A short book about writing with AI without losing your voice. It explains why detectors flag people who wrote every word themselves, what "sounds like AI" actually means line by line, and a working method that keeps the writing yours. Free to read here, free to share, no email required.
I run a small company in Navi Mumbai. One of the things I built is an AI detector. People paste text into it all day, and I get to see what happens next.
A student in Pune pastes her essay. It comes back 80% AI. She wrote every word of it herself. She writes me a long email, upset, asking why.
A freelance writer in Lagos pastes a blog post his client rejected. The client ran it through a detector and said "this is ChatGPT". He did use ChatGPT, for the first draft. Then he spent two hours editing. The detector still caught it.
A marketing manager in Manila pastes ten product descriptions. All of them come back flagged. She doesn't care about the score. She cares that they all sound the same and nobody is buying.
These three people have the same problem from different angles. They use AI to write, or they get accused of using it, and they don't understand what is actually going on under the hood. So they either stop using AI (and fall behind), or they keep using it badly (and get caught, or get ignored).
This book is the thing I wish I could send to each of them.
It is short on purpose. You should be able to read it in one sitting, maybe an hour, and then actually use it the same day. There is no theory for the sake of theory. Every chapter ends with something you can do.
What this book is not: it is not a trick to fool your teacher. If you want a way to submit ChatGPT homework and get away with it, this is the wrong book, and honestly the wrong plan. Detectors change every month. Teachers are getting sharper. And the skill you would be skipping, the ability to write clearly in your own voice, is the one thing that keeps paying you for the rest of your life.
What this book is: a plain explanation of how AI writing gets recognised, why real people get flagged too, and a working method for using AI as a helper while the final piece still sounds like you.
Let's go.
People say a piece of text "sounds like ChatGPT" the way they say a song "sounds like the 80s". They know it when they hear it, but they can't tell you why. Once you can name the patterns, you can't unsee them, and you can fix them.
Here are the ones I see most, from reading thousands of flagged samples.
"In today's fast-paced digital world, content is more important than ever."
"Communication is a vital skill in every aspect of life."
These sentences are true and completely empty. Nobody disagrees with them, so nobody remembers them. A human who cares about the topic usually starts with the thing they actually want to say, or with a specific moment. AI starts with a warm-up lap.
AI loves groups of three. "Efficient, effective, and engaging." "Students, teachers, and parents." "It saves time, reduces stress, and improves results." Humans do this too, but not in every paragraph. When every list has exactly three items and every item is the same length, it's a machine.
Read those out loud to a friend and watch their face. Nobody talks like that. Your writing shouldn't either.
AI models use the long dash constantly, sometimes two or three per paragraph, where a person would use a comma, a full stop, or brackets. I'm not saying never use a dash. I'm saying if you didn't type it yourself, remove it.
Every paragraph is four sentences. Every sentence is 15 to 22 words. Every paragraph opens with a topic sentence and closes with a mini-summary. It reads like a textbook that was written by a committee.
Real writing is lumpy. Some paragraphs are one line. Some run long because the writer got excited. Sentences vary. Some are short. Others wander a bit, take a detour, and come back to the point because that's how the thought actually happened in the writer's head.
This is the big one. AI text is careful. It says "many experts believe" instead of "I think". It says "various tools" instead of "I used Notion for six months and gave up". It says "significant improvements" instead of "my score went from 62 to 81".
Humans have opinions. Humans remember specific numbers, names, dates, and the time something went wrong. Humans say "I was wrong about this". That texture is the strongest signal of a real person, and it's the thing AI struggles to fake, because it never lived anything.
"In conclusion, by following these steps, you can..." followed by a restatement of everything above. Humans end when they're done. Sometimes with a joke. Sometimes mid-thought. Rarely with a recap.
"It may be beneficial to consider potentially exploring..." AI is trained to avoid being wrong, so it softens every claim. Write what you mean. If you're not sure, say "I'm not sure, but here's my guess", which is itself very human.
Take any piece of AI text you have lying around. An email draft, a paragraph from an assignment, anything. Go through it with this list and circle every one of these patterns. Most people find eight to fifteen in a single page. Don't fix anything yet. Just see them. The fixing comes in Chapter 4.
If you want a second opinion, paste the same text into the free detector at textsight.ai. It gives you a score and marks the sentences it thinks are machine-written. Compare its marks with your circles. They will overlap more than you expect.
