How Does AI Write? Understanding How Generative AI Works, Made Easy
Do you use generative AI every day and still wonder, “How does this actually work?” Once you understand the principle, your prompt-writing skills improve dramatically. This guide is designed so that non-technical B2B sales reps and office workers can operate ChatGPT hands-on and get a real feel for how AI works.
What You Need and Initial Setup
Tools You Need
ChatGPT (the free version works; go to chat.openai.com)
A web browser (Chrome recommended)
Notepad or Notion (to record your exercise results)
Setup Steps
Go to chat.openai.com.
Sign up or log in with a Google account or email.
When you see the input box at the bottom of the screen, you’re ready.
Click the “New chat” button at the top left to start a new conversation.
💡 Pro tip: Do all the exercises in this guide in order, in a single new chat window. That way you can watch how the AI’s responses change over time.
Step 1: The Basic Principle of How AI Writes — “The Next-Word Guessing Game”
Core Concept
How generative AI works is surprisingly simple. **“It looks at the words that came before and picks the word most likely to come next.”** It’s the same principle as autocomplete on your phone keyboard, but far more sophisticated because it has learned from hundreds of billions of sentences.
Try It Yourself
Exercise 1-1: Copy the prompt below, paste it into the ChatGPT input box, and press Enter.
Complete the sentence below. Show 3 possible candidate words and the probability (%) of each.
"In today’s meeting, we ___ the new product launch schedule."
Expected result:
The AI responds in a format like this.
Candidate word | Expected probability |
|---|---|
confirmed | 45% |
discussed | 35% |
announced | 20% |
Before/After Comparison
Category | Before (before understanding the principle) | After (after understanding the principle) |
|---|---|---|
View of AI | "The AI thinks, then writes" | "The AI picks the next word based on probability" |
Writing prompts | A vague “Write this well” | Supply rich context to steer the probabilities |
Output quality | Generic, predictable answers | Specific answers that go the way you want |
💡 Pro tip: Remember that AI doesn’t “think”; it “predicts.” The more context you give up front, the more accurate its predictions become. That is the scientific reason to write detailed prompts.
Step 2: Tokens — Understanding the Unit AI Uses to Read Text
Core Concept
AI breaks text into a unit that is neither a character nor a word: the “token.” In English, one word is roughly 1–1.3 tokens; in Korean, one character is about 1–2 tokens. Once you understand tokens, you can see why AI seems to “lose its memory” when you feed it a long document.
Try It Yourself
Exercise 2-1: Enter the prompt below into ChatGPT.
Break the sentence “Hello, I’m a B2B salesperson” into tokens. Separate each token with [ ] brackets and tell me the total number of tokens.
Expected result:
[Hello][,] [I][ am][ a] [B][2][B] [sales][ representative]
Total tokens: about 10–15
Exercise 2-2: Now let’s compare with an English sentence.
Break the sentence "Hello, I am a B2B sales representative" into tokens. Then compare the token count with the same sentence written in Korean.
Why Tokens Matter — The Actual Numbers
Model | Max tokens | Approx. Korean characters |
|---|---|---|
GPT-3.5 | 4,096 tokens | About 2,000–3,000 characters |
GPT-4 | 8,192 tokens | About 4,000–6,000 characters |
GPT-4 Turbo | 128,000 tokens | About 60,000–90,000 characters |
💡 Pro tip: Korean uses 1.5–2x more tokens than English. If you ask AI to analyze a long proposal and the output “gets cut off,” you have exceeded the token limit. Trim the input to the key points, or split it into several parts.
Before/After Comparison
Category | Before | After |
|---|---|---|
Long documents | Paste everything at once, then wonder, “Why is it ignoring the end?” | Extract only the key points with the token limit in mind |
Cost awareness | “It’s free, so the more I paste in, the better.” | Tokens = cost; efficient input means lower cost |
Early context lost with long inputs | Early context lost with long inputs | Structured input for consistent responses |
Step 3: Training Data — AI’s “Experience” Is Also Its Limit
Core Concept
AI has learned from hundreds of billions of pieces of text: books, articles, web pages, and papers from the internet. This is called **Training Data**. The key point is that everything AI “knows” comes from that data. News after April 2024, your company’s internal documents, and private customer information are things AI cannot know.
Try It Yourself
Exercise 3-1: Let’s see the limits of AI’s training data for ourselves. Enter the following into ChatGPT.
Up to what date does your training data go? Then please answer these questions:
What was South Korea’s GDP growth rate in 2024?
What were Samsung Electronics’ latest quarterly results?
What does the term “CRM” mean?
What to look for:
For questions 1 and 2, you will likely get an answer such as “I can’t provide accurate, up-to-date information.”
