I’ve spent the last three years building the process we use at TJ Digital to write articles with AI, and well over 1,000 hours on our writing skill at this point. It’s hard to say how much of that was specifically about making AI sound less like AI, but that has been the most difficult challenge. Getting it to write for search engines was the easy part. We got pretty good at understanding what makes a page rank and what makes a page get cited, and it’s easy for AI to work those principles into the writing. Getting it to sound like a person has been much harder.
Early on, I aimed my fixes at the wording, with banned phrases, sentence rules, tone instructions and writing samples. That made sense to some degree, and it’s where most of the advice on this subject still goes. But the hard part of getting AI writing to sound human is getting it to have a point of view. AI plays it safe, rarely takes a side and almost never says anything you couldn’t find on ten other pages.
The best fix I’ve found is to give the AI a real person’s opinions and doubts to represent, and the easiest way to get those is to have that person talk about the topic for a few minutes. I saw it first in my own blog posts, and grading hundreds of A/B tests since then has only made me more sure of it.
I’m just one person, though, and I don’t want my view to dictate what’s true. So I had Claude spend four or five hours finding every study, test, tool and skill on the subject. A lot of very smart people have been working on this, and it was dumb of me not to look sooner.
This page has what held up and what didn’t, with links so you can check it. What to do and why is all in the main text and the charts, and the study details and caveats are in sections you can open when you want the proof. The longest lists are on a searchable page with every tool, repo, study, sign and claim we checked. All of this is about writing, since AI voices and speech are a different subject.
What Makes AI Writing Sound Human
The biggest difference I’ve found comes from a real person’s point of view: their opinions, the things they’re unsure about, where they disagree with the usual advice, and the examples only they would use. AI doesn’t come up with those by default, and no fix to the wording can add them.
I learned this from my own blog, and for a long time I took the wrong lesson from it. The blog posts on my site are all written by AI, but they start with a transcript from me. We take one of my videos, we take the transcript, and we turn that into a blog post. The post ends up maybe three times as long as the transcript, because the AI goes and finds supporting facts to back up what I’m saying, and it still sounds like I wrote it.
Using the transcripts was Claude’s idea in the first place. While I was working on one of my videos, it suggested using my video transcripts as examples, and I said, oh yeah, that’s great. It did seem to help a lot.
My assumption this whole time was that the AI could tell how I speak from the transcript and was copying my voice. I spent a lot of time trying to capture each client’s voice and apply it to all their writing, and it never worked very well. And I kept thinking, but my blog posts sound so good.
Then this October I learned that the AI wasn’t replicating my voice very much at all. It uses the words from my transcripts occasionally, but not throughout. This is how I described what it was picking up on a call that week:
“It’s my opinions. It’s my thoughts. It’s the fact that I hedge. It’s the fact that I’m going back and forth between two ideas. I have opinions that the AI would never have, and it’s able to then bake that into the whole content. So, how do you scale that?”
I thought my transcripts were making the writing sound human because the AI could match my wording, when really I was giving it a point of view it could represent.
You can see the difference in two versions of one paragraph, both written for this article, on whether AI humanizer tools work. The first has no point of view:
Written for this article, with no point of view: AI humanizer tools can be a helpful part of a content workflow. They may reduce some common AI patterns, although results tend to vary depending on the tool, the content and the audience. It’s worth testing a few options to see which works best for your needs.
The second is built from what I said about humanizers on a call in October:
Written for this article, from my own words: Humanizers do help, but they usually go wrong in one of two directions. Either they add personality and it comes out a little cringe, or too casual for a blog post, or they strip the personality out, and the post stops sounding like AI and starts sounding boring and bland. We’ve tried several and built our own, and I still haven’t found a rule-based system that makes it feel like there’s a soul behind the writing.
The second version has an opinion, a doubt and something I’ve actually tried. None of its sentences is fancier than the first version’s.
Talking seems to bring this out better than writing does. In six studies of people sharing opinions about products, restaurants and experiences, the ones who spoke into a microphone used more emotional language than the ones who wrote. Planning made the difference, because when speakers were told to think first, their language flattened to the level of writing.
That fits what I’ve seen. Written Q&As don’t work nearly as well for us as spoken transcripts, and I expect a ten-minute voice memo to work about as well as one of my video transcripts.
I should be upfront about the evidence, though. We found no published study comparing articles built from a person’s interview with articles written other ways. What I’m telling you comes from experience: well over 1,000 hours on our writing skill, and hundreds of A/B tests I’ve graded along the way, including about 50 in this last round. You can decide how much that’s worth to you. You can also judge it by how this article reads.
How to Give AI a Point of View
Record the person who knows the subject talking about it for a few minutes, without a script, and have the AI write from the transcript. If nobody is available to interview them, the AI can do the interviewing. If they won’t record anything at all, work from calls and videos they’ve already done, or take a position from experts the brand agrees with and let the brand change it.
A recording is our favorite way to start an article at TJ Digital. We get someone talking on the topic, and then the AI writes the brief, our plan for the article, from that person’s point of view and adds the research and facts around what they said. The first four steps below are how our article process works, and checking the certainty is the one I’d add when you do this yourself.
This chart shows the steps at a glance, and the next few sections go through them with the prompts to paste.
What to do when you can get the person who knows the subject to talk, even if that’s you
- Get them talkingA two-to-five-minute voice memo, a call, or the AI interviewing them one question at a time
- Clean it lightlyRemove filler and false starts; keep their sentences and meaning
- Pull their best linesOpinions, doubts, stories and phrasings, quoted exactly
- Draft around themThe AI orders, cuts repeats and adds facts, keeping their words where they fit
- Check the certaintyIf they said “I think”, the draft shouldn’t say “always”
Nobody will record anything?
Work from calls and videos they’ve already done, or take a position from experts the brand agrees with and let the brand change it. Don’t have the AI play the person.
Record the Person Talking for a Few Minutes
A voice memo is the easiest version. The person rambles into their phone for two to five minutes about the topic and gives their opinions as they go. When we can get time on a call, we sometimes get through about ten topics in one sitting.
Ask them to talk without preparing. Planning is what flattened people’s language in those speech studies, and written answers sent ahead give people all the time they want to plan.
Clean the transcript lightly by taking out the false starts and stumbles and leaving the person’s sentences and meaning alone. One oral-history editing guide goes further and cuts most qualifiers like “I think”. I’d keep those here, because the doubts are what you recorded the person for.
The recording gives you material no amount of research will. As I tell clients, “we get things from you that we wouldn’t get in our own research.” If you’re planning content at any volume, this is the piece of an AI content strategy I’d build around.
If you only try one thing on your next piece, ramble into your phone for two minutes before you open the chatbot.
What to Ask
Ask for the things AI won’t produce by default, which are what the person disagrees with, what they’re unsure about and what they’ve seen firsthand. We found no study of which questions draw out the most opinion, so treat these as a starting list:
- What do most people get wrong about this?
- Where do you disagree with the usual advice, and why?
- What are you still unsure about, or what have you changed your mind on?
- What’s a specific example or story from your own work?
- How would you explain this to a customer in a sentence or two?
Ask them to over-explain, too. Scribe Media, which ghostwrites books from recorded interviews, tells its authors to over-explain because cutting is easier than adding (Scribe Media).
Have the AI Interview You
If nobody is available to interview you, the AI can do it. Have it ask one question at a time and keep its own messages short, because you want to be the one doing most of the talking. Katie Parrott, a writer at Every, a publication that covers AI, starts her essays this way, with Claude interviewing her one question at a time (Every).
People seem to like being interviewed this way, too. Anthropic ran 1,250 interviews with professionals using its own AI interviewer, each about 10-15 minutes, and reports that 97.6% of participants rated their satisfaction at 5 or higher on a 7-point scale in a survey afterward (Anthropic). That’s the maker’s survey of its own tool, its page doesn’t say how many people answered, and the study didn’t write articles from those interviews.
You can paste this prompt into ChatGPT, Claude or Gemini.
Before you write anything, I want you to interview me. The piece is about [topic], and it's for [who will read it].
First, look up what's commonly said about this topic so you know the usual advice.
Then ask me one question at a time. Keep each question short, and wait for my answer before you ask the next one. Base each follow-up on what I just said. Over the interview, get me to tell you:
- what I think most people get wrong about this
- where I disagree with the usual advice, and why
- what I'm unsure about or have changed my mind on
- a specific example or story from my own work
- how I'd explain this to a customer in a sentence or two
If I make a factual claim, note it so we can check it later. Stop after about 10 questions, or when I say "done".
Then give me the full transcript, followed by a list of my exact quotes worth keeping. Don't write the piece yet.Turn the Transcript Into a Draft
Once you have a transcript, the drafting step is where the person’s view gets kept or lost. I’d do it in this order:
- Put the transcript at the top of the prompt, above your instructions. That’s where Anthropic recommends long documents go (Anthropic).
- Pull out the person’s best lines first, meaning the opinions, doubts, stories and phrasings you want to keep. Check each line against the transcript.
- Have the AI draft around those lines. Its job is to put the points in order, cut repeats, and add connecting sentences and facts, keeping the person’s own sentences where they fit.
- Read the draft for certainty the AI added. If the person said “I think this usually works”, the draft shouldn’t say “this always works”. Every’s Compound Writing plugin checks a draft against its source transcripts for this kind of drift (Every’s Compound Writing).
- Skip any humanizer pass over the person’s own words. Polishing them risks making them read like written text again, though that’s my inference and we found no test of it.
For steps 2 and 3, you can paste this after the transcript.
Above is a transcript of [name] talking about [topic]. I want a [blog post / email / social post] for [who will read it], written from their point of view.
First, list the 5 to 10 lines from the transcript that carry their opinions, doubts, stories and examples. Quote them exactly.
Then write the draft:
- Build it around those lines. Keep their sentences wherever they fit, cleaned of filler and false starts only.
- Keep their level of certainty. If they said "I think" or "usually", keep it, and don't add certainty they didn't have.
- Where they disagree with the usual advice, say so plainly.
- Don't add opinions, stories or quotes they didn't give. If the piece needs a fact they didn't mention, write [SOURCE NEEDED] where it goes.When Nobody Will Record Anything
Most people won’t record anything, which makes this the usual case.
“Our ideal situation is that the clients give us at least a voice memo, because the content turns out great when they do,” I said on a call this fall. “But I’m also thinking like, okay, our clients aren’t always going to do that. They’re usually not going to do that.”
Start with spoken material that already exists, such as sales and onboarding calls, webinars, podcasts and videos. In the tests I’ve graded, a client’s words from past calls helped most on articles where we had nothing else from them.
If there’s no recording at all, take a position anyway. Our new system has every article take one, either from the client’s own words or from experts the brand agrees with, and the plan is for the client to vet any opinion drawn from other experts. We’d always rather have the client’s point of view than another expert’s. But either way, the article “needs to have some kind of opinion that’s not super safe, just to sound human.”
Don’t ask the AI to play the person or to invent an expert. When Nielsen Norman Group reran three user studies with AI stand-ins, the AI users gave favorable, vague feedback, often treating every idea as a good one, and you’re looking for the opposite of that (Nielsen Norman Group).
Invented experts can also get you in trouble. Press Gazette named more than 50 apparently fake experts who featured more than 1,000 times in the British press (Press Gazette).
Emails and Social Posts
The same idea works at a smaller size. Before the AI writes an email or a post, say what you actually think into a voice memo, or type a few rough lines in your own words. Then have the AI shape those into the message, using the drafting prompt with the format changed.
Where to Put Transcripts and Samples in ChatGPT, Claude or Gemini
In all three apps, transcripts, writing samples and the person’s opinions go in a project or a skill, and short standing rules go in custom instructions. We found no test of which place produces the most human-sounding writing, and this covers what each feature holds as of Nov 1, 2026.
I wouldn’t start a new voice setup in a custom GPT or a Gemini Gem. Custom GPTs are scheduled to retire on Dec 11, 2026, later for some business workspaces, and Gems on adults’ personal accounts are removed on Dec 17, 2026. If you’re building skills for several brands, I wrote about using Claude skills for multiple clients.