Most people imagine a detector as a search engine that checks if your text exists somewhere. That's a plagiarism checker, and it's a completely different thing. An AI detector has never seen your text before. It's making a guess about how it was produced.
Here is the guess, in plain language.
Large language models write by predicting the next word. At every step they pick the most likely word given everything before it. So the text they produce is, by definition, made of likely words in likely orders.
A detector runs the same kind of prediction in reverse. It reads your text and asks, at every word, "how surprised am I by this word?" The technical name is perplexity. Low perplexity means the text is very predictable. High perplexity means it keeps surprising the model.
Human writing is surprising. We pick odd words. We reference things the model doesn't expect. We make small grammar slips. We put the punchline in a strange place. All of that raises perplexity.
AI writing is smooth. Every word is the expected word. Perplexity is low, and the detector says "this looks machine-made".
The second thing detectors measure is variation, sometimes called burstiness. Humans write in bursts. A long, winding sentence, then a short one. A dense paragraph, then a two-line one. AI tends to keep a steady pace, like a metronome.
Detectors look at the spread of sentence lengths, the spread of paragraph lengths, and how much the predictability score moves from sentence to sentence. Flat lines are suspicious.
How surprised the model is by each word, reading left to right. Taller means more surprising.
AI writingEvery word is the expected word, so the line barely moves. Low perplexity.
Your writingOdd words, detours, a slip here and there. It keeps surprising the model.
How much the sentence lengths vary. Each bar is one sentence.
AI writingA steady pace, like a metronome. Flat lines are suspicious.
Your writingLong, then short, then long again. Real writing is lumpy.
An illustration of the two signals a detector measures. Not a reading of any real text.
The newer detectors go further. They are trained on millions of examples, half human and half AI, and they learn patterns that nobody has a name for. Word choice, punctuation habits, how ideas are ordered, how often certain phrases appear. These models are more accurate than pure perplexity checks, and they are also harder to explain, which is part of why people get frustrated with them.
Now the important part. Every method above is measuring how predictable and how even your writing is. Not who wrote it.
So who writes predictable, even text without using AI?
People who learned English as a second language and were taught to write "correctly". Textbook grammar, safe vocabulary, standard sentence shapes. Their writing is predictable precisely because they worked hard to make it correct.
Students who were taught the five-paragraph essay format. Intro with thesis, three body paragraphs with topic sentences, conclusion that restates. That format is practically a template for low-variation text.
Technical and legal writers, who are supposed to be precise and repetitive.
Anyone writing in a formal register about a common topic.
In 2023, researchers at Stanford tested several popular detectors on essays written by non-native English speakers for the TOEFL exam. More than half of those human-written essays were flagged as AI. The same detectors did fine on essays by native speakers. The problem wasn't the students. It was that the detectors mistake "careful and correct" for "machine".
This is why the student in Pune got flagged. Her essay was genuinely hers. It was also careful, correct, formal, and built on a template. To a detector, that looks like the output of a model that is also careful, correct, and formal.
Two things.
First, a detector score is evidence, not proof. If you get flagged for something you wrote, you are not crazy and you are not alone. Chapter 5 covers what to do about it.
Second, the fix for "sounds like AI" and the fix for "gets flagged as AI" are the same fix. Both come from making your writing more like a specific person and less like an average of all people. More surprise, more variation, more you. That's the whole method in the next two chapters.
Write three or four sentences about what you had for lunch yesterday. Don't think, just write, the way you'd text a friend. Now paste it into a detector. It will almost certainly come back human, and with a low AI score. Now ask ChatGPT to "write a short paragraph about what I had for lunch" and paste that. Compare the two scores. You just saw perplexity and burstiness with your own eyes.
I use AI to write. Every day. Emails, product pages, outlines, even parts of this book were shaped with it. I'm not going to pretend otherwise, and I don't think you should either.
The question was never "AI or no AI". The question is where in the process you put it. Most people put it in the wrong place. They ask AI to write the thing, then try to make it sound human afterwards. That's backwards. It's like asking a stranger to write your wedding speech and then editing in a few jokes. The bones are still someone else's.
Here is the order that works.
Timings are for a 600-word assignment. Under an hour, start to finish.
