Which phrases make a message read as AI-written?
Three kinds of thing give a message away: stock openers and sign-offs, a short list of inflated words, and a few sentence shapes, such as a contrast against something nobody said or a trailing clause that comments on the sentence before it. One of them proves nothing. Three in a short message is what readers notice, and each one has a plain swap.
This list comes from the checker Ownhand runs on every draft it writes, plus the public research behind it. None of it proves a person used AI; it is what makes a reader stop believing you wrote it yourself.
Openers and sign-offs
The first and last lines carry the most weight, and they are where model drafts lean hardest on stock phrases. These are the ones the email_cold and other email occasions refuse outright.
| Phrase | What to write |
|---|---|
I hope this email finds you well | Nothing. Start with why you are writing. |
I wanted to reach out | Say the thing you reached out about. |
just following up | Name what you are following up on, with the date or number. |
circling back | Same: name the topic and the ask. |
Please don't hesitate to reach out | Drop it. People know how to reply. |
Would love to connect | Ask for one specific thing, such as a 15-minute call on Thursday. |
I hope this helps | Drop it, or ask whether it answered the question. |
All of them talk about the message instead of saying it. Two real drafts, and what two writers made of them:
These examples are real output from the live API. Each AI draft was rewritten by a demo Hand: a fictional writer (an engineer, a founder, an ops lead, a designer or a consultant) built from a handful of samples.
AI draft: I hope this message finds you well. I wanted to reach out because I believe our platform could help your team streamline invoicing and save up to 10 hours a week.
Demo Hand, designer: Our platform could speed up your invoicing and save your team up to 10 hours a week.
AI draft: I just wanted to quickly circle back on the Q3 budget numbers. Please let me know if you need any additional information from my end!
Demo Hand, ops: Q3 budget numbers. Let me know if you need anything else.
Both rewrites kept every fact and cut the throat-clearing. The invoicing one is half the length of the draft and says the same thing.
Words
Single words are the weakest signal and the easiest fix. The swaps are the ones Ownhand's checker suggests; where it says "drop it", the sentence reads better without the word.
Inflated verbs
A long word where a short one does the job.
| Word | Plain swap |
|---|---|
leverage | use |
facilitate | help, or run |
streamline | simplify, or speed up |
empower | let, or help |
ensure | make sure |
foster | build, encourage, or help |
delve | dig into, or look at |
Importance and certainty
Telling the reader what counts, in place of showing it.
| Word | Plain swap |
|---|---|
crucial | important, or drop it |
pivotal | important |
fundamentally | (drop it) |
genuinely | (drop it) |
arguably | (drop it, or commit) |
Connectors
Essay glue. In a message, the sentence order does this job.
| Word | Plain swap |
|---|---|
moreover | (drop it, or 'also') |
furthermore | (drop it, or 'also') |
additionally | (drop it, or 'also') |
ultimately | (drop it) |
Office abstractions
Nouns standing in for a person, number or plan you could name.
| Word | Plain swap |
|---|---|
stakeholders | the people involved (name them) |
insights | finding, or what we learned |
outcomes | results |
initiatives | plan, project, or program |
actionable | useful, or something you can act on |
robust | strong, or solid |
seamless | smooth, or (drop it) |
Decoration
Words nobody says out loud about a budget or a deploy.
| Word | Plain swap |
|---|---|
tapestry | (say the actual thing) |
testament | proof, or sign |
myriad | many |
plethora | a lot of |
realm | area, or field |
Sentence shapes
Readers notice these even when they cannot say why. A 2025 PNAS study found models lean on certain grammar far more than people do. In press coverage of the Carnegie Mellon study, present participle clauses showed up at two to five times the human rate and nominalizations (turning a verb into a noun) at 1.5 to 2 times.
| Shape | Example | Plain version |
|---|---|---|
| A contrast against something nobody said | It wasn't a bug. It was a choice. | We did it on purpose. |
| A trailing clause that comments on the sentence | The bridge opened in 1932, reflecting the city's ambitions. | The bridge opened in 1932. |
| Announcing that something is important | This matters because | Give the reason and let the reader judge. |
| A setup before the point | Here's the thing | Start with the thing. |
| A summary line at the end | At the end of the day | Stop after the last real sentence. |
| Stock metaphors | move the needle, load-bearing, at the heart of | Say what changes, and by how much. |
This pull request draft stacks a listed word on a clause tacked on with ensuring
. The rewrite says what the code does.