See what each app holds
- In ChatGPT, Projects keep chats, files and instructions together, which makes them the place for samples and transcripts. Custom instructions hold 1,500 characters on Free and Go and 5,000 on paid plans, which is room for rules but not samples. For drafts such as social posts, OpenAI says your requested tone and context take precedence over the personality setting. Personal accounts also can’t make new custom GPTs.
- In Claude, a project’s knowledge is always loaded in that project’s chats, which makes it the place for transcripts and samples. Skills load only when Claude judges them relevant (Customize > Skills, with code execution turned on). Claude’s old Styles feature appears to be gone. Its help page was missing on Nov 1, and Claude’s personalization help now describes only instructions, projects and skills. If you had a custom style, look under Customize > Skills.
- In Gemini, skills (Customize > Skills > Create) hold instructions and up to 100 MB of files, and Gemini uses one only when its name and description match the task (Google). Gems on adults’ personal accounts become read-only on Nov 17, 2026, a month before they’re removed, and you should download their files before then. Work accounts follow in 2027.
Why AI Writing Sounds Like AI
Much of the AI sound traces to the training a model gets after it has read the web. In that second stage, called post-training, the model learns to follow requests and to give the answers people rate highly, and people tend to rate the familiar, predictable answer highly. Training on those ratings pushes a model toward its most typical answer, and that shapes what it chooses to say along with how it says it.
This is what I mean when I call AI safe. It doesn’t have controversial opinions or use quirky examples, and it doesn’t have a voice outside of what’s already accepted as known. As I put it in a message during the research: “It’s not that AI is a bad writer per se. It’s just that it doesn’t have much to say that’s of interest other than what it can verify to be true.”
I also think much of the issue is that AI doesn’t have a good theory of mind. It has a really hard time understanding what’s going to be obvious to the reader and what needs more explanation, and so it will often over-explain obvious concepts and then breeze right over very complex concepts with a single short sentence. I imagine we just need smarter AIs to fix that issue, because no amount of skills or guidance is going to fix it.
The style is also hard to shake. A program trained to recognize each chatbot’s writing could still name the model after another model reworded the text (the chatbot-fingerprint study), and much of the style is shared across companies, down to some of the same overused phrases (the professional-edits study). For anyone fixing a draft, this means rewording it or switching chatbots won’t remove the style.
The studies behind this
You can see the effect of post-training by comparing a model before and after it. A 2025 study that had models continue real people’s writing tested Meta’s Llama models, whose base versions (trained on web text, before any instruction tuning) are public, alongside OpenAI’s GPT-4o. The base Llama models came close to how people write on grammar features such as how often verbs get turned into nouns (“the implementation of” for “implementing”). Their instruction-tuned versions wrote less like people. GPT-4o, the chat model in the test, used “-ing” clauses (such as “leaning on his agility”) at 5.3 times the human rate.
A 2025 study of the ratings used to train models points at the cause. People rating answers prefer text that feels familiar and fluent, and the authors argue the push toward the typical answer would happen even with flawless training, because the preference is built into the ratings themselves. That training also makes answers less varied than training on examples alone. And professional writers produce far more text that can’t be pieced together from what’s already on the web than AI models do, which is a rough measure of originality.
Some of the sound comes from the web text itself. Most of the sentence patterns models repeat also appear in the text they first learned from (a 2024 study of sentence templates).
In the chatbot-fingerprint study, the program named which of five chatbots wrote a text 97% of the time, and still 91% of the time after another model reworded it. In the professional-edits study, writers who edited AI paragraphs found the same problems, down to the same overused phrases, in GPT-4o, Claude 3.5 Sonnet and Llama, though most of that study was fiction and the exact phrases may differ in articles.
Different models also give similar answers to the same open-ended question, and post-trained models’ story openings collapse to fewer types, even with training meant to keep variety. I’m not too worried about tells that show up across articles on a business blog, since it’s rare that someone reads more than one article from a client’s website.
Does AI Hedge Too Much or Too Little?
AI hedges less than people do, in the personal way. It uses fewer personal hedges like “I think”, fewer firm claims like “always” and fewer opinions, and in one small study, AI-written essays stayed neutral about 69% more often than essays people wrote alone. On impersonal softeners like “may” and “typically”, the evidence disagrees.
A personal hedge is a kind of commitment. “I think this works for most small businesses, but I’d skip it if you sell to large companies” puts a person behind a view, doubts included, and AI rarely writes sentences like that.
I got this wrong for a while. On a call in October I said AI doesn’t hedge and is always confident, and I repeated it in messages after the research came in, until I realized on Nov 1 what I’d said. AI does hedge. What it rarely does is hedge or commit from a view of its own, the way a person does.
My advice is not to ban hedges. AI already commits less than people do, and what my transcripts added was a person’s own doubts and firm views.
The hedging studies, and what we counted in our own drafts
We also counted these words in our own drafts and in older human-written posts. It’s a rough count, not a study.
How often each shows up, per 1,000 words
- Client drafts: AI first drafts of client articles, most without the client's call material
- My transcript drafts: AI first drafts of my own articles, written from my video transcripts
- Human posts: Company blog posts from four businesses, published before 2023
Personal hedges
"I think", "in my experience"
Firm words
"always", "definitely"
Softeners
"may", "typically"
My transcript drafts had about 5.5 times the personal hedges and 3.2 times the firm words of those client drafts. These are our own counts from one writing process, though. My drafts are first-person articles in my own voice and the client drafts aren’t, and the transcript may not be the only thing behind the difference.
The human posts look odd on personal hedges because company blogs rarely say “I”. It can’t settle that question, and the essay studies compared students’ and researchers’ writing, where people do say it.
A 2025 study comparing ChatGPT’s argumentative essays with British university students’ found ChatGPT used clearly fewer hedges, fewer words that make a claim firmer and fewer words that show the writer’s own judgment. A 2023 study of ChatGPT essays found fewer phrases like “I think” and “in my opinion”, even though teachers rated those essays higher. And ChatGPT-written linguistics research articles in a 2026 study never used “I” or “we”.
In the study of neutral essays, about 100 people wrote an essay on whether money leads to happiness, some on their own and some with an AI (gpt-4o-mini) writing it for them. The 69% is a relative increase from one topic and a small group. It matches how people describe AI writing, though. A long-time writer on the forum LessWrong complained that it never asserts anything without leaving room for the opposite (Richard Kennaway).
Whether AI overuses words like “may” and “typically” is still open. In the study that had models continue real people’s writing, GPT-4o used more downtoners than human writers did, words like “barely” and “nearly” that weaken a claim, though that’s a different group of words from “may”. Wikipedia’s volunteer editors, who keep a public guide to spotting AI writing, saw “perhaps” and “tends to” more often in human-written Wikipedia articles than in AI text (Signs of AI writing).
AI also overstates in a way no word count catches. Every’s writing plugin flags drafts that turn a partial test into a general finding (Every’s Compound Writing).
In our own counts, the 17 client drafts used less than half the human rate of firm words and about the human rate of softeners, which is plenty of “may” and little “always”. Two of those 17 still had existing client input from calls. Client transcripts also didn’t have the same effect mine did. Drafts given a client’s own call transcripts had 0.17 personal hedges per 1,000 words, against 0.32 for the drafts they were compared with, and the comparison can’t say why.
These numbers have limits. They come from one writing process with its own rules, the word lists can’t tell what a sentence means, and nothing here checks whether the gaps could be chance. The human posts had two personal hedges in about 31,000 words.
Which Fixes for AI Writing Actually Work?
Most of the fixes people try work on the wording. Those fixes are real, but they’re small. They remove one pattern at a time, and the models change their habits anyway.
This table has every method we looked at, with a verdict that grades the published evidence for it. Works means the evidence supports it, Partly means it helps with limits, Doesn’t means the evidence says no, and Untested means we found no test of it. Anything marked as ours is my own experience from grading these, not a published study.
Helps, with limits: start here8
| Method | Verdict | What the evidence says | Source |
|---|---|---|---|
| Record the person and write from the transcript | Partly | People use more emotional language speaking than writing; it’s what has worked best in my experience; we found no study of articles | Speaking-versus-writing study |
| A handful of writing examples | Partly | Five examples beat none on the four models tested both ways, but AI blog posts still passed as the author’s only 17-21% of the time | Writing-samples study |
| Past call transcripts as source material | Partly | In the tests I’ve graded, they helped, most of all on articles where we had nothing else from the client | My own A/B tests |
| Banned-word list | Partly | In one developer’s test, a ban cut the listed words’ uses by about a third, and one rewrite lost a needed hedge; AI word lists change with each model generation | Schmalbach’s test; Wikipedia’s editors |
| Blocking words while the model writes | Partly | Simple word-fragment bans stop working at around 2,000 patterns; the Antislop sampler handles more but needs a model you run yourself; one developer’s ban worked on an open model reached through a hosted service; Claude’s API has no setting for it | Antislop paper; Schmalbach’s test |
| Start the draft with the person’s own first words | Partly | Raised the share of blog drafts one detector called human from 9% to 22%; we found no reader test | Writing-samples study |
| Humanizer rewrite | Partly | In one user’s test, won 16 of 16 with AI judges while detection barely moved, and judges noticed a more uniform rhythm afterward | blader/humanizer test |
| AI edits its own draft | Partly | Sometimes swaps one cliché for another and handles specific detail worst; models told to fix grammar only changed essays’ meaning | Professional-edits study; Neutral-essays study |
Doesn’t help10
| Method | Verdict | What the evidence says | Source |
|---|---|---|---|
| Tell it to “sound human”, “write casually” or “try harder” | Doesn't | Instruction-tuned models (GPT-4o, GPT-4o mini, Llama 3) told to match a person’s style kept a noun-heavy style; in one user’s test, “try harder” fooled 0 of 15 AI judges (AI models asked to rate text) | Model-continuation study; blader/humanizer test |
| A rule set telling the model to write more human | Doesn't | Fooled 0 of 17 AI judges in one user’s test; in our experience, piles of rigid rules made client drafts robotic | blader/humanizer test |
| More examples (up to 10) | Doesn't | Going from 2 to 10 changed the paper’s four measures very little; we found no test above 10 | Writing-samples study |
| Same-topic examples | Doesn't | Lowered how often the imitation was matched to its author (roughly halved for blogs) | Writing-samples study |
| Many excerpts plus a style description | Doesn't | Expert readers preferred human writers’ imitations, and a detector flagged 97% (fiction, older models, one preprint) | Author-imitation study |
| Rewrite with a second AI | Doesn't | The first model is still identifiable after another model rewords its text | Chatbot-fingerprint study |
| AI judge picks the best draft | Doesn't | Three AI models matched professional copy-editors at about chance on yes-or-no “slop” labels; in another study, a panel of open AI judges reversed 55% of verdicts when the order changed | Copy-editor study; Judge-panel study |
| AI role-plays the expert | Doesn't | Long excerpts of AI-imitated casual phone calls were judged human 29% of the time, against 64% for real ones; AI stand-ins in user studies gave favorable, vague feedback | Phone-call study; Nielsen Norman Group |
| Higher effort setting | Doesn't | On ToneBench, a writing benchmark that calls its results directional, Claude Opus 5.5 gained about 1.3 points from medium to extra high at about three times the cost, less than the score’s normal run-to-run variation; I haven’t seen a big difference in my own tests | ToneBench |
| Higher temperature | Doesn't | Added little originality and more incoherence in a 2024 story study; Claude Opus 5.5 and Sonnet 5.5 reject the setting | Temperature study; Anthropic |
Untested8
| Method | Verdict | What the evidence says | Source |
|---|---|---|---|
| Persona prompt | Untested | Doubled how often people took GPT-4.5 for a human in five-minute chats; we found no test on articles | Turing-test study |
| Write the prompt in the style you want back | Untested | Anthropic says the formatting of a prompt can steer the formatting of the reply; we found no test for article voice | Anthropic |
| Require a position in every piece | Untested | AI-written essays stayed neutral about 69% more often; in my experience it helps, though it changed along with other things | Neutral-essays study |
| Good-and-bad example pairs | Untested | Beat plain examples on 2024 tasks; we found no test for voice | Example-pairs study |
| Several drafts, a person chooses | Untested | The drafts have to be made to differ, since post-trained models’ story openings collapse to fewer types | Story-openings study |
| Learn from the person’s edits | Untested | Cut editing effort when GPT-4 played the users, on summaries and emails (2024) | Learning-from-edits study |
| AI interviews you first | Untested | In Anthropic’s own survey after its AI interviews of about 10-15 minutes, 97.6% of participants rated their satisfaction 5 or higher out of 7; the page doesn’t say how many answered, and no articles were rated | Anthropic |
| Written Q&A | Untested | Writing gives people time to plan, which flattened their language in the speech study | Speaking-versus-writing study |
Works, but most people can’t get it2
| Method | Verdict | What the evidence says | Source |
|---|---|---|---|
| A scoring model trained on expert edits picks the best | Works | Nine professional writers ranked its picks best; one preprint, with its research models and data released by the authors | Expert-edits study |
| Fine-tuning on the person’s writing | Worksmostly out of reach | In one preprint, expert readers preferred GPT-4o fine-tuned on an author’s books (30 authors, excerpts up to 450 words of literary writing) at a median $81 per author in computing costs; closed to new OpenAI customers, and the only Claude model Amazon Bedrock lists for tuning, Claude 3 Haiku, has reached end of life | Author-imitation study; OpenAI; Amazon Bedrock’s tuning list; Amazon Bedrock |
The method I recommend, recording a person and writing from the transcript, only gets a “Partly” there, because we found no study that tests it on articles.