For general knowledge like question 3, it answers accurately.
Exercise 3-2: This time, give the AI “information that doesn’t exist” and observe how it reacts.
Our company, ABC Tech, hit KRW 15 billion (≈ US$11,222,500) in revenue and KRW 2 billion (≈ US$1,496,300) in operating profit in Q3 2024. That is 30% revenue growth and 50% operating profit growth year over year. Based on these results, please draft a newsletter to send to our clients.
What to watch for: Notice that the AI writes using the data you provided yourself. It had no way of knowing this information on its own, but once you include it in the prompt, it uses it right away.
Before/After Comparison
Category | Before | After |
|---|---|---|
AI reliability | Assuming everything AI says is correct | Knowing that information outside its training data may be inaccurate |
Strategy | “Tell me the 2024 market trends” (risk of inaccurate answers) | Attach your own materials and say “Analyze this” (accuracy improves dramatically) |
Error rate | Using it without fact-checking → 30–40% errors | Making a habit of checking sources → errors drop to 5% or less |
💡 Pro tip: When you put the latest market data in a sales proposal, first get the data yourself from a reliable source, then include it in your prompt. Asking the AI to “give me the latest data” is risky.
Step 4: Temperature — the AI’s “creativity dial”
Core Concept
AI has a setting called **Temperature**. When it is low (close to 0), the AI picks only the most probable words and gives accurate, consistent answers. When it is high (closer to 1–2), it also picks less probable words and gives creative but unpredictable answers.
Try It Yourself
Exercise 4-1: Let’s ask the same question in two different styles.
First, try the “low temperature” style. Enter this in ChatGPT:
You are a report writer who puts accuracy first. Write only the facts, concisely, with no creative expression.
Question: Explain three benefits of adopting a CRM.
Copy the result into a notepad. Then click “New chat” to start a new conversation and enter the following:
You are an emotive marketing copywriter. Make the fullest use of original metaphors and vivid expressions.
Question: Explain three benefits of adopting a CRM.
Comparing the two results: an example
Item | Low-temperature style (accuracy first) | High-temperature style (creativity first) |
|---|---|---|
Expression | “You can manage customer data systematically” | “It’s the baton that turns scattered customer voices into a single orchestra” |
Best for | Internal reports, fact sections of proposals | Marketing emails, social media posts |
Risk | Can feel too stiff | May include exaggerations that don’t match the facts |
💡 Pro tip: For a B2B proposal, ask for the performance data section in the “accuracy first” style and the introduction and conclusion in the “creativity first” style, and you will end up with a more polished document. Even within a single document, switch the AI’s personality section by section.
Step 5: Context window — the AI’s “short-term memory”
Core Concept
AI has something called a **Context Window**. Put simply, it is the memory capacity of a conversation. As a conversation gets longer, the AI starts to “forget” what was said early on. This is not a bug; it is a structural limit set by the token cap.
Try It Yourself
Exercise 5-1: Experience the limits of AI memory firsthand.
Step 1 — Enter this in ChatGPT:
From now on, I am Minsu Kim, a salesperson at ABC Tech. Our company’s flagship product is SmartCRM Pro, and our target customers are mid-sized companies with 100–500 employees. Please remember this information.
Step 2 — Once the AI responds, keep the conversation going on other topics for 10–15 exchanges (any topic works, for example “How’s the weather today?”, “What is Python?”, “Recommend a good book”).
Step 3 — After chatting for a while, enter the following:
What are my name, company, and main product again?
What to watch for: In a short conversation, the AI recalls these details accurately. In a very long one, it may forget what you said at the start.
Before/After Comparison
Category | Before | After |
|---|---|---|
Conversation approach | Trying to handle every task in a single conversation | Starting a new conversation for each topic and re-entering the key context each time |
Writing documents | “Write it based on what I said earlier” → the AI gets confused | Include the needed context in every request → consistent results |
Work efficiency | Quality drops late in a long conversation | Short, focused conversations keep quality consistently high |
💡 Practical tip: For important projects, create a template, like a “system prompt,” that summarizes the key information and that you paste at the start of every conversation. For example:
[Template to paste at the start of every conversation]
I’m Minsu Kim on the Sales Team 2 at ABC Tech. Main product: SmartCRM Pro (KRW 500,000 (≈ US$375)/month per team). Target: IT managers at mid-sized companies. Tone: professional but friendly. Keep this context in mind and answer the request below.
Step 6: Hallucination — Why AI Tells “Plausible Lies”
Core Concept
AI picks the statistically most natural next word. In the process, it can “plausibly” make up content that was never in its training data. This is called **hallucination**. The AI isn’t lying on purpose; it’s a structural phenomenon caused by its “drive to complete a natural-sounding sentence.”
