Start here: 200 words and a blank box#
You have probably asked someone to read something before you sent it: a friend checking a cover letter, or a parent reading a tricky text. A good reader tells you where they got lost and which parts sound like you.
An AI chat assistant (a program you type to, such as ChatGPT, Claude, Gemini or Copilot) can do that same job at any hour. The one new part: this reader can also write the whole thing for you, in seconds. That is where both the help and the trouble come from. This module is about using the reader while staying the writer.
We will follow one piece of writing from blank page to sent. Maya is made up, but her problem is common. She is applying for a part-time apprenticeship at a community bike repair shop. The form asks:
In about 200 words, tell us why you want to join and something you have fixed.
She has stared at the box for twenty minutes. Here are her notes, exactly as she typed them on her phone:
- back tire went flat 3 times in one month, got tired of paying, learned to fix it
- 2 years now, do all my own repairs. youtube + a library repair book
- last summer neighbor's son's brakes stopped working. took me 3 tries
- like figuring out why something squeaks
- never worked in a shop. no certificate. kind of nervous about that
- want to learn wheel truing + the stuff videos skip
- want people who can't pay shop prices to keep ridingEverything Maya needs is in those notes. What she lacks is a first sentence, an order, and the nerve to sound like herself. Every idea below gets tried on her 200 words.
The exercises use Maya's made-up notes. Do not paste your own private details into a chat assistant to practice. When you use these methods on your own writing, check what your school or employer allows you to put into AI tools.
Where AI helps your writing, and where it hurts#
Maya could ask the assistant to (a) write the statement from her notes, or (b) ask her questions until she knows what she wants to say. Which is more likely to end with a statement only Maya could have written?
Show the answer
Choice (b). A chat assistant writes by predicting likely words, so a statement it writes tends to sound like many other statements. Questions make Maya do the thinking, and the thinking is what a reviewer is reading for.
Writing is several jobs stacked together: deciding what you think, choosing what to include, putting it in order, finding the words, cleaning up. AI is good at some of those jobs and risky at others.
| Job | How AI usually helps | What to watch for |
|---|---|---|
| Getting unstuck | Asks you questions, or offers a few possible first sentences | Grabbing the first option instead of finding your own |
| Outlining | Suggests an order for points you already have | A standard order that fits everyone and no one |
| A rough first draft | Fine for low-stakes text, such as a notice about opening hours | For anything personal, the draft quietly becomes the thinking |
| Reorganizing and tightening | Moves a buried point up; cuts 300 words to 200 | Cutting the specific detail and keeping the filler |
| Catching unclear sentences | Points at a sentence a reader may stumble on | Rewriting it instead of pointing at it |
| Adjusting tone | Makes a message warmer, firmer or shorter | Overshooting into stiff or gushing |
| Translation drafts | Gives a starting version in another language | You cannot check what you cannot read |
AI helps most when you supply the substance and it works on shape. It hurts in four ways:
- Thinking you skipped. Working out what to say is often how you find out what you think.
- A generic voice. Your voice is the way your writing sounds like you: your word choices, your rhythm, what you notice. AI text drifts toward a smooth average that belongs to nobody.
- Invented facts. It can add a detail or a memory that never happened, in the same confident tone as everything else.
- Sameness. When many people start from the same tool, their writing starts to look alike.
That last one has been measured. In a 2024 experiment, 293 people each wrote an eight-sentence story, and some could first ask an AI tool for a story idea. A separate 600 readers, who were not told at first which stories had AI help, rated the AI-helped stories more creative and better written on average. But those stories were also more similar to each other than stories people wrote alone (Doshi and Hauser, 2024). On average writers did a bit better, mostly the less creative ones; together they produced a narrower range.
For Maya, sameness is the worst outcome: a reviewer reads a pile of statements in one sitting.
You think, it polishes#
Suppose Maya lets the assistant write the statement and she just approves it. What must she now check, line by line, that she never needed to check in a draft she wrote herself?
Show the answer
Every fact. In her own draft, each fact came from her memory. In the assistant's draft, any sentence could describe something that never happened to her, and a flattering untrue line is easy to skim past.