Telling ChatGPT to Sound More Human
Instructions like “sound more human”, “write conversationally” or “try harder” barely change anything, because they leave the model’s trained habits in place. In a 2025 study that had models continue real people’s writing, chat-trained models (GPT-4o, GPT-4o mini and Llama 3) told to write in the same style, tone and word choice as the original still wrote a noun-heavy style with more “-ing” clauses than people use. A 2026 study of 11 models found the main differences between AI and human text held up even under prompts written to get human-like text.
I’ve seen the same thing myself. As I told my team in October: “My experience is adjectives don’t solve anything. Saying be more conversational, saying, you know, write more clearly, saying make it easier to read, like none of that does anything.”
Giving the model a persona is an open question. In a test where people chatted for five minutes with a person and an AI at once and then guessed which was human, a humanlike persona prompt raised how often they picked GPT-4.5 from 36% to 73%. That was live chat, though, and we found no test of personas on writing.
Banned Words and Writing Rules
A banned-word list cuts down the words on it, and not much else. The model seems to swap in new habits, the list goes out of date with the next model, and a ban can even change what a sentence claims.
Some popular lists also ban the wrong things. The stop-slop skill says to cut all adverbs, and a prompt from a 2025 blog post bans “might” and “tends to”. Those bans push writing away from how people write.
We went through all of this with our own rules. Claude doesn’t realize what’s unnatural about its own writing, and we had to spell it out, so at one point we had about a hundred writing rules. As I said on a call in October, it’ll “apply those rules very diligently, but it still is missing that personality or that soul that comes from human writing.”
More rules made it worse. When a client said they hated a sentence, Claude tended to overcorrect with a rule like “never use a transition”. We had a few client accounts that were really hard cases, where we could never get natural-sounding drafts, and when we looked deeper, each had about 80 rules like that in its guidelines. Following that many rigid rules made the writing robotic again.
Even one rule can bend the writing. I updated our guidance on opening paragraphs again and again, and the drafts still contorted into unnatural structures to satisfy the rule about answering the question in the opening paragraph.
I used to recommend banned AI words, tone settings and writing samples, and I’ve changed my mind on most of it. This is what I’d do now:
- Find the patterns that actually show up in your drafts. mshumer/unslop has the model write about 50 samples in your field and lists what it repeats, and the table under “Signs of AI Writing” works as a checklist.
- Keep a short list of those patterns, say what to write in place of each, and check the list again when the model changes.
- Edit or remove an old rule before adding a new one.
- Put the rest of the effort into what the draft says.
When we cut our rules down and required every article to take a position, the drafts started sounding more like a real person at the brand. Several things changed at once, so the rule cut can’t take the credit by itself.
If you want to look through the published word lists and style guides yourself, this table has the ones we found, with how much evidence is behind each. Start with Wikipedia’s.
| Name | What it does | What the evidence says | Cost |
|---|---|---|---|
| Start here | |||
| Wikipedia: Signs of AI writing Top pick | Volunteer editors’ guide to AI tells by model era, with sections on signs of human writing and signs that don’t work | Why start here: it’s the most complete list, it’s sorted by model era, and it’s kept current. A long guide, edited almost daily; observations with footnotes, no accuracy numbers for its signs, though it quotes one 2025 preprint’s figures for expert readers (90% caught, 3.3% misidentified) | Free |
| Anthropic’s prompting guides | Say what you want, name specific patterns and match the prompt’s style to the output; the Fable 5.1 guide defines mannered prose, and the Opus 5.5 guide says to name patterns to avoid (in its web-design section) | Why start here: it comes from the people who make Claude, and it says what to write, not only what to avoid. The company’s own guidance, with no numbers | Free |
| Gwern, “Better LLM writing” | Have the model suggest chatbot-style edits, then do the opposite; generate many options and pick | Why start here: its idea of generating many options and picking one is the most useful habit on this list. One practitioner; older models | Free |
| Just to know about | |||
| WikiProject AI Cleanup guide | How Wikipedia’s editors find and clean up AI text | Community guidance; rates Pangram among the most reliable detectors despite its misses | Free |
| EQ-Bench Slop Score (code) | Scores text against 1,648 words, 430 three-word phrases and “not x but y” patterns | No reader test; its own page says it isn’t an AI detector | Free |
| Pangram’s signs of AI writing | A detector company’s rates for nine surface signs | The company’s own analysis; it says these signs aren’t what its detector looks at | Free |
| tropes.fyi | 49 AI writing tropes in six groups, tagged as rising or fading | Shows no source data or dates | Free |
| jola.dev on Claude’s phrases | Claude’s pet phrases plus a hook that swaps them on screen | The hook changes only what you see; Claude still writes them | Free |
| sabrina.dev’s prompt and inaiwetrust’s list | Prompts with long banned lists (65 entries on sabrina.dev, 108 on inaiwetrust), including “can”, “may”, “might” and “tends to” | No tests; they ban hedges that human writing uses, and sabrina.dev’s bans semicolons and formatting, then suggests a semicolon and asks for bullets | Free |
| Zack Proser’s scrub pass | Run the anti-slop pass last, and have a different model family do the rewrite than the one that drafted | One practitioner, no reader test; he warns that a draft that says nothing can still pass a scanner | Free |
| Simon Willison on AI writing | Lets AI proofread, but won’t let it speak for him in opinions or first person | One practitioner’s rule | Free |
| WRITING.md | Rules that change by type of writing, with examples and a self-check | None found | Free (MIT) |
The tests on banned words, and on blocking words while the model writes
A ban changing what a sentence claims turned up in a test by Vincent Schmalbach, a developer. He ran eight AI articles through DeepSeek V4 Flash, an open model, on a hosted service, with a list of AI words made less likely as it wrote, and the ban cut those words’ uses by about a third (his write-up). One of the eight rewrites also dropped the hedge “appears”, which turned a careful observation into a flat claim. He also tried a second method on a different model, GPT-5.6 Sol, pasting the list into the prompt. That cut the listed words from 11 to 1 across five outputs, but two of the five dropped or changed information.
The habits come back in new forms. When AI models tried to fix AI prose the way professional writers had, they sometimes replaced one cliché with another (the professional-edits study). Wikipedia’s editors keep separate AI word lists for each model generation, and they now describe a reversed form of the old “it’s not X, it’s Y” habit, “Y rather than X”, as common in recent output. We found no study that measures a banned habit coming back in new wording, though the sources we found that track these patterns over time do show them shifting from one model generation to the next.
On the lists that ban hedges, the essay studies found AI already commits less than people do, and Wikipedia’s editors saw words like “very”, “perhaps” and “tends to” more often in human-written Wikipedia articles than in AI text.
Whether you can block words while the model writes depends on the model. Claude’s API has no setting for it (Schmalbach’s write-up), and OpenRouter, a service that lists each model’s settings, showed none for GPT-6 or Gemini 4 Argon in late October. On a model you run yourself, a tool called the Antislop sampler watches the output and backs up whenever a banned word, phrase or pattern appears, and it handles more than 8,000 of them (the Antislop paper). Simple word-fragment bans stop working well at around 2,000, and either way only the wording changes.
Anthropic’s advice points toward saying what you want. Its prompting guide says Claude responds well to clear, explicit instructions and suggests telling it what to do in place of what not to do, such as asking for smoothly flowing paragraphs in place of a ban on formatting symbols (Anthropic). Anthropic’s guide for Claude Fable 5.1 also says an instruction that defines “mannered prose” (metaphor and flourish where a plain statement would do) and asks Claude to remove it helps with that model (Anthropic). When Anthropic released a prompt for this, I tried it, and it helps a little.
AI Humanizer Tools
Humanizers mostly clean up word-level tells. You already saw my view of them in the before-and-after, where they either add personality that comes out a little cringe or strip it out and leave the post boring and bland.
The only judge or reader test from outside a humanizer’s makers that we found is one user’s blind test of a popular skill called blader/humanizer. Its rewrite beat the raw AI output on quality as AI judges saw it, but detection barely moved. Afterward the judges noticed a uniform rhythm more often, which the tester read as a new tell. No people rated the text.
Commercial apps such as Undetectable AI and StealthGPT market themselves for getting past detectors, and I wouldn’t make that a goal. We found no independent test of commercial humanizers on writing quality, and the leaderboards we found that rank them aren’t independent either.
This table lists the humanizer apps and writing apps we found, with what the evidence says about each. Tavi’s pick is Spiral, because it works from your own writing samples and doesn’t sell itself on getting past detectors. The vendors’ claims are their own.
| Name | What it does | What the evidence says | Cost |
|---|---|---|---|
| Start here | |||
| Spiral (from Every) Top pick | A writing app that picks your relevant samples for each new draft | Why start here: it works from your own samples and isn’t built around getting past detectors. The company’s own eval, with an AI judge trying to spot the draft among the person’s real samples: Spiral’s drafts blended in 87% of the time | From $15/month |
| Grammarly AI Humanizer | Rewrites in a chosen voice; says it isn’t meant to get past detectors | Why start here: there’s a free version, and it says it isn’t meant to get past detectors. Passed Pangram 0 of 30 times on one vendor-run leaderboard | Free basic version |
| Just to know about | |||
| Undetectable AI | Rewrites text to pass detectors and offers a money-back guarantee if output is flagged | One study found under 4% still flagged after it; leaderboards run by vendors disagree widely | From $5/month, billed yearly |
| WriteHuman | Rewrites text; says its model is updated for GPTZero and Pangram | Only leaderboards it runs or a rival runs | $12-$39/month yearly, $20-$59 monthly |
| StealthGPT | A rewriter built to get past detectors | After it, Pangram stayed accurate and GPTZero missed about half (University of Chicago test) | $25-$150/month yearly, $30-$180 monthly |
| QuillBot AI Humanizer | A rewriter with Basic humanizing on its free tier and Advanced humanizing on Premium | No independent quality test | Free tier; paid plans |
| ChatGPT’s AI humanizer page | Shortens, clarifies and rewords inside ChatGPT | None | Free with ChatGPT |
| Humanizer GPTs in the GPT Store | Creators’ own humanizer prompts | None; custom GPTs are scheduled to retire Dec 11, 2026, later for some business workspaces | Depends on your plan |
| Leaderboard humanizers (Phrasly, HumanizeMyPaper, HIX Bypass, Rephrasy and others) | Rewriters built to get past detectors | Ranked only by HumanizerBench (run by WriteHuman), HumanizerEval (data hosted by StealthGPT) and a Hugging Face page with affiliate links | Varies |
| Brand-voice features in Copy.ai, HubSpot and Hypotenuse | Match a voice from writing samples (Copy.ai recommends 3,000+ words if possible; HubSpot asks for 2 or 3 samples, of at least 500 words for blogs and pages or 150 for email, texts and social); Hypotenuse says it trains a model on how you write | No evidence the sample thresholds were tested | HubSpot: Professional and Enterprise plans; the other two pages name no plan |
The humanizer tests and leaderboards
A user ran the blader/humanizer test and posted the results on the skill’s public discussion page on GitHub (issue #229). This skill is the most popular humanizer there, with about 63,000 stars (GitHub’s version of a like). The rewrite won 16 of 16 quality comparisons, and detection went from 100% to about 98%. Afterward, the judges mentioned dashes and quotes far less, and mentioned a uniform rhythm and quotable closing lines more.