Before you open ChatGPT, answer three questions on paper or in a plain notes app:
What am I actually trying to say? One sentence. If you can't do it in one sentence, you're not ready to write, and AI won't fix that.
Who is going to read it, and what do they already believe? Your professor, your client, a stranger on Google.
What do I know about this that a generic model doesn't? A number, a story, a mistake you made, a conversation you had, something you noticed. This is your raw material. It's the part that can't be generated.
This takes ten minutes and it's the most important ten minutes. Everything downstream depends on it.
Now open the AI. But don't say "write me an essay about X". Say things like:
"Here's my main argument: [your sentence]. What are the three strongest objections to it?"
"I have these five points. Which order makes the most sense and why?"
"Here's a paragraph I wrote. Where is it unclear?"
"Give me ten possible opening lines for this. Make them short."
"Explain [concept] to me like I'm 15, so I can then explain it in my own words."
Notice what's happening. The AI is helping you think. It's generating options, poking holes, filling gaps in your knowledge. You're still the one deciding what goes on the page.
If you must have a first draft generated, fine, but feed it your Stage 1 material. Paste your one-sentence argument, your audience, your specific stories and numbers. Then say: "Draft this. Plain English. Short sentences. No introduction paragraph. Don't summarise at the end. Use my examples exactly as I gave them." You'll get something usable instead of something generic.
This is the stage everyone skips, and it's why they get flagged.
Take the AI draft. Read it out loud, one paragraph at a time. Then close the window and type what you just read, from memory, the way you would say it to a friend. Don't copy. Retell.
You will lose some polish. Good. You will also lose the balanced triplets, the "delves", the perfect rhythm, and the empty opener, because your mouth doesn't produce those things naturally. What comes out is your sentence structure, your word choices, your rhythm.
This is slower than copy-pasting. Roughly twice as slow. It's still three or four times faster than writing from a blank page, and the result is something you can defend, because you actually wrote it.
Now you go through with the checklist in the next chapter. Add specifics. Cut filler. Break the rhythm. Insert an opinion. This is where a competent piece becomes a piece with a pulse.
Last step, paste it into a detector. Not to "pass". To find the sentences you rushed. A good detector highlights the sections that read most machine-like, and nine times out of ten those are the paragraphs you didn't retell properly in Stage 3. Rewrite only those. Scan again. Done.
I built the free scanner at textsight.ai around exactly this loop. Scan, see which sentences are flagged, fix those, rescan. There's also a humanizer for when you're truly out of time, but I'll say this plainly: the humanizer is a shortcut, and the retelling method in Stage 3 produces better writing every single time.
For a 600-word assignment:
Stage 1, ten minutes. Stage 2, ten minutes. Stage 3, twenty minutes. Stage 4, ten minutes. Stage 5, five minutes. Under an hour, and you end with work that's yours.
Compare that to the two hours the freelancer in Lagos spent editing a draft that was never his to begin with, and still got rejected.
Pick something small you need to write this week. Do Stage 1 right now, before you close this book. Three questions, one sentence each. Keep the note. You'll use it in Chapter 9.
This is the chapter to bookmark. It is a checklist you run over any draft, AI-assisted or not, to make it read like a person wrote it. I've timed it on hundreds of pieces. For a page of text, it takes about twelve minutes once you've done it a few times.
Do the steps in order. Each one fixes a specific detector signal from Chapter 2 and a specific tell from Chapter 1.
Read your opening paragraph. Ask: does the piece still make sense if I remove it? If yes, remove it. Most first paragraphs are a warm-up. The real opening is usually your second paragraph, and it's better.
If you need an opener, start with one of these: a specific moment ("Last Tuesday my client sent me a one-line email"), a number ("Eleven of the fourteen essays were flagged"), a direct claim ("Most AI detectors are measuring the wrong thing"), or a question you actually want answered.
Same test. If your closing paragraph starts with "In conclusion", "Overall", "To sum up", or "By following these steps", it goes. End on your strongest point, or a single line that lands.
Search for each of these and replace or delete: delve, tapestry, landscape, testament, underscore, pivotal, multifaceted, navigate, robust, seamless, leverage, foster, realm, embark, unlock, elevate, harness, crucial, vital, essential, comprehensive, in today's, ever-evolving, game-changer, cutting-edge.
Replace with the plain word. "Crucial" becomes "important" or, better, gets cut because the sentence already shows it's important. "Leverage" becomes "use". "Navigate the landscape" becomes "deal with".