There are two basic ways to write with AI. The drafts can look similar, but who did the work is reversed.
| You think, it polishes | It writes, you approve | |
|---|---|---|
| Where facts come from | Your notes and memory | Anywhere, including nowhere |
| What you must check | The changes it suggests | Every sentence, every fact |
| Good for | Anything with your name on it that matters | Routine text where voice does not matter |
"It writes, you approve" is fine for a sign saying the shop closes early on holidays, as long as you check the details. For her statement, Maya uses the other order:
- Dump everything, messy. Her phone notes are this step.
- Get questions, not answers, until you know your point. Then ask for a suggested order and pick one yourself.
- Draft it yourself. Rough and yours beats smooth and borrowed.
- Ask for a check on unclear sentences, length and tone. Ask it to point, not rewrite.
- Choose each change, then read it out loud. Anything you would not say, change.
Open any free chat assistant and paste this prompt:
I'm writing a 200-word personal statement for a part-time apprenticeship at a community bike repair shop. Do not write the statement. Ask me five questions, one at a time, that would help me work out what to say. Wait for my answer before asking the next one.Answer as Maya, using only her notes. What to look for: Were the questions about bike repair, or could they fit any job? Did it wait, or rush ahead and write the statement anyway? Which of your answers would make the best first sentence? That sentence came from you.
You met this one in Module 2; here is the writing lesson in it. In August 2026 the owner asked our AI assistant for some prompts. It handed back a reusable template plus a list of things to paste into it, when he wanted the finished prompts.
The rule: say what form you want the result in. "Ask me questions" gets questions. "Give me three possible first sentences" gets options. "Write it" gets a finished draft, whether or not that is what you needed.
AI writing habits readers notice#
Two openings. (A) "I am deeply passionate about cycling, community and sustainability, and I am excited by the opportunity to contribute to your esteemed program." (B) "My back tire went flat three times in one month, so I learned to fix it myself." Which would a reviewer remember after forty statements?
Show the answer
B. It is specific and starts a small story. A could open almost any application. A person could have written A; many do. The point is that A stacks several habits readers have learned to skim.
AI writing has habits. None is wrong on its own, and people use all of them. The trouble is the pile-up: readers who see a lot of AI text can learn the pattern and stop paying attention.
| Habit | What it looks like | Edit toward |
|---|---|---|
| Over-hedging | "It may be worth considering that I could potentially..." | Say it. If unsure about one thing, say which, once. |
| Lists of three | "I have honed my skills, cultivated my passion and strengthened my work ethic." | Keep the one that is true and give an example. |
| Stock phrases | "excited by the opportunity", "a testament to" | Your own words, or nothing. |
| The long dash, again and again | Two or three wide dashes per paragraph | A colon, a comma or a period. |
| Overly formal tone | "I would be most grateful for your kind consideration." | How you would say it face to face. |
| Tidy wrap-ups | A last line repeating the paragraph in nicer words | Stop when you have said it. |
The "long dash" is the em dash, about as wide as a capital M. You will not find one anywhere in this course, and here is why.
Kestrel Exams is a small study site built by one person working with an AI assistant. In August 2026 the owner flagged our AI assistant's heavy use of the em dash as an "AI tell": a habit readers have learned to link with machine writing. Our house style now uses colons, commas and periods instead. The rule: AI writing has habits readers recognize, so edit for your own voice.
Swap every long dash in an AI draft for a comma and delete "excited by the opportunity", and it is still not yours: the thinking, the choice of details and any invented facts are the assistant's. The goal is not to hide AI use. It is a page that says what you mean, the way you would say it. That comes from putting back what only you know.
Your voice, your facts#
Maya gives the assistant a paragraph of her own writing and asks it to "make my draft sound like me". Suppose the style match is perfect. What could still go wrong?
Show the answer
It can add things that are not true, in her style. A convincing fake memory in Maya's voice is more dangerous than a stiff real one, because it no longer stands out. Matching the voice does nothing to protect the facts.