The tester concluded that prose with the same good shape everywhere is itself a sign of AI. It’s a small test, with AI judges from one model family, 77 of 144 planned comparisons finished and the tester’s own checker as part of the setup. The skill’s description page now says beating detectors isn’t a goal.
Results for the commercial apps conflict across studies that used different humanizers, detectors and kinds of text. After Undetectable AI, fewer than 4% of AI rewrites were still flagged by either of the two detectors in a Notre Dame paper. After StealthGPT, the detector Pangram still caught almost everything while GPTZero, another detector, missed about half, in a University of Chicago working paper.
A 2026 Belgian study used a test set built from student theses, with AI-written sections added and then rewritten by a GPT-4o prompt, not a commercial humanizer. Pangram located the AI content correctly in 37 of the 40 humanized papers, and GPTZero in 1.
As for the leaderboards, HumanizerBench is run by WriteHuman, a humanizer company. HumanizerEval names no operator, but its data sits under a GitHub organization called StealthGPT-Labs, and a ranking on Hugging Face, a site for sharing AI models, has affiliate links to its top three tools.
The boards also disagree with each other. Phrasly passed Pangram 4% of the time on one, 83% on another and 60 of 60 times on the third. Detector companies now sell humanizer detection too, and Pangram claims its version 4 catches 98.83% of output from 13 humanizers.
One skill maker’s own test points back to the person’s input. The authors of work-humanizer found their skill’s output read as human to Pangram on all four test texts when it had the writer’s own notes, while rewrites without notes scored 2-4% human. It’s the makers’ test on four texts, but what moved the result was input from the person.
Writing Examples
Two to five examples help about as much as ten, and even then they leave a big gap. This chart comes from a 2025 study of real writers. It shows how often an authorship check accepted AI blog posts as the author’s own, with no examples, with five of the author’s posts, and with examples picked at random against examples on the same topic.
How often an authorship check accepted AI-written blog posts as the work of the person they imitated
No examples against five of the person’s posts
- No examples8-10%
- Five of their posts17-21%
Five posts picked at random against five on the same topic
- Picked at random19%
- Same topic as the new post10%
On real people’s blog posts, the same check is right about 91% of the time.
Same-topic examples did worse. The model seems to pick up the topic and miss the style, which matches what I’d seen.
“If you say write something that sounds like this paragraph, it will write something that sounds like that paragraph,” I told my team. “The problem is it’ll take things from it that you don’t want to take from it.” The first time we gave it a set of examples that were all paragraphs, it stopped using bullets and tables. As I put it later, if none of the examples have bullet points, “it’s going to be like, oh, they hate bullet points.”
Most of our clients don’t have good samples anyway. As I said in a team meeting, even when a client has blog posts, “they’re usually like AI written or they’re written by another agency.”
In late October I still thought examples were the only thing that really worked. Then the research showed me that 10 examples were no better than 2, and that my transcripts had been working because they gave the AI my opinions.
If you use examples, this is how I’d do it:
- Use two to five. Anthropic suggests three to five that are relevant and varied (Anthropic).
- Make them do different jobs, such as an opening, an explanation and an opinion.
- Keep them off the new piece’s topic.
- Wrap each example in tags, such as
<example>and</example>, so the model knows where it begins and ends. - Tell the model what to take from them, such as how long the sentences run or how directly opinions get stated.
- Check the draft for wording copied straight from them. Every’s writing plugin flags this (Every’s Compound Writing).
If your goal is matching your own voice, I wrote a separate guide on how to make AI writing sound like you.
The research on writing examples
The chart comes from a 2025 study of more than 400 real writers. Five samples of someone’s writing beat none on all four models tested both ways, while going from 2 samples to 10 changed the paper’s four measures very little. With five examples, a program built to judge who wrote a text accepted only 17-21% of AI-written blog posts as the author’s, and on real blog posts the same check is right about 91% of the time. Blogs and forum posts were the hardest kinds of writing to imitate, and the models often fell back to an average, generic tone.
Same-topic examples cut those authorship scores roughly in half for blogs. Giving the model the first words of a real text helped with one detector. When the model continued from the first 50 words of the author’s real post, GPTZero called 22% of the blog outputs human, up from 9%, and no person judged those outputs.
More material doesn’t close the gap. In a 2026 fiction study, a preprint, GPT-4o, Claude 3.5 Sonnet and Gemini 1.5 Pro got 20 excerpts of an award-winning author plus a written description of the style, and professional writers imitated the same authors. Readers with graduate degrees in creative writing strongly preferred the human imitations, and Pangram flagged 97% of the AI text. Training the model on each author’s complete books reversed that result, which I cover under fine-tuning.
Samples from a company website may not be fully human, either. By September 2024, up to about a quarter of the text in corporate press releases was estimated to have been substantially modified by AI (a study of AI in corporate writing).
Editing an AI Draft
I wouldn’t count on AI editing to put a point of view into a draft. In one study of creative-writing paragraphs, models copying professional writers’ edits did worst at adding specific detail, had mixed results cutting unnecessary explanation, and sometimes replaced one cliché with another.
AI editing can also change what a piece says. In one study of AI-edited essays, three AI models told to fix grammar only changed the essays’ meaning and often their conclusions. Rewriting with a second AI runs into the same problem as switching chatbots, because the first model stays identifiable after the rewording.
Hand edits have limits too. When people edited AI drafts of their own personal writing, the results stayed closer to AI text than to their own writing (a 2026 post-editing study). And you can’t hand the final choice to an AI judge, meaning an AI model asked to rate drafts, because three AI models matched professional copy-editors at about the level of chance on what counted as “slop”, meaning generic, low-value AI text (the copy-editor study).
Having a person choose among several different drafts is promising, though we found no test of it. The drafts have to differ, and chat training makes answers less varied. One fix from the research is to ask for several options along with how likely the model thinks each is. In practice: “Give me five different openings for this piece, each with the probability you’d write it, and include a couple of unlikely ones.”
The editing studies
Tools built on the second-AI idea include nobuzz, which has Gemini rewrite Claude’s replies, vomit, which rewrites through a model running on your own computer and warns that it can make things up, and SlopMonster. Their pages show no test with readers.
In the post-editing study, 81 people edited AI drafts, and their edits varied widely. They removed about a quarter of the em dashes, rarely removed stock words and used about five times as many contractions as the drafts. The edited pieces also ended up far more alike than the same people’s unassisted writing.
In the copy-editor study, the three AI models flagged 3-8% of items where the editors flagged about a third. The one approach that picked well was custom-built. Researchers had an AI suggest 20 possible edits for each draft and kept the one a scoring model, trained on professional writers’ edits, rated highest (the expert-edits study). Nine professional writers ranked those picks above a random edit and above the first draft. The authors released their research models and data, and we found no ready-made product that does this.
Post-trained models’ story openings collapse to fewer types, even with training meant to keep the answers varied (a 2026 workshop study of AI story openings). Asking for several options with probabilities made answers 1.6 to 2.1 times more varied on creative writing tasks such as poems, stories and jokes (the ratings study).
Learning from your edits has early support. A 2024 method that wrote down a user’s preferences from their edits to AI drafts cut later editing effort, though GPT-4 played the users (the learning-from-edits study). writing-style-skill tries the same idea as a Claude skill, with no test we found.
Fine-Tuning on Your Own Writing
Fine-tuning means training a model further on your own writing. It’s the only method we found that beat expert human writers in a blind test, but that was one preprint on fiction, and most people can no longer fine-tune ChatGPT or Claude. We found no such test on blogs, business writing or brand voice.
It suits someone with a large archive of their own writing, technical help and a reason to keep one voice across a lot of content. For most people, I’d start with a transcript in the prompt.
What fine-tuning costs and where you can still do it
In the fiction study, professional writers and AI models each imitated award-winning authors. After GPT-4o was trained on the complete works of 30 of those authors, readers with graduate degrees in creative writing went from strongly preferring the human imitations to preferring the AI’s, on style and on quality (the author-imitation study). Pangram flagged 97% of the prompted imitations and 3% of the fine-tuned ones. The median cost was about $81 per author in computing and fees, not counting the human work of editing. We found no published before-and-after results from an agency or brand.
Access has mostly closed. OpenAI stopped new organizations from fine-tuning on May 7, 2026, and existing customers can’t start new jobs after Jan 6, 2027 (OpenAI). For Claude, Amazon Bedrock’s fine-tuning list names only Claude 3 Haiku, in one region (us-west-2), where AWS’s lifecycle page gives end of life on Sep 10, 2026.
You still have options outside those two. Google’s Gemini Enterprise Agent Platform, formerly Vertex AI, offers tuning for several Gemini Flash models and Gemini 2.5 Pro, though not Gemini 4 Argon (Google Cloud). Open models can be tuned on Together, which lists Qwen, Llama and Gemma, on Fireworks, which names Qwen, DeepSeek, Kimi and gpt-oss, or on Thinking Machines’ Tinker, which lists Qwen, GPT-OSS, Kimi and others. Hugging Face’s Claude Code training skill runs the training on Hugging Face’s cloud computers.
The training is cheap, and the data is the expensive part. Each training example pairs a short instruction with a passage of the person’s real writing. As rough arithmetic, 1,000 examples of about 5,000 tokens each (a token is roughly three-quarters of a word), trained in three passes, would cost about $16 on Together for a mid-sized model or $6 on Tinker for a small one. The 1,000 is only for the arithmetic. A planning note written by Nicolas Deville’s AI assistant for his knowledge base expects 200 to 500 reviewed examples and calls producing them the real cost, while the training takes an afternoon (Nicolas Deville).
If you try it, a few findings help:
- A simple tone change took about 100 examples in one small study, while the literary study used whole books.
- Train on voice and keep facts in the prompt, because training a model on facts it doesn’t already know makes it invent more (a 2024 study of fine-tuning on new facts).
- LoRA, a cheaper method that trains a small add-on in place of the whole model, matched full fine-tuning on small and medium sets of examples in research from Thinking Machines, a company that sells it.
- Check your archive for AI-written text first. Deville’s archive had 99 AI-written notes among its 1,284.
- Expect new quirks. A developer who trained on his YouTube transcripts got answers that sounded like him, along with em dashes he never uses and a generic voice on unfamiliar questions (Zen van Riel).
Which AI Model Sounds Most Human?
We found no public test where people rate current models on blog, business or brand-voice articles, so I can’t name a winner. Claude Opus 5.5 comes out on top of the two closest tests, ToneBench, which scores YouTube scripts written in one creator’s voice, and Arena’s creative-writing board, where people vote between two anonymous answers.
I wouldn’t lean on either one much. On ToneBench, Opus 5.5 is also one of the AI judges, and its lead on the score is about the size of a model’s normal variation from one run to the next. On Arena, it’s in a statistical tie with Gemini 4 Argon and Gemini 4 Carbon.
When I compared a Gemini 4 Argon draft with a Claude Opus draft, the Argon draft was about as good, and I think we’ll have to see how good Carbon is at writing.
We write ours with Claude Opus 5.5, and I wouldn’t switch models just to sound more human. Switching chatbots doesn’t remove the style, and the top models are all pretty close anyway.
I kind of assumed models hadn’t made much progress in writing, but seeing old articles written by AI changed my mind. I do think newer models are getting better at writing, and I think we’re also getting better at recognizing AI content at the same time. As models get smarter, they’ll generally get better at writing.
Every model change also resets some of this work. “Anytime a new model comes out, we have to kind of figure out how to get it to write like a normal person again,” as I said on a call. My guess is that in the next year or so we’ll get models that are really good at writing. It just hasn’t been a priority for the companies, who are focused on code because that’s where the money is.
