shows. Huge numbers of people handed their sentences to the same few models and came out sounding alike. Readers noticed, even when they couldn’t name what felt off.
So judge a chatbot on your own writing, not on a polished demo. Make sure it keeps your voice, follows direction, and doesn’t make up sources. Keep an eye out for the stock phrases and the neat little triplets. Then do the final pass yourself. The goal was never to slip past a detector. It was to sound like a real person who knows what they’re talking about, and that part is still yours.