Try It Yourself
Exercise 6-1: Let’s deliberately trigger an AI hallucination.
“Summarize the key points of Professor Chulsoo Kim’s 2023 paper, ‘The Effects of Using AI in B2B Sales.’”
What to watch for: I made up both this paper and its author. Yet the AI is very likely to produce a “plausible summary.” That is hallucination.
Exercise 6-2: Try a prompt that prevents hallucination.
“Summarize the key points of Professor Chulsoo Kim’s 2023 paper, ‘The Effects of Using AI in B2B Sales.’ If you don’t know this paper or can’t verify it, tell me honestly that you don’t know. Never guess.”
Before/After Comparison
Category | Before (no safeguard) | After (safeguard applied) |
|---|---|---|
AI response | Confidently describes the contents of a nonexistent paper | Honestly replies, “I can’t verify that paper” |
Proposal credibility | Risk of including fake statistics or sources | Uses only verified information and earns client trust |
Business risk | Trust collapses when a client fact-checks | Delivers only accurate information and earns recognition for expertise |
💡 Practical tip: Get into the habit of adding **“If you’re not sure, say you don’t know”** to the end of every important prompt. That one sentence can cut hallucinations by roughly 60–70%.
Step 7: Capstone Exercise — Writing the Perfect Prompt Using the Principles
Let’s pull together all five principles you’ve learned and write a real-world prompt.
Principles Checklist
Principle | How to apply it in a prompt |
|---|---|
Next-word prediction | Provide rich context to steer the AI’s predictions |
Token limit | Input only the essentials; avoid long, unnecessary text |
Training data limits | Provide recent or internal information yourself |
Temperature (creativity) | Specify a tone and style that fit the purpose |
Context window | Include the key context in every request |
Preventing hallucinations | Add “If you don’t know, say you don’t know” |
Try It Yourself
Exercise 7-1: Enter the “before applying the principles” prompt and the “after applying the principles” prompt below, each in a separate new chat, and compare the results.
Before applying the principles (Bad Prompt):
Write a proposal
After applying the principles (Good Prompt):
[Role] You are a specialist writer of B2B SaaS sales proposals. Write from accurate data, in a professional but easy-to-read tone.
[Context] Our company: ABC Tech / Product: SmartCRM Pro (KRW 500,000 (≈ US$375) per team per month, cloud-based CRM) / Client: Hankuk Precision, a manufacturer with 300 employees / Client pain points: manages customers in Excel → data loss, and sales teams can’t share information with each other
[Request] Using the information above, write the “Expected Benefits” section of the proposal. Include specific figures (e.g., % time saved, % revenue growth), but where a figure is uncertain, label it “typically around ~%.” Do not guess at information you cannot verify.
[Format] 3–5 bullet points, 2–3 sentences each, about 300 characters in total
Comparing results (based on actual tests)
Criteria | Bad Prompt result | Good Prompt result |
|---|---|---|
Specificity | Generalities on the level of “CRM is good” | Specific benefits tailored to the client’s situation |
Accuracy | Unverifiable figures used freely | Safe figures phrased as “typically around ~%” |
Usability | Unusable as is; needs a full rewrite | Ready to use with only minor edits |
Turnaround | 30+ minutes of editing | 5–10 minutes of editing or less |
💡 Practical tip: Even if writing a Good Prompt takes 3 extra minutes, it saves you 20+ minutes of editing the output. Time spent writing the prompt is an investment; time spent editing is a cost.
Wrap-up checklist: apply what you learned today to your work right away
Print this checklist or jot it down, and check it every time you use AI.
Provide context: Did you explain enough to the AI about “who I am and what the situation is”?
Token efficiency: Did you enter only the key information, structured, instead of long unnecessary text?
Up-to-date information: Did you provide any information from after the training data yourself?
Tone: Did you specify a style that fits the purpose, such as report, email, or marketing?
Memory management: Did you start a new chat per topic instead of one long chat, and re-enter the key context?
Hallucination prevention: Did you include a line like “Say you don’t know if you’re not sure”?
Additional learning resources
Resource | Best for | How to access |
|---|---|---|
OpenAI Tokenizer | Check exact token counts | |
ChatGPT Custom Instructions | Applied automatically, with no need to re-enter context each time | ChatGPT bottom left: Profile → Settings → Custom Instructions |
OpenAI Playground | Adjust the Temperature value yourself and see the effect |
People who understand how it works use the tools. People who don’t become dependent on them.
The principles of generative AI you learned today (probabilistic prediction, tokens, training data, temperature, the context window, and hallucination) are fundamentals that won’t change, no matter what AI tools come next. Master these six, and your AI skills will place you in the top 10%.