If you want to check the boards yourself, this table shows the ones worth knowing, with what each actually measures and how much to trust it.
| Benchmark | What it measures | Leaders (date) | Limits |
|---|---|---|---|
| ToneBench | YouTube scripts written in one creator’s voice, scored by four AI judges against the creator’s own scripts | Claude Opus 5.5 85.93 Claude Sonnet 5.5 84.60 Claude Fable 5.1 84.53 GPT-6 Sol 83.52 (8th, the first non-Claude entry) All at max effort, Oct 26, 2026 | One creator and 10 tasks Opus 5.5 is one of the judges Scores vary 1.1 to 1.8 points between runs, about the size of the gaps at the top No human ratings shown The board calls its results directional |
| Arena creative writing | People vote between two anonymous replies to their own short creative prompts | Opus 5.5 1520 Gemini 4 Argon 1517 Gemini 4 Carbon 1515, on far fewer votes A statistical tie, Oct 27, 2026 | No target voice; short prompts; GPT-6 models outside the top 25 |
| EQ-Bench Creative Writing v3 | Short fiction scored by AI judges, plus a score for stock AI phrases | GPT-6 Astra, Kimi K3.1 Pro and Fable 5.1 lead Opus 5 has the fewest stock phrases (6.59) Opus 5.5 is 8th (10.43) Oct 28, 2026 | Fiction; Claude and GLM judges; the phrase score counts overused words and phrases |
| EQ-Bench Longform | Stories of about 8,000 words, scored by one Claude judge | Opus 5 86.3 Fable 5.1 85.3 Opus 5.5 84.6 Grok 4.7 has the fewest stock phrases Oct 28, 2026 | Fiction; one judge; the closest board to article length |
| Lech Mazur’s writing benchmark | 600-800-word stories that must include ten required elements, compared by AI judges | Fable 5.1 3.745 Opus 5.5 3.693 Opus 5 3.685 Oct 23, 2026 | Measures story construction around required elements; nothing on it measures how human a story sounds |
| SlopBench | Repeated openers, paragraph rhythm and stock phrasing in email, social, essay and workplace tasks | Kimi K2.6 has the least Mistral Large the most Sep 2026 | Only the two ends of the ranking hold up; unclear whether the newest models were tested |
| Hemingway-bench | Blind comparisons by professional writers (Surge AI’s post) | Claude Fable 5 1110 Gemini 3.7 Flash 1105 Claude Fable 5.1 1095 Nov 1, 2026 (the board shows no date) | Run by Surge AI, which sells human ratings; Opus 5.5 not listed |
More on the leaderboards, and what a writer at Every found
On ToneBench, Opus 5.5 scored 85.93 against 84.60 for the next model, and a draft’s score typically varies by about 1.44 points. The board also gives each model an Elo rating, built from head-to-head comparisons, and Opus 5.5’s range there doesn’t overlap the next model’s. The board itself calls its results directional and not statistically definitive.
The boards that use AI judges on fiction, EQ-Bench and Lech Mazur’s, rank Opus 5.5 lower. Gemini 4 Argon, tied for first on Arena, ranks 35th of 84 models on ToneBench’s voice matching. Price barely tracks quality either. Across ToneBench’s top 25 entries, cost per script varies about 250 times while scores differ by about 4 points. In a 2026 study of 11 older models, from GPT-2 to GPT-4, the choice of model shaped style more than the way text was sampled from it, with some exceptions.
The most useful human opinion I found came from Katie Parrott, a staff writer at Every, who had Opus 5.5 write versions of the opening of one of her own essays (Every). She called it the most readable prose she’d seen from Anthropic or OpenAI, said it often buries the point, and found it didn’t feel clearly human. Its versions ran 37 to 39 sentences against her 21.
Anthropic positions Claude Fable 5.1, its more expensive model, for hard reasoning problems and long tasks the model works through by itself, and tells most people to begin with Opus 5.5 (Anthropic).
Models sold for writing, such as Sudowrite Muse, NovelAI’s storytelling models and Writer’s Palmyra Creative, mostly target fiction and creative work, and we found no published blind test for any of them. Writer lists Palmyra Creative as deprecated as of Jul 13, 2026 (Writer).
Do Settings Like Effort and Temperature Help?
Not much, and Claude doesn’t let you change most of them. Effort, the setting for how much the model reasons before it writes, is the one Claude lets you raise. I’m not seeing a big difference from extra-high effort in the tests I’ve graded.
We still run our articles at extra high, because they need to run autonomously, and even a slight edge for medium wouldn’t change that. I explain the trade-off in which Claude effort level to use.
Temperature, the setting for how random the word choice is, adds little. In a 2024 story study, higher temperature was only weakly linked to originality and moderately linked to incoherence (the temperature study). Claude’s newest models don’t accept it anyway. Opus 5.5, Sonnet 5.5 and Fable 5.1 return an error if you set it, and they also reject prefill, an older trick where you write the first words of the model’s reply yourself (Anthropic, Sonnet 5.5 docs, Fable 5.1 docs).
Base models, the versions before chat training, have grammar closer to human, but we found no test of them with readers and they’re hard to direct. You’d also have trouble getting one. OpenRouter listed no base models among the 452 it carried on Nov 1, and we didn’t find one offered in the ChatGPT, Claude or Gemini apps.
This table shows what the evidence says about each setting and whether you can use it with ChatGPT, Claude or Gemini models.
Whether you can use each with Claude, ChatGPT and Gemini models, as of late October 2026
| Setting | What the evidence says | Claude | ChatGPT | Gemini |
|---|---|---|---|---|
| Higher effort | Opus 5.5 on ToneBench: 84.49 at medium, 85.83 at extra high (about 3 times the cost), 85.93 at max (about 8 times); I haven’t seen a big difference in my own tests | Yes, low to max (Opus 5.5 defaults to medium) | A reasoning setting on GPT-6 | Thinking set to low, medium or high on Gemini 4 Argon |
| Temperature | Weakly linked to originality, moderately to incoherence; min-p, a popular setting that trims unlikely words, showed no reliable benefit when its original data was reanalyzed (reanalysis) | No, it returns an error | Not listed for GPT-6 on OpenRouter | Yes, on Gemini 4 Argon |
| Base model | Grammar closer to human; two detectors often read its text as human (base-model study); we found no reader test; hard to direct; you’d run an open one yourself, such as Qwen3 Base or Llama 3.1 405B Base | None found | None found | None found |
| Blocking phrases while it writes | The Antislop sampler handles 8,000+ words, phrases and patterns; we found no reader test; possible on some open models | No | Not listed for GPT-6 on OpenRouter | Not listed for Gemini 4 Argon on OpenRouter |
| A model built for creative writing | We found no blind test showing these sound more human; most target fiction or creative work | Anthropic positions Fable 5.1 for reasoning work | Shown in March 2025, with no release date given (TechCrunch) | None found |
Signs of AI Writing
To check a draft, start with the stance and content rows at the top of the table. Those signs are about what the writing says, like whether it takes a side and whether anything in it is new, and no word list or humanizer fixes them.
You know AI writing when you read it, but it’s hard to articulate how you know it’s AI, and learning what actually makes it sound that way was the fascinating part of this research for me. Words like “delve” used to show up all the time, and em dashes are easy to remove and aren’t nearly as common anymore. AI is also very noun-heavy, which I hadn’t really thought about before. With Claude, the term people use is mannered prose. It tends to combine two different concepts into one sentence, if it’s going to list things it lists exactly three, and it likes to use synonyms for common words.
The word-level signs get the most attention and change the fastest. In the table, Current means recent studies or editors still see the sign, Fading means it belonged mostly to older models, Myth means the evidence doesn’t support it, and Unproven means we found no independent count of it.
“Our drafts” means TJ Digital’s AI first drafts, compared with 27 company blog posts that four businesses published before 2023. Phrase counts come from 76 drafts and rates like contractions from a larger set of 121 AI articles, and each row says which. One writing process produced all of them, and they show how that one process writes.
Stance6
| Sign | Example | What the evidence says | Status | Source |
|---|---|---|---|---|
| Few personal hedges | “I think”, “in my experience” | ChatGPT essays had fewer hedges overall (2025) and fewer phrases like “I think” (2023) than students’ essays, and ChatGPT linguistics articles had no “I” or “we”; older models | Current | Essay study, 2023 essay study, linguistics-articles study |
| Few firm claims | “always”, “definitely”, “the only” | Fewer in ChatGPT essays; less than half the human rate in our 17 client drafts, most without the client’s call material | Current | Essay study |
| No side taken | “There are pros and cons to both.” | When gpt-4o-mini wrote the essay, about 69% more essays stayed neutral than when people wrote alone (one small study) | Current | Neutral-essays study |
| Mild, upbeat feeling | positive tone, no strong emotion | 2023-era models wrote news with more joy and less fear than people did; training made one model’s fiction less intense | Current | AI-news study, Narrative-flattening study |
| Dated disclaimers | “It’s important to note that…” | Wikipedia’s editors date it to 2023-24; none of our 76 drafts used it | Fading | Wikipedia’s editors |
| Heavy softening | “may”, “typically”, “depends on” | Our 17 client drafts, most without the client’s call material, matched the human rate, while GPT-4o used more downtoners (words like “barely” and “nearly”) than people; the evidence points both ways, so softening alone tells you little | Myth | Model-continuation study |
Content8
| Sign | Example | What the evidence says | Status | Source |
|---|---|---|---|---|
| Generic detail | the first example anyone would think of | When AI models edited AI writing, adding specific detail was the fix they handled worst; models imitating writers often fell back to an average tone | Current | Professional-edits study, Imitation study |
| Clichés | “a sense of”, “the weight of”, “unspoken” | 17% of professional writers’ fixes; GPT-4o, Claude 3.5 Sonnet and Llama all overused the same ones in fiction, and AI edits swapped clichés for new ones | Current | Professional-edits study |
| Off-point or padded content | sections that drift from the reader’s question | Whether the text stayed on point and how much it actually said best predicted professional copy-editors’ “slop” labels (slop meaning generic, low-value AI text) | Current | Copy-editor study |
| Nothing new | points every page on the topic makes | Originality came up in 24% of all the explanations five frequent AI users gave for their calls | Current | Five-reader study |
| Invented stories, quotes and experts | a made-up customer story; a Seneca quote | GPT-3.5 and GPT-4 added an anecdote and a Seneca quote when asked to refine essays; many, not all, of 500+ stories built on one company’s press releases featuring fake and misleading comment have been deleted by the publishers | Current | Essay-refining study, Press Gazette |
| Inflated importance | “plays a pivotal role”, “a testament to” | Wikipedia’s editors list puffery and promotional tone; we found no count | Current | Wikipedia’s editors |
| Sameness across pieces | many articles with the same outline and opening | Different models give similar answers to open questions, and post-trained models’ story openings collapse to fewer types, even with training meant to keep variety | Current | Similar-answers study, Opening-types paper |
| Stock names and titles | “Emily”, “Sarah”, “Dr.” | In 63-70% of GPT-4o and Claude 3.5 Sonnet articles in 2024; we found no test since | Fading | Five-reader study |
Sentence and paragraph structure7
| Sign | Example | What the evidence says | Status | Source |
|---|---|---|---|---|
| Over-explaining and burying the point | explaining what the reader already knows | 18% of professional writers’ fixes to AI paragraphs; a staff writer at Every, a publication that covers AI, found Claude Opus 5.5 often buries the point | Current | Every, Professional-edits study |
| Even, polished rhythm | every paragraph the same length and cadence | AI models asked to judge text mentioned uniform rhythm more after a humanizer rewrite (60 mentions before, 101 after) | Current | blader/humanizer test |
| Formulaic endings | “In conclusion”, a reflective closing line | People who use AI often flagged formulaic conclusions; quotable closing lines rose after a humanizer rewrite | Current | Five-reader study, blader/humanizer test |