AI draft: This pull request introduces comprehensive retry logic for the payments webhook, ensuring robust handling of transient failures. Max retries are set to 5.
Demo Hand, consultant: This pull request adds retry logic to the payments webhook. Transient failures are retried up to 5 times. Review welcome.
What is missing
Some tells are absences, and no ban list fixes them. Model drafts tend to have no contractions, sentences of one length, and no asides or opinions. A person's message has rough edges: a fragment, a two-word sentence, a lowercase start in Slack. If the draft is three times longer than the last thing you sent the same person, they will notice before any listed word.
Formatting
- Markdown where it does not render: bold labels, headings and bullets in an email or a Slack DM.
- A list for two items, or a heading over one paragraph.
- Long dashes used as a break in every other sentence. Wikipedia's guide lists their overuse as a sign of model text.
- Leftovers: a bracketed placeholder for your name, a knowledge-cutoff note, a pasted citation marker.
Check a draft in two minutes
- Read the first and last lines. Delete any opener or closer from the table above and start with the point.
- Search for the words in the word tables. Use the swap, or cut the word.
- Look for the shapes: a contrast, a trailing clause, a closing summary. Rewrite each one as a plain statement.
- Strip Markdown and dashes if the message is going to email or chat.
- Compare the length with what you normally send this person, and cut to match.
Where Ownhand fits
Ownhand's write tool runs this checker on every rewrite. When a draft trips it, the model gets the list of problems and tries again, up to two more times; anything left is reported to your agent. It also checks the occasion, such as a cold email over 125 words or Markdown in a plain-text email.
Removing tells only gets a draft to neutral. Your Hand, built from things you wrote, supplies the rest: your length, your openers, your punctuation. It keeps learning from what you send (see how Ownhand learns from what you send).
You do not need it if you write most messages yourself and only want a check: use the steps above, or paste a draft and these tables into any chatbot and ask it to flag matches. For why drafts end up like this in the first place, see why AI drafts sound like AI.
Follow-up questions
Does one AI-sounding word mean a message was written by AI?
No. Every word on these lists is ordinary English that people used long before chatbots. Readers notice a cluster: a stock opener, two or three inflated words and a tidy closing line in what should have been one sentence.
Can I just tell ChatGPT or Claude not to use these words?
It helps a little. A ban list in custom instructions removes the exact words, but the model reaches for a near synonym and keeps the sentence shapes. Checking each draft after it is written, and asking for a fix when it fails, works better.
Are long dashes a sign of AI writing?
They are one of the best known. Wikipedia's guide notes that model output uses them more often than casual human writing of the same kind. Careful writers use them too, so treat a dash as one signal among several.
Which tells should I fix first in an email or a Slack message?
The first line and the last line. They are read first, and their stock versions are the most recognized. Cut the padding and the closing offer, start with the point, then match the length to what you normally send that person.
Sources
Checked on 2 October 2026. If something here is out of date, email hi@thalientlabs.ai and we will fix it.
- Wikipedia: Signs of AI writing: word lists, collaborative closers such as offers of more help, knowledge-cutoff disclaimers, overuse of long dashes
- Reinhart et al., Do LLMs write like humans? (PNAS, 2025): grammatical and rhetorical style of model text against human text
- TechXplore: coverage of the Carnegie Mellon study: participle clauses at two to five times the human rate, nominalizations at 1.5 to 2 times
- Ownhand source code: the word list, sentence patterns and form-letter phrases the write tool checks