Find every list of three. Make one of them a list of two, or four, or turn one item into its own sentence. Not all of them. Just enough that the pattern breaks.
Find your longest sentence in each paragraph. Cut it in two. Find two short sentences next to each other and join them. Then add one very short sentence somewhere. Three words. Like this.
Read the paragraph out loud. If it sounds like a metronome, keep going until it doesn't.
This is the step that matters most. For every paragraph, add one detail only you could know. A number. A name. A place. A date. A brand. A thing that went wrong. Something you saw, heard, or measured.
"Many students struggle with time management" becomes "In my hostel, the guy in the next room used to start his assignments at 11pm the night before, every time."
"Various tools can help" becomes "I tried Notion, then Trello, then went back to a paper diary."
If you can't think of a specific for a paragraph, that paragraph might not need to exist.
Somewhere in the piece, say what you think. "I think this is overrated." "Honestly, the second method never worked for me." "This part is boring but you need it." AI avoids taking sides. You don't have to.
Replace long dashes with commas, full stops, or brackets. Replace "may be beneficial to consider" with "should". Replace "it's important to note that" with nothing. Just say the thing.
Here's a paragraph straight from an AI draft:
"In today's competitive job market, having a strong resume is crucial. A well-crafted resume can help you stand out, showcase your skills, and land interviews. It is important to note that recruiters often spend only a few seconds on each resume, so it is essential to make a strong first impression."
After the twelve-minute edit:
"Recruiters spend about seven seconds on a resume. I know because I sat next to one at a hiring event in Bangalore and timed her. Seven seconds is enough to read your name, your last job title, and maybe one line. So that one line had better be good. Mine used to say 'responsible for various projects'. It now says 'shipped a payments feature used by 40,000 merchants'. Same job, different resume."
Same message. One is furniture, the other is a person. Paste both into a detector if you want to see the score difference, but you don't need a tool to feel it.
Take the AI text you marked up in Chapter 1. Run all eight steps. Set a timer for fifteen minutes since it's your first go. When you're done, read the before and after side by side. Then scan both. Most people find they like the second version better. That matters more than the score.
If you're a student, you're dealing with two separate problems and it helps to keep them apart.
Problem one: you want to use AI to study and write better, without getting into trouble.
Problem two: you might get accused of using AI when you didn't, or when you used it in a way you thought was allowed.
Every college, and often every professor, has a different policy. Some ban AI completely. Some allow it for brainstorming but not drafting. Some allow anything as long as you disclose it. Some haven't decided and are making it up as they go.
Don't guess. Find the policy in your syllabus or course handbook. If it isn't written down, email the professor and ask. Keep the reply. A one-line email saying "Is it okay to use ChatGPT to help outline my essay?" and a one-line "Yes, but write the essay yourself" is the best insurance you will ever have.
Whatever you write, write it in a tool that keeps history. Google Docs has version history built in (File, then Version history). Microsoft Word does the same if the file is on OneDrive. This is not optional anymore.
Version history shows the essay growing over time, with the deletions, the rewrites, the paragraph you moved. An essay pasted in as one block at 2am has no history. If you're ever accused, the first thing you'll be asked for is your drafts. Have them.
Also keep your Stage 1 notes from Chapter 3, your outline, and any chat logs with the AI if you used it for brainstorming. A folder per assignment. Takes thirty seconds.
If your course allows AI with disclosure, do it properly. A line at the end like: "I used ChatGPT to generate an outline and to check my grammar in paragraphs 2 and 4. All arguments, examples, and writing are my own." Specific, honest, boring. That's what a professor wants to read.
Some courses want a citation format instead. APA and MLA both have guidance now for citing generative AI. Your library website will have the exact format.
Your university may run your essay through Turnitin or a similar tool. You can't control their tool or their threshold. What you can control is whether your essay has the tells from Chapter 1.
So scan it yourself first, not to game the number but to find the paragraphs that sound like a template. Most student essays that get wrongly flagged have the same shape: five paragraphs, topic sentences, "In conclusion". The fix is the twelve-minute edit. Add your own examples, break the format a little, take a position.
The free scanner at textsight.ai is enough for this. Paste, look at which sentences it marks, fix those. Do it once for each assignment and it becomes a habit.