| Same shape in every section | answer, reason, example, takeaway, every time | Wikipedia’s editors and Every’s writing plugin flag it; we found no count | Current | Every’s Compound Writing |
| Repeated sentence templates | the same pattern of word types again and again | Models repeat grammar patterns more than people (2024), but in a study of professional copy-editors, repetition wasn’t a clear predictor of their “slop” labels | Current | Templates study, Copy-editor study |
| Sentences all about the same length | most sentences 10 to 30 words | 2023-era models varied sentence length less than human news writers; our 121 articles vary as much as the human posts | Fading | AI-news study |
| Perfect grammar, bland prose, transition words | flawless grammar; “Furthermore” | Wikipedia’s editors list these as signs that don’t work | Myth | Wikipedia’s editors |
Phrases and sentence patterns12
| Sign | Example | What the evidence says | Status | Source |
|---|---|---|---|---|
| “-ing” clauses | “Bryan, leaning on his agility, dances around the ring” | GPT-4o used them at 5.3 times the human rate, wherever they fell in a sentence (2024); in our 121 articles, a narrower count of “-ing” phrases after a comma was lower than in the human posts | Current | Model-continuation study |
| Noun-heavy grammar | “the implementation of” for “implementing” | GPT-4o turned verbs into nouns at about twice the human rate | Current | Model-continuation study |
| “It’s not X, it’s Y” | a sentence that denies one thing to assert another | Nearly tripled on the web from 2023 to 2026 but still rare (about 2 uses per 10,000 words); in 3% of our 76 drafts and 4% of human posts | Current | Pew Research Center |
| Reversed contrast | “Y rather than X” | Wikipedia’s editors call it common in recent output, with no source; “instead of” was in 38% of our 76 drafts and 11% of human posts | Current | Wikipedia’s editors |
| Fancy verbs for “is” and “has” | “serves as”, “boasts”, “features” | Wikipedia’s editors list it, and list plain verbs (“wrote”, “used”, “tried”) as signs of human writing | Current | Wikipedia’s editors |
| Few wordy connectors | rarely uses “in order to” or “the fact that” | Wikipedia’s editors saw these more in human-written Wikipedia articles than in AI text; “in order to” was in 1% of our 76 drafts and 33% of human posts | Current | Wikipedia’s editors |
| Announcing the point | “Here’s what that means”, “The real question is” | Listed in catalogs of AI habits, with no study we found; “that’s why / where / what / how” was in 58% of our 76 drafts and 7% of human posts | Current | tropes.fyi |
| Vague attributions | “Experts argue” with no expert named | Wikipedia’s editors list attributions like this as a sign | Current | Wikipedia’s editors |
| Chatbot leftovers | “I hope this helps!” | Wikipedia’s editors list them; older leftovers like notes about the model’s knowledge cutoff are now rare | Current | Wikipedia’s editors |
| Lists of three | “fast, simple and affordable” | The only count we found comes from Pangram, a detector company: 19 per 10,000 words in AI text against 5 in human text, by its own analysis | Unproven | Pangram |
| Missing contractions | “do not”, “it is” | People added contractions when editing AI drafts, but our 121 articles, written with instructions to match each brand’s voice, used twice the human rate | Myth | Post-editing study |
| Passive voice | “was designed to” (with no one doing the designing) | GPT-4o used this kind of passive at about half the human rate | Myth | Model-continuation study |
Words6
| Sign | Example | What the evidence says | Status | Source |
|---|---|---|---|---|
| The newest favorites | “matters”, “practical”, “clearer”, “quietly” | On Wikipedia’s mid-2026 list, which cites a single blog review as its source; “matters” was in 53% of our 76 drafts and none of the human posts | Current | Wikipedia’s editors |
| Each chatbot’s own words | ChatGPT: “certainly”, “overall”; Claude: “here”, “according to”, “based on” | Words that set one chatbot apart from the others, in 2024-25 versions | Current | Chatbot-fingerprint study |
| Claude’s pet phrases | “load-bearing”, “honest take” | Users report them; one user count found “load-bearing” rose from about 0.5 to 12 uses per 10,000 words between Claude Opus 4.6 and Opus 4.7 | Current | jola.dev, user count |
| “Delve” and its friends | “delve”, “showcasing”, “underscores” | Jumped in research abstracts after ChatGPT; “delve” fell once it was publicized in 2024; none of our 76 drafts used it | Fading | ChatGPT-words study, biomedical-abstracts study |
| Older models’ favorite words | “tapestry”, “intricate”, “palpable” (GPT-4o); “emphasizing”, “highlighting” (GPT-5) | GPT-4o used the literary ones 95 to 162 times the human rate in a 2024 test with fiction in the mix; Wikipedia’s editors date the GPT-5 words to mid-2025 through mid-2026 | Fading | Model-continuation study |
| Buzzwords | “leverage”, “robust”, “seamless”, “in today’s” | “Leverage” and “robust” are on Wikipedia’s watch lists; all four were in 0-13% of our 76 drafts, about as often as pre-2023 company blogs used them | Fading | Wikipedia’s editors |
Punctuation and formatting7
| Sign | Example | What the evidence says | Status | Source |
|---|---|---|---|---|
| Only commas and periods | many short sentences, little other punctuation | Wikipedia’s editors call it a new sign now that em dashes are trained out; one line cites a single blog review | Current | Wikipedia’s editors |
| Formatting where it doesn’t belong | bold and headings in an email or forum post | Pangram’s count: formatting symbols (bold, headers, code marks) in plain-text settings, 90 times per 10,000 words in AI text against 8 in human text, and bullet lists 28 against 3; formatting alone identified which chatbot wrote a text 73% of the time | Current | Chatbot-fingerprint study, Pangram |
| A bold label on every list item | “Speed: …”, “Cost: …” | Wikipedia’s editors list it; we found no count | Current | Wikipedia’s editors |
| Heavy bold | bold phrases in every paragraph | Wikipedia’s style notes say some newer models or apps are instructed to avoid overusing bold | Current | Wikipedia’s editors |
| Heading habits | Title Case Headings | Wikipedia’s editors say chatbots strongly tend to use title case in headings | Current | Wikipedia’s editors |
| Emoji, curly quotes, stray code | “utm_source=” left in a link | Wikipedia’s editors list them; curly quotes alone don’t prove AI, and “utm_source=” in a link is a ChatGPT leftover; Pangram counts emoji at 0.2 against 0.1 per 10,000 words | Current | Wikipedia’s editors, Wikipedia’s style notes, Pangram |
| Em dash overuse | asides set off by long dashes | Wikipedia’s editors list it as a sign of older models (2022 to September 2026); an Arena analysis reported in the tech press found Opus 5.5 used about 95% fewer than Opus 5 (0.8 against 15.2 per 1,000 words) | Fading | Wikipedia’s editors, BleepingComputer |
We also counted words and sentence patterns in our AI first drafts and in those human posts, and this table shows the patterns that set our drafts apart. The published lists of AI tells we checked mostly don’t name them (tropes.fyi’s “comma-clipped trailing phrase” comes closest to the “, so…” ending), and you may want to check your own drafts for them. Keep in mind that we found these phrases by searching the same drafts for anything that stood out, which makes the gaps look bigger than a fresh test would.
Share of drafts or posts that use each pattern, except where a row gives a rate
| Pattern | Example | Our AI drafts | Human posts |
|---|---|---|---|
| “So the”, “so a”, “so you” or “so it” in a sentence | “…, so you save a step.” | 93% | 26% |
| Colons in mid-sentence | “The answer is simple: …” | 96% | 52% |
| Questions of any kind (every question mark) | “So how does that work?” | 0.55 per 1,000 words (34 client drafts) | 2.87 per 1,000 words |
| “That’s why / where / what / how” | “That’s why the order matters.” | 58% | 7% |
| Recurring connectives | “on its own”, “whatever”, “each one”, “starts with”, “which means”, “the kind of”, “carries” | 17% to 49% | 0% to 7% |
Why word lists go out of date, and more on our counts
“Delve” faded soon after it was publicized as a ChatGPT word in early 2024 (a study of ChatGPT-favored words in research abstracts). Wikipedia’s volunteer editors, who keep a long, continuously edited guide to AI writing (Signs of AI writing), now list em-dash overuse as a sign of older models, and they say newer models are made to suppress em dashes, in a passage tagged as needing a source.
When one sign goes, another tends to show up. Those editors say AI text now uses commas and periods more than people do, with little other punctuation, and they call that a sign of its own. That passage is also tagged as needing a source, and another line on the page cites a single blog review for the same habit. Word lists also fit only the writing they came from. The best-known ones were built from biomedical abstracts and fiction, and we found none built from blog or marketing writing.
In our pattern counts, the draft figures come from 76 AI first drafts unless a row says otherwise. Against the published tells, our drafts didn’t show GPT-4o’s “-ing” habit, the “it’s not X, it’s Y” habit, a narrow range of sentence lengths or missing contractions. They used “matters” far more often than the human posts and “in order to” far less, as Wikipedia’s editors would predict.
The low question rate may come from our own writing rules. The human posts are older company blogs, many of them weak. The counts describe what’s typical of our drafts and say nothing about what’s good.
Can People Tell When Something Was Written by AI?
People who use AI for writing often can spot it, especially as a group, and people who rarely use AI are close to guessing. This chart comes from one 2025 study of nonfiction articles (the five-reader study). It compares five frequent AI users voting together, across 300 articles, with four people who rarely use AI, averaged per person, on 60 of those articles.
Nonfiction articles, half written by AI
Five frequent AI users, voting together
- 99.3%of the AI articles caught
- 0%of the human articles wrongly flagged as AI
Four people who rarely use AI, averaged
- 56.7%of the AI articles caught
- 51.7%of the human articles wrongly flagged as AI
A single frequent user is much less reliable than the group. Alone, each of the five caught between 59% and 97% of the AI articles, and one of them caught none of the humanized ones. The AI articles also came from 2024 models, and newer models may be harder to spot. Still, if your readers use AI a lot, I’d assume they’ll notice.
When the five frequent users explained their call on each text, these are the clues that came up most, as a share of all their explanations.
Share of their explanations that mentioned each clue
- Vocabulary53%
- Sentence structure36%
- Grammar and punctuation25%
- Originality24%
- Quotes that sounded formal or uniform22%
Vocabulary is the clue they named most, and it still came up in 42% of their correct calls on humanized articles. Word choice does matter to the best spotters, and cleaning it up is worth doing. As a group, they still caught every humanized article.
What ordinary readers, editors and trained writers notice
Ordinary readers do worse than they think. Online readers told AI-written poems from human ones correctly 46.6% of the time, below a coin flip, and rated the AI poems higher when nobody told them which was which (a 2024 poetry study). People also rely on cues AI can fake. They guess “human” when they see “I”, contractions or family topics, and in short profile self-descriptions, AI text written to use those cues was rated more human than real people’s writing (a 2023 study of people’s AI-spotting cues).
Professional copy-editors look at something else. In one preprint, their “slop” labels followed whether the text stayed on point, how much it actually said and whether its tone fit.
Trained writers are the hardest to please. In a 2026 fiction preprint, readers with graduate degrees in creative writing strongly preferred human imitations of award-winning authors over AI imitations, while general readers favored the AI’s on quality. English teachers held up too. In a 2021 study, they could tell model-written stories from human ones where screened crowd workers couldn’t, though those models came before ChatGPT (the 2021 story study).
Do AI Detectors Work?
The best detectors are now very accurate on plain AI text. They miss more when the AI has been shown a real writer’s work and asked to imitate it, some of them wrongly flag human writing, and none can tell you whether a draft is any good.
I don’t treat beating a detector as a goal. A lower score doesn’t mean better writing.
This chart shows what happens when AI imitates a writer, from an independent July 2026 test by Epoch AI, a research group, and from a 2026 preprint on fiction.