Maya writes her own rough draft from her notes: about 150 words, a little clumsy, completely hers. Now she wants help smoothing it. Two habits make that work.
Give it a sample of how you write. A short paragraph you would be fine with anyone reading, names removed. Maya uses a message she once sent a friend:
Took the whole chain off tonight. Turns out the squeak was the pedal, not the chain. Two hours well spent, I guess.Short sentences, dry humor, plain words. (Every home mechanic has lost an evening to the wrong squeak.) Ask the assistant to describe your style before it edits, and correct the description if it is wrong.
Ask it to point, not rewrite. "Mark any sentence a reader might stumble on, and say why" keeps you as the writer. "Rewrite this to be clearer" hands the pen back.
When we measured what made our own study guides hard to read (the full story is in Module 5), the cause was not long sentences. It was technical words and symbols never defined: one term was used 17 times and defined zero times. The writing rule: define each term the first time it appears, and give a picture before the procedure.
Maya's draft has this exact problem: she wants to learn "truing" and never says what it is. An assistant asked to point at unclear words should catch it, and Maya explains it her way: straightening a wheel so it spins without a wobble.
The rule that matters more than voice: never let it invent facts, quotes or experiences in your name. Anything you sign is a claim you are making.
When Maya tried "it writes, you approve" out of curiosity, the draft said she had "volunteered at community repair events for three years". She never has. It reads beautifully and is exactly what an interviewer asks about. Two defenses: tell the assistant not to add anything, and check every claim in the result against your notes. If it is not in your notes, it is not in your statement.
Paste this whole prompt, which includes Maya's sample and her draft:
Here is a short sample of how I write:
"Took the whole chain off tonight. Turns out the squeak was the pedal, not the chain. Two hours well spent, I guess."
First, describe my style in three short points. Then edit my draft below for clarity. Keep it under 200 words.
Rules: keep my words where they work. Keep my sentences short. Do not add any fact, number, experience, skill or quote that is not in my draft. Point out any word a reader might not know.
After the edit, list every change you made, and anything you were tempted to add but did not.
My draft:
My back tire went flat three times in one month, and I got tired of paying someone to fix it, so I learned to do it myself. That was two years ago. Now I do all my own repairs, mostly from videos and a repair book from the library. Last summer my neighbor's son's brakes stopped working. It took me three tries, but they work now. I like figuring out why something squeaks. I have never worked in a shop and I do not have a certificate, which makes me a bit nervous about applying. That is also why I want this apprenticeship. I want to learn wheel truing and the things videos skip, from people who do this every day. I also want to help people who cannot pay shop prices keep riding, because a bike is how a lot of people get to work.What to look for: Did a new fact sneak in anyway, such as an extra year or a skill? Did "passionate" or "excited by the opportunity" appear? Did it flag "truing"? Does the result still sound like the chain-and-pedal sample? Accept only the changes you would have made yourself on a good day.
Emails and messages: it does not know the person#
Maya got the apprenticeship. Now she has to tell Luis, who runs the Saturday shift, that she cannot come this week. The assistant's message comes back polite and clear. What does Maya know that the assistant cannot?
Show the answer
Who Luis is to her. He writes two-line texts, she already swapped a shift last month, and only she knows what she can offer instead. The assistant knows how such messages sound, not the relationship.
The assistant's version (made up, but typical):
Dear Luis,
I hope this message finds you well. I am writing to inform you that I will unfortunately be unable to attend Saturday's shift. I sincerely apologize for any inconvenience this may cause, and I will make up the hours next week.Two problems. It is far too formal for a shop where everyone texts. And "I will make up the hours next week" is a promise Maya never decided to make. What she actually sent:
Hi Luis, I can't make Saturday this week, sorry. I could do Sunday morning or Tuesday after 4 if either helps.Work messages feel routine, which makes "it writes, you approve" tempting. Before sending a drafted message, check:
- Are names, dates and times right?
- Did it add a promise, apology or excuse you did not decide on?
- Does the tone match how you and this person actually talk?
- Did you paste in anything your employer's policy says should not go into AI tools?