Stay calm. This happens to honest students constantly, especially those writing in English as a second language. The detector is a tool with a known error rate, not a judge.
Do these in order:
Ask, in writing, which tool was used and what score it gave. You have a right to know what you're accused of.
Bring your version history, your notes, your outline, and any email where the professor approved AI for outlining or grammar.
Offer to discuss the essay in person. Someone who wrote the thing can talk about it for twenty minutes without notes. Someone who pasted it can't. This is the single most convincing thing you can do.
Point to the published research on false positives, politely. The 2023 Stanford study on non-native writers is a good one to mention. Most academic integrity offices are aware of it now.
Don't say "the detector is wrong" and stop there. Say "I understand why it flagged this, here's the proof it's mine".
One more thing. The best use of AI as a student has nothing to do with writing the assignment. It's asking it to explain the thing you didn't understand in the lecture, three different ways, until one clicks. It's asking it to quiz you. It's asking "what would a tough examiner ask about this topic?" None of that ever touches your submitted work, and all of it makes you better.
Clients are running detectors now. Not all of them, but enough that you should assume they might. Here's how to work with that instead of against it.
Have an answer ready for "do you use AI?" that is honest and confident. Mine, when I hire writers, is what I want to hear: "Yes, for research and first outlines. Every sentence you get is written or rewritten by me, and I scan my own work before I send it." That answer wins more work than "no, never" because nobody believes "no, never" anymore.
Put it in your proposal or onboarding doc so it's never a surprise.
When you send a draft, include a screenshot or a line: "Scanned with TextSight, 8% AI, mostly in the product-spec paragraph you asked me to keep technical." You've just answered the question before they asked, you've shown you're careful, and you've explained the one section that might read mechanical. Clients love this. It costs you thirty seconds.
Most freelancers who get rejected for "sounds like AI" made the same mistake. They generated a draft, edited it, and sent it. The bones were still the model's. Go back to Chapter 3. Generate, then retell in your own words, then edit. Your writing will have a voice, and voice is the only thing a client can't get from ChatGPT for free.
If a client wants "just AI content, cheap", let them go, or price it as what it is: generation plus a light edit, at a low rate, with no guarantee it reads human. Your real product is the retelling and the specifics. The interview with their customer. The number from their dashboard. The thing you noticed on their competitor's site. Charge for that.
Writers who produce a lot of content develop habits that look machine-like even without AI. Same opener every article. Same structure. Same closing CTA. Same three-item lists. Vary it on purpose. A quick scan on a random sample of your old work is a good way to see your own patterns.
I'm a Marathi speaker. I learned English at school, then properly at work, and I still catch myself writing "do the needful" and "revert back". So this chapter is personal.
Non-native writers get flagged more. Not because their writing is bad, but because it's careful. You were taught to use the safe word, the correct structure, the standard phrase. Every one of those choices makes your writing more predictable, and predictable is what detectors are built to catch.
Here's the frustrating part: the "correct" English you worked hard for is now working against you. And here's the freeing part: the fix is to relax it, which is easier than tightening it.
Your spoken English has personality. It has your accent in the rhythm, your regional phrases, the way you order ideas. When you write, you strip all of that out to be safe. Stop stripping. Let a bit of how you actually talk stay on the page.
If you'd say "this method is good but slow, honestly", write exactly that. Don't upgrade it to "this methodology, while effective, presents certain temporal limitations". The first one is human. The second is what a detector and a bored reader both hate.
Grammarly, QuillBot and the AI "improve my writing" buttons push everything toward the same smooth, average English. Use them to catch actual errors, the wrong tense, a missing article. Don't accept every suggestion. Every accepted "rephrase for clarity" makes your writing sound a little more like everyone else's, and a little more like a machine.
A trick: run the grammar check, then read your text out loud. Anywhere it no longer sounds like you, put your version back, even if the tool underlines it.
A slightly odd phrase, a missing comma, a word order that's a bit Indian or Filipino or Nigerian, these are not failures. They're fingerprints. Detectors read them as human, and most readers don't mind them at all as long as the meaning is clear. Perfect is not the target. Clear and yours is the target.
The strongest thing you can put in any piece of writing is a detail from your own world. A price in rupees. A street name. A festival. The way your college canteen works. A generic model trained mostly on American internet text doesn't produce those, so they read as strongly human. Don't translate your life into a generic one to sound "international". The local detail is the interesting part.
