In short: AI text sounds human when it carries a real stance (opinions, doubts, first-hand examples) and when the model can read a person's voice instead of being told about it. Banned-word lists and "write like a human" barely help. That is how T.J. Robertson summed up hundreds of graded A/B tests on 1 November 2026. We turned it into three changes to Hermes 3000: style samples, "My take" for nonfiction and whitepapers, and no more invented sources. Measured: with "My take", 12 to 21 of the author's word sequences end up in the text instead of none, and the author's doubts stay in.
Anyone who writes with AI knows the tone: smooth, balanced, somehow correct, and nobody's. A nonfiction chapter that explains everything and claims nothing. A paragraph that reads as if a very diligent intern wrote it, one who has read a little about every topic.
The usual answers are tricks: lists of banned words, "write more casually", a second model that "humanises" the text. We wanted to know which of these actually work. The best answer we found does not come from research. It comes from an agency's day-to-day practice.
Where the idea comes from: T.J. Robertson
T.J. Robertson founded the digital agency TJ Digital in Henderson, Nevada, and has spent three years building its process for writing articles with AI. On 1 November 2026 Robertson published what came out of it: How to Make AI Sound More Human. It rests on hundreds of graded A/B tests, about 50 of them in the latest round, plus a review of the studies on the subject. Robertson's earlier study of which pages AI answers cite had already helped us make this blog easier to find in October.
The core of the new article in one sentence: give the AI a real point of view from a real person. Tinkering with the wording does not replace that.
| Helps (with limits) | Does not help |
|---|---|
| Let the person talk for 2–5 minutes and write from the transcript | "Sound human", "write casually", "try harder" |
| Two to five varied writing samples from that person | Piles of rules prescribing human writing |
| Offer several versions and let a person choose | More than about ten examples, examples on the same topic |
| Mark missing facts instead of letting the model invent them | A second model rewriting the text; an AI acting as judge |
| A short list of patterns that really show up in your own drafts | Long, generic banned-word lists: the model just finds new stock phrases |
Two points convinced us in particular. First, the reasoning: language models are trained to give the most expected answer, which includes the most expected opinion. You cannot coax a stance out of a model; you have to hand it one. Second, the honesty about testing. According to Robertson, neither AI detectors nor AI judges reliably tell whether a text sounds human. What counts is a blind comparison by people who read a lot of AI text.
What we found in Hermes
We read our own writing prompts against the article. Some things already fit: nowhere do we say "sound human", our variants feature offers three versions of a selected passage for you to choose from, and for model comparisons we have a blind human jury in the Text Arena.
Three things did not fit:
- The style analysis threw your text away. It read your sample, turned it into a description ("Write in short sentences …") and kept only the description. According to Robertson, that is the weaker half.
- The nonfiction prompt asked for "impersonal and objective". That is exactly the stance-free tone people recognise as AI text.
- The whitepaper prompt asked for source references "[Source, Year]". Without material to draw on, that is an invitation to invent studies and years, in texts that go out under your name.
Change 1: no more invented sources
Whitepaper texts now take figures, studies and sources only from the material you provide: your knowledge collection and your book's source and evidence entries. If a statement needs support the material does not give, Hermes phrases it cautiously and adds [source needed]. You see the gap and fill it, instead of missing a plausible fake.
In our test we had a chapter on the four-day week written with no material at all. Before, the prompt would have asked for source references. Now the text contained not a single invented study, percentage or year, and "[source needed]" nine times instead.
Change 2: style samples that go into every generation
In the book structure, under "Writing Style & Audience", you now find style samples: two to five short passages from your own writing. Hermes passes them along every time it writes, with the instruction to take on sentence length, rhythm, word choice and tone, but not the content and not a single sentence. The style analysis now asks whether it should keep the analysed passage straight away.
