- AI
- A
Everything is generated by GPT! A guide on how to recognize AI text and how to make it indistinguishable from human writing.
Everyone has already joked and remembered that if the text contains — , it was written by ChatGPT. And if it doesn't, then a human wrote it?
This article is the most detailed guide on the Russian internet on how to distinguish Gen AI–generated text from human-written text, and how to use GenAI yourself to write more human-like text. I will break down real techniques, markers, mistakes, and equip you with valuable knowledge.
Subscribe to the channel https://t.me/meetdeadlines!
Theory
How LLM "thinks" in reality
It’s a crafty T9, it predicts the next words rather than understanding meaning
Imagine you’re talking with an extremely observant interlocutor, Artem (as Artems often are). He never tries to read your mind—he just carefully listens to your every word. He hears the start of your phrase and, thanks to years of conversations, can guess exactly what you’ll say next. He doesn’t care what you’re actually thinking.
The model recognizes word importance in its own way
While you’re speaking, this Artem is only noting the words that most often shape your train of thought—not every word. It’s not guaranteed that they're the words you would choose! Sometimes he latches onto what seems unimportant to you, and other times he might miss real subtext—because his view is shaped by thousands of conversations with different people.
The model doesn’t retrain itself on your task, for it any prompt is just a hint
If you ask Artem to cook paella, he’ll make it the way he once saw on some travel show, and even if it doesn't taste right, he’ll explain why that’s just fine. Whatever you say is just a hint for him about what answer is expected, but Artem never actually learns anything new or changes his habits.
The task you give isn’t truly understood, nor remembered
You discussed something important to you, but Artem has ADHD and ten minutes later has completely forgotten what was talked about at the start. His memory is phenomenal only for short snippets and things he once studied, but if you chat with him long enough, the details from the first minutes slip away.
When you set a role, it only mimics, never lives through it
Artem is, by default, a know-it-all. He thinks of himself as an expert in everything, though in reality he may only have a patchy understanding as he likes to study everything in bits and pieces. This guy instantly takes on any role you give him, but has no real, lived experience. He can’t share real-life examples, but he can spin up something that sounds plausible without trouble.
Ask him to act as a doctor, a blogger, or even a wizard—he’ll immediately get into character and speak the right lingo. But don’t be fooled: he’s never actually treated patients, run a blog, or performed miracles—he just mimics what he’s seen or read.
You’re understood from the first time, and nothing is carried over afterward
You can tell him about yourself and your preferences a thousand times, but he’ll never draw conclusions for the future. Every time is a blank slate, with no memory of previous conversations—only here and now. In this sense, he’s always keen but always a new companion.
The model’s imagination isn’t about creativity, but weirdness
When it seems like Artyom suddenly starts joking or coming up with unusual ideas, it’s not an epiphany or genius. At that moment, he just decided to play around and say the very first things that come to his mind, picking increasingly weird options to surprise you with something unconventional. Sometimes it’s great and makes you smile, and sometimes you wonder what kind of nonsense he just blurted out.
In short – nobody taught the model to think, it was taught to guess
The way LLMs work is well illustrated by the old game Akinator https://ru.akinator.com/. If you didn’t play it as a child, try it now if you’re curious—after reading, the core concept will click!
Bonus
If you study this terminology, you’ll be the smartest in the chat when discussions about artificial intelligence and how ChatGPT is replacing everyone begin: token, tokenization, next token prediction, transformer, model weights, attention, attention heads, context window, data corpus, agent mode, short-term memory, user memory, temperature+top-k/top-p sampling, statelessness, reasoning, finetuning.
There is a huge difference between how we think, speak, and write
When we think, our thoughts appear in fragments, can intertwine, and sometimes even argue with each other – it's a vivid internal dialogue.
When we begin to speak, we often jump around, add emotion, pause, stumble, and the meaning gets constructed on the fly, sometimes changing in the process.
In writing, we try to collect these bits into a more coherent and understandable text. We remove some things, add others, organize, and search for the right words. We compile.
Written speech brings structure, but preserves association and the urge to show depth of thought, add shades of meaning, examples, emotions. As a result, human texts are rarely perfectly smooth—you can trip over them—but they always reflect the path that thought took, with twists and turns and the prism of a personal worldview or the aftertaste of a delicious cherry pie still influencing the way things are told.