If you were taught the five-paragraph essay, the "firstly, secondly, thirdly, in conclusion" format, know that this format is close to what a detector expects from a machine. You don't have to throw it away, but loosen it. Start with an example instead of a thesis. Let one body section be twice as long as the others because it matters more. Skip the conclusion or make it one sentence.
Write a paragraph about your hometown, in English, without stopping to fix anything. Then write the same paragraph the "proper" way you'd write for an exam. Scan both. The first one will score more human, nearly every time. That's the gap you're closing.
You don't need many. Here's my working kit, what each one is for, and where it goes wrong.
Google Docs, free. Every draft lives here from the first line, so version history exists automatically. Word on OneDrive does the same. Never write straight into the chat window or a form field.
ChatGPT, Claude, Gemini, whichever you like. Free tiers are enough. Use it the way Chapter 3 describes: objections, outlines, options, explanations. The moment you notice you're copy-pasting whole paragraphs, stop and go back to retelling.
Grammarly free, or the built-in checker in Docs. Fix real errors. Reject rephrases. Turn off the "tone" and "clarity" suggestions entirely if you can.
Hemingway Editor, free on the web. Paste your text and it highlights long sentences, adverbs, and passive voice. Don't chase a grade. Use it to spot the two or three sentences that need breaking up.
This is where TextSight comes in, and yes, I built it, so read this part with that in mind.
The free tier at textsight.ai needs no card. The current limits are on the pricing page, and they change, so I won't quote a number here that goes stale. Paste your text, and it returns an overall score plus sentence-level highlighting so you can see exactly where it thinks the machine is. That's the feature I care about. The overall number is a headline; the highlighted sentences are the to-do list.
There's also a humanizer that rewrites flagged text to read more naturally. I'll be honest about it: it's for when you're out of time and the text is already yours in substance. It will not turn a fully generated essay into something you can defend, and it shouldn't. The retelling method always produces better writing than any rewriting tool, including mine.
Other detectors exist. GPTZero, Originality, Copyleaks, ZeroGPT. They'll give you different numbers for the same text, sometimes wildly different, because they weigh the signals from Chapter 2 differently. Don't panic about that. Use one consistently, look at the sentences it flags rather than the score, and fix those.
Paraphrasing spinners that swap every third word for a synonym. They produce text that reads worse than the AI draft and often scores worse too, because now it's predictable and awkward. Any tool that promises detectors will never catch you is selling you a race you can't win. The detectors update, the trick stops working, and you've learned nothing.
Let's do one properly. Say you're writing a 400-word blog post for a small bakery's website about why they switched to sourdough. Here's the whole method in one go.
What am I saying? "We switched to sourdough because customers kept asking, and it turned out to be better for the business too."
Who's reading? Locals searching for bread near them, and regulars who want to know what changed.
What do I know that a model doesn't? The owner told me the old bread had a 15% waste rate. Sourdough keeps three days, so waste dropped to under 5%. The first batch was a disaster, too sour, and a regular named Mrs Fernandes told them so. They now do a Saturday sourdough class that sells out.
I paste those notes and ask: "Give me five possible opening lines using the Mrs Fernandes detail." I pick one. I ask: "What would a reader want to know that I haven't covered?" It says price and whether the old bread is still available. Good, I'll add both.
If I just asked the AI to "write a blog post about switching to sourdough", I'd get this:
"In today's health-conscious world, more and more consumers are seeking out artisanal bread options. Sourdough, with its rich history and numerous health benefits, has become a popular choice. At our bakery, we are proud to announce our transition to sourdough. This decision was driven by customer demand, a commitment to quality, and a desire to reduce waste. Our sourdough is crafted with care, using traditional methods and natural fermentation. Whether you're a longtime customer or new to our bakery, we invite you to experience the difference."
Every tell from Chapter 1 is in there. Empty opener. Triplet. "Crafted with care." No specifics. Closing invite. Any detector will flag it, and more importantly, no one will remember it.
"Mrs Fernandes told us the first batch was too sour. She was right.
That was in March. We'd switched to sourdough because about a dozen regulars had asked for it over the winter, and because our old sandwich loaf was going stale by the second day. We were throwing out roughly 15% of what we baked. It hurt.