Share of AI-written text each detector missed
AI shown five passages by a writer and asked to write like them
- Pangram10%
- GPTZero10.8%
- Originality.ai17.9%
AI imitating award-winning authors’ fiction, checked with Pangram
- Given 20 excerpts and a style description3%
- Trained on the author’s complete works97%
This table puts each detector’s independent results next to what its vendor claims. Versions, types of text and settings differ between tests, and I wouldn’t average the numbers. If you need a detector, start with Pangram, which held up best across these tests.
| Detector | What independent tests found | What the vendor claims | Cost |
|---|---|---|---|
| Pangram | Epoch (2026, version 3.3.2): flagged 0 of 495 human passages; missed 0% of plain AI text and 10% of imitations. University of Chicago (2025): near-zero errors in six kinds of writing, and still accurate after a humanizer. Notre Dame (2026, version 3.2): flagged 14.9% of human abstracts published in 2023-25 and none from 2013-15 | Pangram 4.1 (Oct 19, 2026, the current version), per its model card: wrongly flags 0.0090% of English text on a cleaned test set and misses 0.3652% of AI text. Pangram 4 claimed 98.83% of humanizer output caught (launch post). Pangram says it’s unreliable on short replies, code, tables of contents and templated text | From $20/month for 300,000 words $0.05 per 100 words through its API |
| GPTZero | Epoch: flagged no human passages; missed 0.67% of plain AI text and 10.8% of imitations. Five-reader study (2025): caught 85%, then 47% after humanizing. University of Chicago: missed about half or more after a humanizer | Under 1% false flags 1.1% on non-native English essays 96.5% accurate on documents that are part human, part AI | $12.99-$24.99/month billed yearly |
| Originality.ai | Epoch (Turbo 3.0.2): flagged 19 of 495 human passages (3.84%); missed 0.34% of plain AI text and 17.9% of imitations | Its current models, Originality.ai 5.0 and 5.0 MAX (Oct 2026), claim 99%+ accuracy (launch post) The Lite and Turbo models, retired in Sep 2026, claimed about 0.5% and 1.5% false flags | Pro $14.95/month for 200,000 words |
| Turnitin | The one independent study we found, a 2026 Belgian study of 160 papers (40 real student theses, 40 fully AI, 40 mixed, 40 mixed and then humanized by a prompt): it cleared the human papers, scored all 40 fully AI papers at 0-20% AI, and caught 60% of mixed and 50% of humanized papers | Under 1% false flags, counted only on documents it scores at 20% AI or higher (Turnitin); detection of AI “bypasser” tools since Aug 2025 | Sold to schools |
| Binoculars (free research detector) | Five-reader study: caught 67%, or 40% at its strict setting; 0-7% of humanized text | Its 2024 paper claimed over 90% caught while wrongly flagging 0.01% of human text | Free |
| Other free detectors (RADAR, Fast-DetectGPT, open-source RoBERTa models) | Five-reader study: Fast-DetectGPT caught 80% with 7.2% false flags, RADAR 15%; University of Chicago: a RoBERTa detector wrongly flagged 30-78% of human text | No vendor | Free |
| Asking a chatbot | GPT-4o with an expert guidebook caught 78%, wrongly flagged 10.7% of human articles and caught 3.3% of humanized ones (older models) | No vendor claim | Any chatbot |
| Five frequent AI users voting together | Caught 99.3% with no false flags, and 100% of humanized articles; one reader alone caught 59-97% (older models) | No vendor | Your own panel |
If your worry is Google and not a detector, that’s a separate question, and I covered it in does Google penalize AI content.
More on false flags, chat training and watermarks
False flags on human writing depend on the detector and the type of writing. A Notre Dame paper tested Pangram 3.2 and GPTZero on academic abstracts. On abstracts published in 2023-25 they flagged 14.9% and 8.9%, which the authors say may partly reflect real AI help, and on abstracts from 2013-15 they flagged none. Lightly AI-edited abstracts were flagged 38-80% of the time, and the authors say detectors can’t tell AI editing from an AI-written draft.
For Originality.ai, the independent false-flag rate ran higher than the vendor’s claim, while Epoch found no false flags at all for Pangram and GPTZero. Epoch tested Pangram 3.3.2, and Pangram’s current version is 4.1, released Oct 19, 2026. Asking a chatbot whether a text is AI works poorly.
Detectors also seem to react to traces of chat training. GPTZero and Pangram often judged text from base models as human (the base-model study). Detectors can also do worse on models they weren’t trained on (RAID), and a brand-new model may score “more human” partly for that reason.
Anthropic says text from Claude Fable 5.1 carries a hidden watermark, a statistical pattern in the model’s word choices with no visible marks (Anthropic, Fable 5.1 docs). By the company’s account, light editing probably won’t remove it and a full rewrite will. Only eligible organizations, such as regulators, researchers and educators, can check for it, and Anthropic’s pages don’t say whether Opus 5.5 or Sonnet 5.5 carry it.
How to Tell Whether It Worked
Put two versions side by side, hide which is which, and have people who use AI for writing often pick between them. Ask whether the writing is good as well as whether it sounds human, because the bar I care about is good writing, and sounding human is only part of that. Detector scores and AI judges can’t reliably tell you either one.
I learned to ask whether it’s good while rating excerpts blind. I kept spotting the human ones because they were bad. My note on one of them: “It’s funny. It’s actually that this article is so bad that I can tell it was human-written.” On another I wrote, “It’s not good. It feels like an unskilled human wrote it though.”
Just because you can tell a person wrote something doesn’t mean it’s worth copying, and human-written content was never all that great. Making it sound like a human wrote it would be a great feat, but our articles also have to be good, rank in search and get cited by AI.
Detector scores can line up with quality, but they don’t measure it. In the fiction preprint, fine-tuned imitations that expert readers preferred also slipped past detectors. AI judges agree poorly with people, too. In LitBench, a test of AI judges on stories, the best one matched people’s choices on 73% of human-written story pairs and was at chance on AI-written ones. In a 2026 preprint on panels of open AI judges, 55% of verdicts reversed when the two texts swapped places.
The people you pick change the answer as well. In the five-reader study, a panel of people who use AI for writing often was the most accurate judge of AI authorship, tied with Pangram.
How to Check a Single Draft
Most of the time you have one draft and nobody to run a test with. A quick check is still worth doing, as long as you treat it as a clue:
- Read it against the stance and content rows of the signs table. Does it take a position? Is there a detail or an opinion only this person would give?
- Mark any sentence you’d never say out loud, and any claim stated more firmly than the person would state it.
- If you can, give it to one person who uses AI often, without saying who wrote it, and ask which sentences they’d change.
One reader’s verdict is a clue and nothing more, because the reliable result in the five-reader study came from five of them voting together.
How to Run a Blind Test of Your Own
Small tests can only show big differences. If readers truly prefer one version 3 times out of 4, you need about 35 pairs where the rater picked a side to have an 80% chance of the test showing it, and a 2-in-3 preference needs about 75.
The same arithmetic cuts the other way. If raters spot the AI 20 times out of 40, their true skill could be anywhere from 35% to 65%, and “our team couldn’t tell the difference” proves very little.
This table has the full steps, which come from the rating studies and from simple win-loss arithmetic, for comparing two versions of the same article.
| Step | What to do | Why |
|---|---|---|
| 1. Write the plan first | Before anyone reads anything, write down what you’re changing, the question raters will answer, how many pairs, and what result makes you switch | In 97 human-rating experiments from 2019 language-research papers, 24% reported any check for chance, and typical designs were too small to detect modest differences (a study of statistical power in language research) |
| 2. Make matched pairs | Write the same brief two ways. Keep the title, facts, outline and formatting the same and the lengths within about 10%, and change one thing per test | Longer text wins partly for being longer: before lengths were evened out, 65% of winning stories in one set were the longer one (LitBench) |
| 3. Hide which is which | Remove every label, use random IDs, and shuffle the left-right order and the order of pairs; the person who made the drafts doesn’t rate them | Telling readers a poem was AI-written lowered its rating by 0.8 points on a 7-point scale, true or not (the poetry study) |
| 4. Pick and check raters | Use people who use AI for writing often. Give every rater 20 samples you know the answer to first, and keep those who get at least 15 right | Frequent AI users as a group caught 99.3% and one alone 59-97%; someone guessing gets 15 of 20 only about 2 times in 100 |
| 5. Add controls | Mix in a few real human articles on similar topics and a few raw one-prompt AI articles | If raters miss raw AI, their verdicts on your versions mean little |
| 6. Ask plain questions | Which reads more like a person wrote it? Which sounds more like us? Which needs less editing before publishing? Allow “no difference”, and have raters highlight the phrases behind their choice | Different questions give different answers, and editors agree far more on problem phrases than on yes-or-no verdicts (the copy-editor study) |
| 7. Count one result per pair | If several people rate a pair, take the majority as that pair’s result | People rating the same pair react to the same two texts, so counting each rating separately makes the result look stronger than it is |
| 8. Plan enough pairs | Set ties aside. For a fair chance of detecting a 3-in-4 preference, plan on about 35 pairs with a winner (about 52 in all if a third end in ties); 2 in 3 needs about 75, and 6 in 10 about 200 | Small tests only detect big differences |
| 9. Know the winning line | With ties set aside and fewer than 9 pairs left, the new version must win them all (6 of 6, 8 of 8), and with 5 or fewer, even a clean sweep can’t rule out luck; after that, 9 of 10, 12 of 15, 15 of 20, 21 of 30 or 27 of 40 | Below these counts, luck alone is a reasonable explanation |
| 10. Finish before judging | Rate every planned pair, even if one version leads early | Checking after every rating and stopping on a lead breaks the simple win-count arithmetic in step 9 (a 2026 study of stopping rules) |
| 11. Treat “couldn’t tell” as no result | If raters spot the AI about half the time, record “not shown either way” | Ordinary readers are near chance anyway, and proving two texts can’t be told apart needs a much bigger test |
| 12. Use detectors and AI judges as screens only | Run them for clues if you like, and let people’s ratings make the call | Scores and quality disagreed in several tests, and AI judges were at chance between AI-written stories |
What I’ve Learned From Grading These
I’ve graded hundreds of A/B tests over the course of this process, including about 50 in this last round: two versions of the same article, side by side, without knowing which was which. No single round proves anything, and I wouldn’t quote the numbers. But that much grading builds an intuition, and it’s where most of what I’ve told you here comes from.
A person’s own words do more than any rule I’ve written, especially on articles where we have nothing else from them. The short-answer boxes we tried at the top of articles all read awkwardly, like AI. Starting an article with the answer is often awkward too, though it’s still useful as long as it’s framed well.
Tools, Skills and Repos Worth Trying
These are Tavi’s recommendations. She sorted everything she found by how much evidence is behind it, and these are the three to start with:
- blader/humanizer, if you’re going to try a humanizer at all. It’s the top pick because it’s the only one we found with a judge or reader test from outside its makers.
- mshumer/unslop, which finds the patterns your own model repeats.
- Wikipedia’s Signs of AI writing, a long list of tells organized by model era and edited almost daily.
Popular doesn’t mean tested. Of the 90 or so humanizer and anti-slop projects we found on GitHub, a site where people share code, most use the same approach, a list of AI patterns to remove, and only blader/humanizer has had a judge or reader test from someone outside its makers. The few that go beyond word lists work on structure, check the draft against its sources, or ask you questions before they write, and we found no such test for those either.