Why it’s so hard to imitate human speech
Human thinking is like a shaman’s vision: we hold many semantic threads in our minds, can feel emotions, memories, unrelated images all at once, and loop back to the start of an idea. LLM architecture operates only with vectors, where the entire text is just a sequence of numbers split into chunks. The attention mechanism made a breakthrough possible, but it limits things: the model doesn’t know what experience, taste, or multi-layered memory are—it can’t really link different layers of a conversation or feel depth.
The model learns text as a matrix of tokens, and inside each one there is no actual knowledge, only the probability of encountering one piece of text next to another. For it, anything outside that sequence loses meaning. A human makes associations, remembers, builds new links; the neural network just moves from one token to the next, estimating probabilities, almost never looking back. But there are those who see it differently https://pmc.ncbi.nlm.nih.gov/articles/PMC10068812/
Practice
Theory is, of course, wonderful, but let's actually get a feel for it in practice. Below is a simple quiz in the form of "guess which text was written by a human." In each section you will find two short texts—try to figure out which was written by a human and which by AI.
14 Key Signs That Distinguish AI Text from Human Text
1. Logic of Argument Development
1) A person can't do without water, because it's the main element of metabolism and you just can't go without it.
2) Drinking water is good for your health because water is important for all bodily processes.
Explanation
In human speech, thoughts are constructed sequentially, but can develop unpredictably, sometimes jumping or leaving things unsaid. We often think in loops. Models often write too straightforwardly, without live twists or real mental leaps.
2. Distribution of Emphasis
3) My breakfast consists of oatmeal, tea, and a banana. Oatmeal is very healthy, and tea is a good pick-me-up in the morning.
4) I prefer to make a smoothie with kefir, oatmeal, and banana for breakfast. It's not only fast and healthy, but also fun.
Explanation
In speech, a person easily shifts focus: might start with one thing, then get distracted by a detail, and then return. A model more often writes evenly and doesn’t emphasize anything in particular or make vivid digressions. Another important factor is that people distribute emphasis by sentence or phrase block, while LLMs do so throughout the whole text.
3. Maintaining Genre
5) Don't worry, you'll do great, I know it! You're really a master of your craft, and if the employer isn't blind, they'll definitely hire you!
6) Good luck at the interview, you'll do fine, just be confident and prepared. Remember, an interview is like a date, and you're great at those!
Explanation
Even in a short text, a person intuitively chooses a style and tone—jokes, ponders, writes like an anecdote or a recipe, tries to convey a specific feeling, while a model often delivers something generic, fit for anyone. A real text always breathes with the mood of the chosen genre and is bound to not appeal to everyone.
4. Personal Lens and Experience, Subjectivity
7) My favorite season is summer, because it’s warm and you can relax outdoors, especially nice to lie on the grass in the park.
8) I love summer because you can wear fewer layers and eat tons of fruit, I’ve especially liked sour cherries since childhood.
Explanation
A person inevitably filters any story through themselves: inserts personal memories and associations, even if the text is very short. A model writes impersonally, hardly ever sharing its own, sticking to generic phrases. People also love to state life positions and points of view, even in dry scientific articles, and to look for those like themselves.
5. Intuitiveness
9) I often wake up later than I should and get ready for English class in a rush, then dash for breakfast and log on for the first work meetings.
10) In the morning I wake up, wash up, have breakfast and get ready for work to start a productive day.
Explanation
A person often writes without thinking through every word; unexpected turns of phrase and comparisons surface, and sometimes even the idea itself has no clue where it will go next. The model, however, tends to construct text consecutively and predictively, not allowing for uncontrolled dispersion or sudden intuitive twists.
6. Rhythm and Symmetry
11) I love wandering in the park. There are lots of trees, the air is cleaner, you even somehow feel more alive. Walking in the park is good for your health and mind.
12) I like walking in the park. There are many trees, fresh air, and beautiful paths. Walking in the park is good for your health and mood.
Explanation
Even though a human's text may sound disjointed, in reality it's more wave-like: there are short and long phrases, pauses, repetitions, sudden breaks, all layered over a single leitmotif. The text follows a unified rhythm, but there's lots of local and unmeasured arrhythmicity. AI usually produces even, single-step rhythm—every sentence is similar in length and structure.