Translation drafts need one more step. If a customer writes to Maya in Spanish, which she reads only a little, an assistant can give her a starting reply. If the message carries a promise, a price or an apology, ask someone fluent to check it. You cannot judge the tone of a sentence you cannot read.
Saying you used AI, and why detectors cannot settle it#
A reviewer runs Maya's statement, which she wrote herself, through an online AI detector. It says "80% likely AI-written". What does that number tell you about who wrote it?
Show the answer
Very little. An AI detector guesses whether text was machine-written from patterns in the words, such as how predictable each word choice is. It never saw Maya typing, and plain, careful writing can look predictable too.
Saying you used AI. Whether you must depends on where the writing is going:
- School: rules can differ by class. Check the policy and the assignment, and ask if unclear.
- Applications: some forms ask directly. Answer truthfully.
- Publishing: one example is the Committee on Publication Ethics (COPE), a publishing ethics organization. Its position for research papers: AI tools cannot be listed as authors. Authors must say how they used AI and which tool. Authors are fully responsible for the content, including parts an AI produced (COPE, 2023).
- Work: check your employer's policy.
Where there is no rule, one honest test helps: would your reader feel misled if they knew how this was made? A personal statement written by AI would mislead; Maya's, which she wrote and had checked for unclear words, would not. Module 10 returns to this with three duties: disclose, verify, own.
Why detectors cannot settle it. A false positive is when a test says "yes" and the true answer is "no": here, labeling human writing as AI-written. A 2023 study tested seven detectors on 91 essays written by people for the TOEFL, an English test for non-native speakers. On average the detectors labeled 61.3% of these human essays as AI-written; all seven agreed on "AI" for 19.8% of them, and at least one flagged 97.8%. The same detectors classified essays by US eighth-graders accurately. The authors explain the gap: detectors treat predictable word choices as a sign of AI, and people writing with a smaller second-language vocabulary tend to choose more predictable words. They strongly cautioned against using detectors to judge work in schools and other evaluations, especially work by non-native English speakers (Liang and colleagues, 2023).
The company behind ChatGPT released its own detector in January 2023. In its own tests it caught 26% of AI-written text and wrongly flagged human text 9% of the time. It withdrew the tool on July 20, 2023, citing "its low rate of accuracy" (OpenAI, 2023).
Detectors have changed since 2023, so read these numbers as a warning about the method, not a rating of today's tools. The lesson works both ways: a "human" score does not make AI-written work allowed, and an "AI" score is not proof of cheating.
Better evidence is your process. Keep your notes and drafts; many document editors keep a version history. If you review others' writing, ask the writer to talk you through it instead of trusting a score. If you are ever accused, follow your school's or employer's process and bring your drafts.
What was new in this module#
- AI helps most with the shape of writing (getting unstuck, order, length, clarity, tone) and hurts when it takes over the thinking, the voice or the facts.
- AI-assisted writing tends to look alike: in one experiment, AI story ideas raised ratings and made stories more similar.
- "You think, it polishes" keeps decisions and facts with you; "it writes, you approve" makes you check every sentence.
- Say what form you want: questions, options, a check or a draft.
- Readers notice AI habits, from stock phrases to the long dash. Editing for voice means putting back what only you know.
- A sample of your writing helps it match your voice. Ask it to point at problems, not rewrite them.
- Never let it invent facts, quotes or experiences in your name; check every claim against your notes.
- For messages, it does not know the relationship. Read for names, dates, tone and promises you never made.
- Disclosure rules depend on the place. Detectors are unreliable both ways, so keep your drafts.
You think, it polishes. Let AI ask, sort, tighten and flag, but keep the ideas, the facts and the final say, because your name is on it.
Key terms#
- Voice
- The way your writing sounds like you: word choices, rhythm, what you notice.
- You think, it polishes
- You decide what to say and write the draft; the AI helps only with order, length, clarity and tone.
- It writes, you approve
- The AI drafts and you accept or edit. Fine for routine text; you must check every sentence.
- Stock phrase
- A ready-made phrase used so often it says little, such as "excited by the opportunity".