Sourdough keeps three days. Waste is now under 5%. That alone paid for the new proofing baskets in six weeks.
The recipe took four tries. Batch one was sour enough to make Mrs Fernandes wince. Batch two didn't rise. By batch four we had the loaf we sell now: a 900g boule, crust that cracks when you press it, and a crumb that's open but not full of holes. It's ₹180. The old loaf was ₹120 and we still bake it on Tuesdays and Fridays for the people who prefer it.
The surprise was the class. We ran a Saturday morning sourdough workshop once, as a favour to a customer who kept asking how we do it. Twelve seats. It sold out in two days, and it's sold out every month since. Turns out people don't just want the bread. They want to know how it works.
Come by on a Saturday and taste the difference. And if it's too sour, tell us. We'll listen. We did last time."
Around 260 words. Every paragraph has a specific. The lengths vary. There's an opinion ("It hurt"), a small story, a price, a name. Scan it and it reads human, because it is. It took about forty minutes including the notes.
Not the tool. Not clever prompting. The notes from Stage 1. Everything good in the final version was in those five lines of notes before the AI was ever opened. The AI helped with the order and the opener. The person supplied everything that mattered.
That's the whole book in one example.
Reading about this changes nothing. Doing it once changes a little. Doing it for a week makes it a habit. Here's the week.
Take three things you wrote recently. One with AI, one without, one you're not sure about. Scan all three at textsight.ai. Look at the highlighted sentences, not the score. Notice which of your own habits show up. Write them down. That list is your personal version of Chapter 1.
Pick the most flagged piece from Day 1. Run all eight steps from Chapter 4. Time it. Scan it again. Keep the before and after.
Take something you need to write anyway. Do Stage 1 notes. Get an AI draft. Close it. Retell it in your own words. Notice how different it feels to type your own sentences from memory versus editing someone else's.
Move all your writing into Google Docs or Word on OneDrive. Make a folder per project or assignment. Check that version history is on. If you're a student, find and save your course's AI policy today.
A paragraph, an email, a note. No AI, no grammar check, no scan. Just you. Then scan it afterwards, just to see. Most people find their raw writing scores more human than their polished writing, and that tells you something about polish.
Students: the line you'd put at the end of an assignment. Freelancers: the line for your proposal. One or two sentences, honest and specific. Save it somewhere you can copy from.
Explain the perplexity idea from Chapter 2 to a friend, in your own words, in under two minutes. If you can do that, you've understood it, and you'll never look at a detector score the same way again.
The tools will keep changing. The models will get better at sounding human, the detectors will get better at catching them, and every few months someone will announce that one side has won. Ignore all of it.
The thing that doesn't change is this: writing that comes from a specific person, with specific knowledge and a point of view, reads as human because it is human. No tool produces that and no tool can take it away. Everything in this book is just a way to make sure that's what ends up on the page.
If you found this useful, two small requests.
First, scan something at textsight.ai. It's free, and seeing your own writing through a detector's eyes once is worth more than anything I can tell you. If you sign up, you'll also get an email from me once in a while with whatever I've learned since writing this, including updates when the detectors change how they work.
Second, send this book to one person who needs it. There is a second book too, on the citations AI invents and how to check them, if that is the part of the problem you keep running into. And if you write English as a second language, Flagged in Translation goes into why careful, translated English gets flagged so often. A classmate who got flagged. A friend who freelances. Anyone who's been told their writing "sounds like AI" and didn't know what to do about it.
Thanks for reading. Go write something that sounds like you.
Dipak Bhosale
Founder, TextSight
dipak@textsight.ai
You are welcome to share this book with anyone who might find it useful, as long as you share it whole and unchanged and do not sell it. Teachers, writing centres and student groups: please link or hand it out freely, no permission needed.
Copyright © 2026 Dipak Bhosale. Published by Lacewing Technologies, Navi Mumbai, India.
This book is for general information and is not legal advice. Check your own institution's or client's policy on AI use before relying on anything here. TextSight is an AI-detection and writing-trust tool. No detector, ours included, can prove that you did or did not use AI, because every score is a probability and can be wrong. The strongest evidence is always your writing process: version history, drafts, and notes.
Chapter 1 asks you to scan something you wrote and compare what the detector marks against what you circled. That exercise is worth more than the rest of the book. The sentence-level highlights show you where the AI signal actually sits.