7) Emotional Dynamics
13) Woke up—and suddenly there's this strange jitter from a raindrop on the window. Damn, sometimes you just want to stop time, just watch it drip, watch the whole city tremble a little from the drizzle. And it seems like nothing special, but inside, everything suddenly flips over. It doesn't matter how much you have to do, doesn't matter that you're late—there's only you and this world, wet and just a little bit new.
14) I woke up in the morning and saw a raindrop on the window. It was an interesting moment, because rain makes the city cooler. I enjoyed watching it, and I felt calm. After that, I started with my chores and got ready for the new day.
Explanation
Human text is full of emotions; the mood can change right in the middle of a sentence. A bit of light happiness, then suddenly anxiety, then unexpected peace or a burst of inspiration. Emotional dynamics is when feelings in the text are alive and on the move, flowing across the spectrum. The model usually writes evenly, without real emotional swings. Yup, that's how much we love them—we're neurotransmitter junkies.
8) Speech Habits
15) Ah, how I love texting with friends in the evenings; we talk about all sorts of gossip and support each other through tough times, and we regularly send each other memes with cats as well.
16) Well, I do like chatting with close friends in the evenings. Even if we only exchange a couple of phrases, we still support each other in tough times. And we have our reelsrelationship too.
Explanation
Everyone has their own filler words, favorite turns of phrase, unconscious repetitions, specific intonation—even weird constructions in sentences. This comes through in how we start phrases, get sidetracked, jump to another thought, or often insert “well,” “basically,” “like,” and other signature quirks. The model has no such habits; it writes smoothly, evenly, and by the book. This is especially noticeable if you've read a person's texts before—they always give you the feeling of “that's just how they are.”
9. Synonym Space and Descriptive Elements
17) My favorite drink is coffee, because it energizes and helps me wake up in the morning.
18) Coffee is an amazing drink; without it, I wouldn't be able to get up in the mornings.
Explanation
People often switch up their words, play with descriptions, throw in unexpected comparisons and images – even in short speeches. The model, however, uses the same formulations, hardly varies vocabulary or descriptions. And, in an attempt to sound convincing, it uses closely related and similarly colored words without any double meanings, which can only be interpreted literally.
10. Humor, irony, provocations, insults
19) Well, damn, you forgot again, huh? I was waiting all morning, thinking: okay, they're finally growing up, this time they’ll give me a proper greeting – but nope, silence. I almost texted you to say that if you're suddenly in the morgue, just let me know somehow, because I’m sitting here lighting a candle for your memory and it’s not even burning. Fine, next year I’ll just sign a card for myself and you can add your name a week later, like always.
20) Happy birthday to me, since no one aside from a couple of truly real friends remembered. You’re actually the only person who can forget to congratulate me and not feel guilty about it, but I love you anyway, thanks, you fool. Waiting for your call whenever the microcosm in your head gets the calendar update.
Explanation
People’s humor and irony often pop up spontaneously, sometimes right on the edge, with a hint, sarcasm, internal playfulness, or even a little provocation. You could actually take offense at something a person writes. The model, on the other hand, jokes too cautiously, explains jokes, or tries not to hurt anyone, and just skips over sharp moments.
11. Approach to compiling text
21) As usual, I’m overloaded and don’t have time to write anything meaningful so until the end of the week I’ll just be replying with memes. There was a request for something personal above. But I’m hardly a role model in some respects. For example, I’ve never been on vacation in my life and I’ve worked officially since I was 14, unofficially since 11 doing little odd jobs. Jokes about “jack-of-all-trades” don’t make me laugh; I’ve been a promoter, hostess, waiter, cook, café administrator, loader, gardener, forester, lab technician at a clinical diagnostic lab, support specialist, call center employee, salesperson, account manager, Android developer – and I ended up as a PM.
22) I’m very busy right now, so until the end of the week I’ll mainly be answering with memes. There was an earlier request to share something personal. I don’t really see myself as a role model, since my background is pretty unusual: I’ve never been on vacation and I started working officially at fourteen, plus I did small jobs from the age of eleven. I’ve tried a bunch of roles: promoter, hostess, waiter, cook, café administrator, loader, gardener, lab technician, support staff, call center worker, sales, account manager, Android developer, and finally landed in project management.
Explanation
For a person, a long text usually comes together slowly, because some thoughts come at once, some are added later, something is inserted from old notes, and some paragraphs are simply taken from sources. A lot changes on the fly—gets cut out, rearranged, old ideas suddenly resurface, sentences jump from place to place, sometimes pieces remain with a different style or even mistakes. As a result, you get a lively, not always perfectly smooth text, where it’s easy to spot different layers and other people’s intonations. Layers of editing are especially noticeable.