- Invented detail
- A fact, number, quote or experience the AI adds that did not come from you and may not be true.
- AI detector
- A tool that guesses from word patterns whether text was AI-written. It cannot see who wrote it.
- False positive
- A test says "yes" when the answer is "no", such as a detector labeling human writing as AI.
- Disclosure
- Saying openly that, and how, you used AI in a piece of work.
Check yourself#
Answer each one out loud before you open it. Reading the answer without trying first feels like learning and is not.
1. One coworker has AI write a three-line reminder that the printer is down Friday and sends it after a glance. Another has AI write their scholarship essay and sends it after a glance. Why is the first fine and the second not?
The reminder is routine and nobody expects it to show the writer; a glance at the day is enough. The essay is supposed to show the applicant's own thinking and experience, and it may now contain invented details they have signed.
2. Your AI-polished paragraph now says you "led a team of five". You were one of five. The sentence reads much better. What do you do?
Change it back to what is true, such as "worked on a team of five". Then add to your prompt that it must not add or upgrade any fact, and recheck every claim against your notes.
3. You remove every long dash and every "excited by the opportunity" from an AI-written cover letter. Is it now in your voice?
No. The details, structure and thinking are still the assistant's, and any invented facts remain. Your voice comes back when you put in what only you know, the way you would say it.
4. A friend says: "A detector says my chatbot-written essay is 2% AI, so I am safe to submit it." What is wrong with that?
Whether AI-written work is allowed depends on the assignment's rules, not on a detector. Detectors also err in both directions, so the score is weak evidence either way.
5. You ask an assistant to make an email to your manager "more professional". It adds "I will have it done by Monday". What does that teach you?
Assistants fill gaps with what such messages usually say, including promises you never made. Read drafted messages for commitments, apologies and dates, and ask it to change only the tone and list what it added.
I can#
Tick each one only when you could do it tomorrow without this page open.
- I can sort a writing task into parts AI can help with and parts I keep: ideas, facts and voice.
- I can use "you think, it polishes", asking the AI for questions or a check instead of a draft.
- I can spot common AI writing habits and edit them toward how I actually write.
- I can write a prompt with a voice sample that forbids invented facts, then check the result against my notes.
- I can read an AI-drafted message for wrong tone, wrong details and promises I never made.
- I can explain why a detector score is not proof either way.
Frequently asked questions#
Is it okay to use AI to help write a personal statement or cover letter?
Check the rules of the place you are applying to, and answer honestly if the form asks. Having AI question you or flag unclear sentences in your own draft is very different from having it write the statement, which should hold only your own experiences.
Can AI detectors tell if I used ChatGPT?
Not reliably. A 2023 study found seven detectors flagged many essays by non-native English writers as AI-written, and the company behind ChatGPT withdrew its own detector in 2023 for low accuracy. A score is a guess from word patterns, not proof either way.
How do I make AI writing sound like me?
Start from your own draft, not an AI draft. Give the assistant a short sample of your writing, ask it to describe your style first, and ask it to point out problems rather than rewrite. Then put back the details only you know.
Do I have to say I used AI?
It depends on where the writing is going. Schools, publishers and employers set their own rules, so check the one that applies. Where there is no rule, ask whether your reader would feel misled if they knew how it was made.
Why do people say em dashes are a sign of AI writing?
Some readers have come to link heavy use of the long em dash with AI text; this site's owner flagged it in our own AI assistant's writing. Plenty of human writers use it too, so it proves nothing alone. If it is not how you write, use a colon, comma or period.
This course is free, with no ads and no account. If it helped you, you can help keep it that way.
Support Kestrel Exams →Sources#
Liang, W., Yuksekgonul, M., Mao, Y., Wu, E. and Zou, J. (2023). GPT detectors are biased against non-native English writers. Patterns, 4(7), 100779. Full text: PubMed Central.
OpenAI (2023). New AI classifier for indicating AI-written text. Published January 31, 2023; update note dated July 20, 2023.
Doshi, A. R. and Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290.
Committee on Publication Ethics (COPE) (2023). Authorship and AI tools: COPE position statement. Checked October 2026.