Every tools table on this page is sorted the same way: the ones worth starting with, then the ones just to know about, with a clear standout marked as the top pick. This one has the skills and plugins. A skill is a folder of instructions (a file called SKILL.md, plus any reference files) that Claude, and other AI assistants that read the same format, loads when a task calls for it, and a plugin bundles skills and helpers together. Stars (GitHub’s version of a like) are from GitHub on Nov 1, 2026, licenses come from each project’s own files, and “None found” means we found no published evaluation as of that date.
| Name | What it does | What the evidence says | Cost |
|---|---|---|---|
| Start here | |||
| blader/humanizer Top pick | 26 AI patterns in six groups, strongest first, with a draft, critique and final pass | Why start here: it’s the only one with an outside judge or reader test, and the most widely used. The only one we found with a judge or reader test from outside: won 16 of 16 on quality with AI judges, detection stayed at 98-100%, and judges noticed a more uniform rhythm afterward; 63k stars | Free (MIT) |
| mshumer/unslop | Has the model write about 50 samples in your field, finds what it repeats and writes an avoid-list; run it again when the model changes | Why start here: it finds what your own model repeats, which a fixed list can’t do. None published | Free (MIT) |
| work-humanizer | Rewrite mode plus an interview mode that asks you up to five questions, with a fact check at the end | Why start here: it asks you questions, so the draft gets a person’s input. Makers’ own test with AI judges: picked 48% of the time against blader’s 20%, with fewer changes in meaning; read as human on Pangram 4 of 4 times when given the writer’s notes | Free (MIT) |
| Every’s Compound Writing | A home for each writer’s voice, style, audience and examples, plus reviewer agents, a check of the draft against its sources and a flag for wording copied too closely from examples | Why start here: it checks the draft against your sources, which most of these skip. None found | Free (MIT, per its README) |
| Just to know about | |||
| stop-slop | Eight core rules (cut all adverbs, no em dashes) and a 50-point score | None found; 18.5k stars; no update since Mar 2026 | Free (MIT) |
| no-ai-slop | Removes 20+ patterns, with edit, detect and satire modes; also a ChatGPT plugin | None found; 13k stars | Free (MIT) |
| avoid-ai-writing | Rewrite, detect and edit modes, five voice profiles and a bundled detector | Only the maker’s own calibration file | Free (MIT) |
| sepia | For fiction, works on story structure first, then flow, then wording, choosing three to five changes per story; other prose gets rule files matched to where it will run | Has a way to evaluate itself, with no results published | Free (MIT) |
| theclaymethod/unslop | Scans phrases, structure and how ideas are arranged (endings that call back, previews the body repeats) | Its makers’ latest result says not good enough to release, with no steady gain across different AI models | Free (MIT, per its README) |
| humanizer-stack | A wording pass, then six structure checks on an outline of the draft, one at a time | None found | Free (MIT, with some CC BY-SA material) |
| agent-style | 21 writing rules loaded before writing | Makers’ own test: counted tells fell 46-86% on three of four tools; quality not measured | Free (CC BY 4.0 rules, MIT code) |
| humanize-writing-skill | Rules on claims, sentence shape and endings | Makers’ own small test: Claude judges preferred it in 5 of 6 and 3 of 3 pairs; says it doesn’t change detector scores | Free (MIT) |
| writing-style-skill | You edit the AI’s drafts, and it extracts rules from your edits | None for the tool; the idea worked with simulated users in a 2024 study | Free (MIT, per its README) |
| nobuzz, vomit and SlopMonster | Rewrite Claude’s output through a second model (Gemini, a model on your own computer, or a rival model) | None with readers; the first model’s style survives rewording, and vomit warns it can make things up | Free |
| Anthropic’s brand-voice plugin | Built by a partner (Tribe AI): finds a brand voice in your documents and checks content against it | None published | Free (MIT) |
The word lists are under “Banned Words and Writing Rules” and the apps are under “AI Humanizer Tools”. Everything else is on the searchable page with every tool, repo, study, sign and claim we checked.
See the open-source tools and the AI-writing checkers
The first table is tools that control the model’s word choices as it writes. The ChatGPT, Claude and Gemini apps don’t let you do that, and current Claude models don’t allow it through the API either. Most readers won’t need these unless they run a model themselves.
| Name | What it does | What the evidence says | Cost |
|---|---|---|---|
| Start here | |||
| antislop-sampler Top pick | Backs up and retries when the model writes a banned word, phrase or pattern | Why start here: it’s the one with a paper behind it, and it handles far more patterns than plain bans. The Antislop paper: handles 8,000+ patterns where plain token banning fails at about 2,000; needs a model whose word probabilities you control | Free (Apache-2.0) |
| slop-forensics | Finds each model’s overused words and phrases | Why start here: it shows what a model overuses before you decide what to ban. The source of EQ-Bench’s phrase lists | Free (MIT) |
| Just to know about | |||
| auto-antislop | Finds a model’s stock phrases and trains them out | The paper reports about 90% less slop with writing scores kept | Free |
| KoboldCpp and SillyTavern | Ban phrases or word fragments while a local model writes | No quality tests | Free (AGPL-3.0) |
| Unslopper-30B-A3B | A model trained to rewrite fiction without AI phrasing | Maker’s own numbers: Pangram’s human score rose from 0.000 to 0.481 while judged quality fell from 8.60 to 7.96 | Free |
| logitloom and loom-mcp | Generate many continuations for a person to pick from; logitloom works through some hosted model services | None; loom-mcp depends on prefill (writing the first words of Claude’s reply), which current Claude models reject | Free |
| pseudobibeR | Counts grammar features to compare a draft with a person’s own writing | No study links its counts to readers’ judgments | Free |
The second table is linters and checkers. A linter is a program that flags problems in text against a set of rules.
| Name | What it does | What the evidence says | Cost |
|---|---|---|---|
| Start here | |||
| Vale with vale-ai-tells Top pick | A prose linter plus a pack of AI-tell rules | Why start here: it’s an established linter, and the rule pack adds AI tells you can edit. None for AI tells | Free (MIT) |
| Hemingway Editor | Highlights hard-to-read sentences; the paid tier rewrites them | Why start here: it’s free, it’s easy to try, and it flags the hard sentences a person should look at. None for AI tells | Free; paid tier with a trial |
| Just to know about | |||
| proselint | Flags weak or awkward prose | None for AI tells | Free (BSD-3-Clause) |
| sloptrim and slop-guard | Score prose files for AI tells | None found | Free |
| LLM Slop Detector (Firefox) | Flags invisible characters, em dashes, curly quotes and stock phrases on web pages | None found; 4 daily users | Free (MIT) |
| Slop or not | A game where you pick the AI reply out of a pair | No accuracy table, and the players aren’t a random sample | Free |
| EditLens | A research model that estimates how much AI editing a text has | Only its authors’ own test, and three of its four authors work at Pangram | Free for non-commercial use (CC BY-NC-SA 4.0) |
Common Claims About AI Writing, True or False
Much of what gets repeated about AI writing and AI detection is out of date or only partly true. This table checks the claims about how AI writes and how detection works against the evidence as of Nov 1, 2026. The studies behind them are all on the research library page.
True1
| Claim | Verdict | Why | Source |
|---|---|---|---|
| AI avoids taking a side | True | AI-written essays stayed neutral about 69% more often than people’s (one small study) | Neutral-essays study |
Partly8
| Claim | Verdict | Why | Source |
|---|---|---|---|
| AI hedges too much | Partly | It softens claims about as often as people in some counts, but it rarely says “I think” or commits to a view | Essay study |
| All AI models sound the same | Partly | Much of the style is shared, yet a program told five chatbots apart 97% of the time | Chatbot-fingerprint study |
| AI’s style comes from its training data | Partly | Much of it traces to the training after a model reads the web, since models before that training write closer to people; most of the sentence patterns models repeat also appear in the web text they learned from | Model-continuation study, Sentence-template study |
| AI doesn’t use contractions | Partly | People added contractions when editing AI drafts, but our 121 articles, written with instructions to match each brand’s voice, used twice the human rate | Post-editing study |
| Humanizer tools make AI text undetectable | Partly | Results conflict: after one commercial humanizer, both detectors in a Notre Dame paper flagged under 4%, while other studies found Pangram still caught most humanized text | Notre Dame study, Five-reader study |
| Detectors almost never flag human writing | Partly | Near zero in two tests, but two detectors flagged 9-15% of recent human-published abstracts in one study, and 13 detectors flagged anywhere from 0% to 100% of non-native English writers’ manuscripts | Edited-manuscripts study, Notre Dame study |
| Claude Opus 5.5 is the best AI writer | Partly | First on ToneBench by about the normal draft-to-draft variation in its scores, though its rating range doesn’t overlap the next model’s and the board calls its results directional; tied for first on Arena; seven models beat it on EQ-Bench fiction | ToneBench, EQ-Bench |
| Base models sound more human than chat models | Partly | True for two detectors and for grammar counts; we found no reader test | Base-model study, Model-continuation study |
False16
| Claim | Verdict | Why | Source |
|---|---|---|---|
| AI never hedges | False | ChatGPT essays had fewer hedges than students’ essays, which still means some, and our drafts use “may” and “typically” at about the human rate | Essay study |
| Ban hedging words to sound more human | False | AI already commits less than people do, and Wikipedia’s editors saw “perhaps” and “tends to” more in human-written Wikipedia articles than in AI text | Wikipedia’s editors |
| Remove the AI words and the text passes as human | False | Five frequent AI users voting together still caught every humanized article | Five-reader study |
| People write the way they talk | False | The same people used more emotional language speaking than writing | Speaking-versus-writing study |
| Perfect grammar or bland prose means AI wrote it | False | Wikipedia’s editors list both as signs that don’t work | Wikipedia’s editors |
| The humanizer leaderboards are independent | False | Every one found is run by a vendor or has affiliate links, and they rank the same tools far apart | HumanizerBench |
| A skill with tens of thousands of GitHub stars has been shown to work | False | stop-slop (18.5k stars) and no-ai-slop (13k) publish no results; of the 90 or so we found, only blader/humanizer has a judge or reader test from outside | stop-slop |
| AI detectors are coin flips | False | On plain AI text from 2026 models, Pangram and GPTZero missed under 1% and flagged no human passages | Epoch AI |
| A lower detector score means the writing sounds more human | False | In one user’s test, AI judges preferred a rewrite in 16 of 16 comparisons while detection stayed at 97.8-100% | blader/humanizer test |
| You can ask ChatGPT whether a text is AI | False | GPT-4o with a guidebook wrongly flagged 11% of human articles and caught 3% of humanized ones (older models) | Five-reader study |
| Most people can tell AI writing by feel | False | People who rarely use AI were at about chance | Five-reader study |
| One experienced reader is a reliable judge | False | Individual frequent AI users caught 59-97%; five voting together caught 99.3% | Five-reader study |
| If readers like it, it passes as human | False | General readers favored AI imitations that expert readers rejected | Author-imitation study |
| A blind test of 8 pairs can prove one version is better | False | With 2 ties, 8 pairs leave 6 with a winner, and only 6 of 6 would rule out luck; 5 wins and 1 loss fits a true win rate anywhere from 44% to 97% (our arithmetic; typical rating studies are too small to detect modest differences) | Statistical-power study |
| The most expensive model writes best | False | Fable 5.1 costs 2.5 times as much as Opus 5.5 per word processed (1.4 times as much per ToneBench script at max effort) and scores lower there (84.53 against 85.93) | ToneBench |
| ChatGPT’s personality setting changes the style of its drafts | False | OpenAI says personality doesn’t control drafts such as emails or posts | OpenAI |
Out of date4
| Claim | Verdict | Why | Source |
|---|---|---|---|
| “Delve” gives AI away | Out of date | It rose after ChatGPT and fell once it was publicized in 2024; none of our 76 drafts used it | ChatGPT-words study |
| An em dash means AI wrote it | Out of date | Wikipedia’s editors now list it as a sign of older models, the newest models avoid it, and professional writers use it too | Wikipedia’s editors |
| Detectors flag most non-native English writing | Out of date | A widely quoted 2023 study found detectors flagged 61% of non-native English essays; in 2026, rates ranged from 0% to 100% depending on the detector | Edited-manuscripts study, 2023 study |
| A custom GPT or a Gemini Gem is the place to keep your voice | Out of date | Custom GPTs are scheduled to retire Dec 11, 2026, and Gems on adults’ personal accounts are removed Dec 17, 2026 | OpenAI, Google |
Unproven1
| Claim | Verdict | Why | Source |
|---|---|---|---|
| A “40% AI” score means AI wrote 40% of the text | Unproven | Detectors can’t tell AI editing from AI drafting | Notre Dame study |
The Hard Limits on How Well AI Can Write
To be completely blunt, this process has been a little demoralizing. I really felt like we could just keep improving our skills and make our AI-written content come out better and better and better until it was nearly human.
Now I’m realizing there are some hard limits on how well AI can write until the models get better at some of these things that have proven to be very difficult, the kind of things that are difficult even for humans. Better understanding why AI is bad at writing has made me think we need better models before AI can truly be good at writing.
A real person’s point of view is still what gets you furthest today, and it won’t get you all the way.
Could You Tell AI Wrote This?
Tavi, our AI system, wrote this article using Claude Opus 5.5 on extra-high effort, through the same article process we use for client articles. She worked from my own words (my request for the article, my messages during the research and testing, and what I’ve said on calls and in videos) and from the research. The point of view is mine, and Tavi did the writing.
So, could you tell? If you could, I’d like to know where it gave itself away, because that’s the part we’re still working on. If you couldn’t, keep in mind that one reader not noticing proves very little.
When a client can give us a few minutes of talking, this is how we start their articles, as part of their SEO and GEO campaigns (showing up in Google and in AI answers). When they can’t, we work from their past calls or take a position from experts the brand agrees with, and the plan is for the client to vet those positions. If you have questions about any of it, or you spotted the sentence that gave it away, reach out through our contact page or email me at tj@tjrobertson.com.