12. Critical Notes (Self-doubt) and Openness to Failure
23) My approach always gives good results, and I’m sure I’m doing everything right.
24) My approach is always correct, I can’t be wrong about anything.
Explanation
There’s always a little worm of doubt inside a person, making you ask yourself a thousand times: what if I’m doing everything wrong, what if I made a mistake and people see it, and then I’ll be ashamed for the rest of my life. We often prepare an excuse in advance in case of failure, justify ourselves even to ourselves, can write and erase, start all over or even be afraid to send a message, because it seems like you’ll say something stupid and everyone will notice. We always doubt ourselves, but appear all the more confident the more we doubt (especially true for narcissists).
13. Grammatical Cleanliness
25) For the most part, I work exactly as much as stated in the contract. Without trickery, without fanaticism—eight hours a day, I’m not 20 anymore after all, and there are things in life besides work. Except, of course, for those rare situations when “everything is totally on fire *” or when I personally promised something to C-level.
26) For the most part, I work exactly as much as stated in the contract. Without trickery, without fanaticism—eight hours a day, I’m not 20 anymore, after all, and there are things in life besides work. Except, of course, for those rare situations when everything is totally on fire * or when I personally promised something to C-level.
Explanation
When a person writes an article or a post, especially in a blog or on social media, you often find imperfect turns of phrase, typos, forgotten commas, words in the wrong order, accidental tautologies or unfinished thoughts. There might be odd sentences that lose part of their meaning when editing, but the author missed them out of fatigue or haste. Sometimes sentences just fall apart or look awkward. Often, the author thinks faster than types, or even wrote the text in parts at different times.
With models, it’s the opposite: the neural network almost always writes smoothly, carefully, without mistakes, with perfectly placed punctuation, and even if it does make a mistake, it does so respectably.
14. Rhetoric of Persuasion, Expertise, and Authorship
27) How to choose a computer chair so your back says thanks: an honest review from a team lead of a certain age, no ads
28) The best tips for choosing a computer chair: experience, screw ups, and real non-paid reviews
Explanation
When someone wants to persuade or present themselves as an expert, they use personal experience, a confident tone, references to their own mistakes, real-life examples, address the reader directly or even argue with them. Authors like to play with the audience. These texts have a distinct personal touch, often featuring a call to action and personal digressions in the style of “if you want to be like me, trust me.”
AI, on the other hand, usually writes dryly, too carefully, restrained, without any personal flair. There’s no personal pressure, no risk, no heated arguments; everything follows academic-style discussion and rhetoric.
Obvious Artifacts
Let’s look at the most common artifacts models often give themselves away with:
Quotation marks – this has been controversial for a few years already, since it’s often easier for us to use “these kinds of quotation marks” rather than these ones « », with spaces inside words for easier perception, but that’s something to keep in mind as well;
Word repetition – sometimes you’ll get 5-10-15 instances of “However” in a single response, which is a classic stylistic translation mistake. HOWEVER :D modern models already perform much better than previous generations. Still, keep an eye out;
Latin in Cyrillic – models sometimes (especially with an overloaded chat context) start outputting English artifacts like “Privet” or “Hello” instead of “Привет”, creating such mistakes. Usually these are text recognition or encoding problems and need “manual review”;
Double dashes — models often use double dashes (—) instead of the regular dash (-) we always use to save time (personally, I don’t even know how to type it on my keyboard);
Bracketed tags/quotes “” () – almost always the information in those can be distorted or look stylized for the whole text, even if it’s a quote or something borrowed/non-original. A classic attempt by the model to present everything in the same tone and emotion.
There’s also an interesting article on common English phrases used by LLMs: https://www.grammarly.com/blog/ai/common-ai-words/.
Differences in LLM Approaches
Each model has its own quirks, which show up in text style, as well as typical expressions, clichés, and approaches to exposition.
ChatGPT is usually polite, overly careful, and tries very hard to avoid any sharp or ambiguous statements. Its responses are often as universal as possible, and it tends toward detailed, structured explanations and recurring templates such as as I mentioned above or it’s worth noting
Grok on the contrary, tries to keep things more informal and provocative, easily using humor and slang, sometimes going out of its way to sound like “one of the guys” for the audience. Its signature style is lively, direct phrases, slightly cocky expressions, and attempts at irony where other models stay neutral.