We measured with a novel scene in a bakery and two samples on completely different topics, short, present tense, first person:
| words per sentence | "I" in the text | 5-word sequences copied from the samples | |
|---|---|---|---|
| without samples (2 runs) | 11.3 / 11.7 | 1 / 1 | 0 / 0 |
| with samples (2 runs) | 9.4 / 9.1 | 23 / 11 | 0 / 0 |
The voice comes through, nothing is copied. While measuring we noticed something we had not expected: the first-person samples also carried over the point of view. That is right when there is no text yet. It is wrong when your novel is told in the third person. A side remark in the prompt ("the existing text takes precedence") did not help. Only a separate, explicit paragraph kept the third person, in two runs out of two. And when a sample comes from the same scene world, its motifs rub off, just as Robertson describes. That is why Hermes asks for varied passages.
Change 3: "My take" for nonfiction and whitepapers
Robertson's most important recommendation is to let the person talk first. That does not need an AI interview. A small transcript is enough. So nonfiction books now have a "My take" field (whitepapers call it "Our position"): once for the whole book under "Writing Style & Audience" and, if you like, for each chapter below its summary. Robertson's guiding questions sit above the field:
- What do most people get wrong about this?
- Where do you disagree with the usual advice, and why?
- What are you unsure about, or what have you changed your mind on?
- An example from your own work?
- How would you explain it to a customer in two sentences?
The easiest way is to press the microphone and talk for two minutes. Rough is perfectly fine. Hermes then writes the text around your statements. Your wording stays where it carries, your experiences stay in the I or we form, and where you are unsure, the text stays unsure. Nothing is invented in your name.
Our test was a section on "Budget planning in small teams", with a made-up but typical take of the kind you dictate in two minutes. It had a clear contrarian opinion ("I consider annual budgets in teams under twenty people to be theatre"), a first-hand example (the agency abolished the annual budget in 2019) and an open doubt (whether it works without someone who has the numbers in their head).
| 4-word sequences kept | "I"/"we" | doubt kept | own example in the text | |
|---|---|---|---|---|
| without a take | 0 | 0 | – | never |
| first version | 2 | 0 | no | in 1 of 2 runs |
| current version (3 runs) | 12–21 | 5–7 | always | in 3 of 3 runs |
The first version taught us something. The stance came through but was depersonalised: "At our agency we abolished the annual budget in 2019" became "In one agency, the classic annual budget was abolished", and "honestly, I'm not sure about that" became "remains questionable". The culprits were two rules in our nonfiction prompt, "impersonal" and "no narrative perspective". They no longer apply once you have entered a take. After that change, the sentence from the take appeared in the text: "I simply consider annual budgets in teams of this size to be theatre."
A small remainder is still there: in one run out of three the model added an I-sentence that was not in the take. So it is worth a quick check that every I-sentence really is yours.
What we do not know yet
We measured whether your voice and your take arrive in the text. We have not yet measured whether readers prefer the result. Our samples are small, two to three runs per variant with one model. And Robertson is clear on this: measurements and AI judges do not replace a blind comparison by people. That is what we do next in our Text Arena, with sections written with and without "My take", and we will report the result here.
How to try it
- Open your book and, in the book structure on the left, "Writing Style & Audience".
- Under "Style samples", add two to five passages of your own, as varied as possible: a dialogue, a description, a thought.
- For nonfiction and whitepapers, press the microphone next to "My take" and answer the guiding questions in your own words.
- Have a section written and read whether it sounds like you.
Frequently asked questions
Does AI text sound more human with style samples?
It sounds more like you: in our test the text took on the samples' sentence length, tense and narrative stance without copying a single sentence. Whether readers prefer it for that is what our blind test is checking.
Will an AI detector stop flagging the text?
That is not the goal. Robertson also advises against writing for detectors. If you want to show what is yours, our provenance record serves you better.
Do I have to dictate?
No, you can type as well. Dictating is just faster and brings out the sentences you would really say.
Is "My take" available for novels?
No, only for nonfiction and whitepapers. In a novel the narrator speaks, not the author. Style samples help there.
What does it cost?
Style samples and "My take" are part of normal writing in Hermes 3000. The first 50 standard pages are free (as of November 2026).