Gemini usually tries to maintain a balance between clarity, informativeness, and brevity. It responds concisely and without unnecessary filler, but sometimes this makes the text too dry. Gemini is more likely than other models to use a technical style and likes to show brief and clear answer structures, avoiding emotional coloring.
Claude stands out by being prone to lengthy, thoughtful, and slightly philosophical reflections, sometimes filling the text with excessive lyricism. It often provides soft and somewhat vague wordings, avoiding hard statements and striving to be as tolerant and correct as possible, which sometimes leads to a lack of specificity and clarity in the answer.
As a result, each model leaves its own trace in the text, and if you learn to distinguish these signature traits, you can easily guess which AI writer you are dealing with. But this article doesn't have enough space to cover all the features, so maybe in the next
Conclusion
Checklist for a quick suspicious text audit
There are no typos, missing commas, or other small errors in the text
All sentences are about the same length and structured similarly
The text is neutral, without personal emotions or subjective evaluations
The author does not doubt, hesitate, or correct themselves
There are no unexpected jumps to another thought or topic in the middle of the text
Repeated words, turns of phrase, and constructions are used
Unusual symbols or quotation marks (« », “ ”, —) appear instead of familiar ones
The feeling that the text was written by a human
Long text
If you ask your colleagues or fellow project managers whether they write a charter, 50%-60% will answer "no, why?". You involuntarily wonder what this artifact exists for if half of them don't use it in practice.
Let’s recall what PMBOK teaches:
The project charter is a document issued by the project initiator or sponsor that formally authorizes the existence of a project and provides the project manager with the authority to use organizational resources for project operations.
That is, the charter is a document that confirms the existence of the project and gives the project manager rights. But here many people have a dissonance—why is it needed? I have a job description, I have a contract, it says what to do. Let’s do without unnecessary bureaucracy!
Also, the charter is preferably written not by the project manager, but by the sponsor with the client. The second common objection: nobody is going to write it, it’s too complicated.
So let's figure out why this artifact is actually needed.
The project charter is the essence of the project itself. As Ivan Selikhovkin says, if you imagine the project as an egg, this is its shell.
The charter defines what exactly we mean by the project. Will it be clear from the contract + terms of reference for whom the project is being done and for what purpose? How do you know if the project goal has been achieved? What risks might cancel it? What does the client expect, what important events should happen during the project? I doubt it.
I'm sure you get the feeling that the text above wasn't written by a human. Try reading it and see if you can find markers that prove the opposite. What gives away a human-written text? Think for yourself. After that, check the checklist.
A brief guide on how not to get caught if you're writing with AI
Use models that allow you to specify the tone, prompt structure, and add personal context separately. The same applies to projects in ChatGPT.
For context, upload large amounts of your authentic texts—for example, export all your old posts from Telegram or other social networks, and fill at least half of the model’s context window (usually 200,000–500,000 tokens) with them.
Turn this text into a vector using n8n or any other available tool.
If you don't have many of your own texts, collect chats or posts you genuinely wrote yourself and upload them as the base.
Take the time to go through a comprehensive psychological test for accentuations and thinking patterns (for example, the Yang test—the link is above) to record your personality traits.
Write a detailed structured prompt using the COSTAR scheme: specify goals, limitations, style, typical tone, and special speech markers separately.
Be sure to explicitly specify the tone of voice and a reluctance to use overly proper grammar—let the text have inconsistencies, filler phrases, and mistakes.
Post-process. Add something personal to each text, something no neural network could know—your experiences, thoughts, doubts, mistakes, and quirks.
Only this way will AI really start to sound human and stop getting caught in the first two paragraphs.
I hope this text helped you get a deeper understanding of everyday text recognition without special services. We’d be happy to see you on the channel https://t.me/meetdeadlines!
Self-check answers
Answers
1 - human
2 - AI
3 - AI
4 - human
5 - human
6 - AI
7 - AI
8 - human
9 - human
10 - AI
11 - human
12 - AI
13 - AI
14 - AI
15 - human
16 - human
17 - AI
18 - human
19 - AI
20 - AI
21 - human
22 - AI
23 - AI
24 - human
25 - human
26 - AI
27 - AI
28 - AI
Write comment