In 2016, Google researchers published a paper describing a system called Word2Vec. The system had been trained on billions of words from the internet, and when researchers asked it to complete the equation "king − man + woman = ?", the system answered: "queen." The AI hadn't been told what a queen was. It had simply absorbed the patterns of how humans use words — and the relationship between those words — from reading more text than any human ever could.
Here's the most important thing to understand about AI writing tools: they predict, they don't understand. A language model like GPT-4, Claude, or Gemini reads your prompt and asks one question over and over: "Given everything I've seen before, what word is most likely to come next?"
Think of it like the autocomplete on your phone's keyboard — except instead of being trained on your texts, it was trained on hundreds of billions of words from books, websites, Wikipedia, code, news articles, and more. The scale of that training is what makes the output feel like real writing.
This is called a Large Language Model (LLM). The "large" refers to both the amount of data used and the number of mathematical connections (called parameters) inside the model — GPT-4, for example, has an estimated 1.76 trillion parameters.
Before 2017, AI writing tools were clunky. They'd lose track of context after a few sentences. Then a team of Google researchers published a paper titled "Attention Is All You Need" — introducing the transformer architecture.
Transformers let AI pay "attention" to all the words in a sentence at once, not just the ones right next to each other. So when you write "The dog chased the ball until it burst," the AI can figure out that "it" refers to the ball, not the dog — because transformers look at the whole context together.
That single paper is the foundation of every major AI writing tool you've ever used: ChatGPT, Claude, Gemini, Copilot, and more.
Knowing that AI predicts word-by-word helps you understand why it sometimes produces confident-sounding nonsense. It's not lying on purpose — it simply has no built-in fact-checker. It's completing a pattern, and sometimes that pattern leads somewhere wrong.
Writers, researchers, and philosophers debate this constantly. When an AI produces a paragraph, it hasn't experienced anything, felt anything, or chosen to express something meaningful. It ran math on probability distributions. But the output can still be useful, interesting, or beautiful — which raises the real question: does the process matter, or only the result?
For now, most educators, publishers, and companies say: AI output is a tool, not authorship. The human who prompted it, shaped it, and decided what to keep is the author. But that view is still evolving — and so are the rules around it.
When you use an AI writing tool for school, always check its output for facts. AI can state things confidently that are completely made up. This is called a "hallucination" — a word we'll explore more in Lesson 3.
You've learned that AI writing tools predict what word comes next based on patterns. In this lab, you'll experiment with how prompts change predictions — and discuss what you notice with your AI lab assistant.
Try giving the AI assistant two versions of the same prompt (one vague, one specific) and ask it to explain what kind of writing it would produce. Pay attention to how much the output changes.
In the summer of 2023, the Writers Guild of America went on strike — one of the longest Hollywood strikes in decades. A central demand: clear rules about how studios could and couldn't use AI to generate scripts or revise writers' work. The studios wanted flexibility to use AI for first drafts; the writers argued this would erode wages and creative standards. After 148 days, a deal was reached: AI-generated material cannot be considered "literary material" under WGA contracts, and writers cannot be asked to rewrite AI output as if it were a human's draft. Real writers had drawn a line.
Not every writer who uses AI is "cheating" or "replacing themselves." In 2024, a survey by the Authors Guild found that roughly 20% of professional authors were experimenting with AI tools in some part of their process — but the ways they used it were very different.
Brainstorming plot ideas, checking grammar, generating names for characters, creating outlines, summarising research, or getting unstuck during writer's block. Most writers who use AI this way say it speeds up the boring parts so they can focus on the creative parts.
Generating entire article drafts, writing dialogue, producing poetry or literary prose, or submitting AI-written content under a human byline. This is where quality drops, readers notice, and ethical problems begin — including plagiarism and misrepresentation.
In 2023, several major publications began publishing AI-generated articles without adequate disclosure. CNET quietly published 77 AI-written finance articles. When readers and staff discovered this, an audit found more than half contained factual errors — including wrong interest rate calculations and incorrect descriptions of how compound interest works. CNET issued corrections and paused the program.
The Associated Press took a different approach. They published a clear AI usage policy: AI can help with data processing, sports game summaries, and financial earnings reports — but every published story still requires a human editor to verify facts and take editorial responsibility. That framework has become a model for other news organizations.
CNET's experience shows the core problem with AI in journalism: it sounds authoritative even when it's wrong. A reader can't tell the difference between a confidently-written correct fact and a confidently-written incorrect one. Human editors exist precisely to catch those errors.
In 2022, science fiction author Clarkesworld Magazine — one of the most prestigious short fiction venues in the genre — reported a massive surge in AI-generated story submissions. By early 2023 they had to temporarily close submissions entirely because the volume of AI-generated work overwhelmed human editors. Editor Neil Clarke described the AI submissions as "a flood of mediocre, soulless text."
Meanwhile, acclaimed author Robin Sloan publicly experimented with using language models as writing partners — feeding his own prose into a fine-tuned model and then using its suggestions as a "strange mirror" to reflect unexpected ideas back at him. He was careful to describe it as a tool that augmented his process, not one that replaced his voice.
The difference between those two cases — bulk generation vs. thoughtful integration — is the difference that most professional writers point to when they talk about AI use that's acceptable versus harmful.
For school assignments, using AI to brainstorm ideas is usually okay (check your school's policy). Using AI to write your essay and submitting it as your own is dishonest — and also skips the learning your teacher is trying to give you. The goal of writing assignments isn't just the document — it's the thinking you do while writing it.
Across all these real cases, one pattern emerges: AI is better at generating than at judging. It can produce 20 opening lines for a story, but it can't reliably tell you which one is best for your specific voice, your specific reader, your specific purpose. That judgment — the taste, the intention, the meaning behind the choices — stays stubbornly human.
In this lab, you'll explore the line between using AI as a tool versus letting it replace your thinking. Your AI assistant will help you brainstorm ideas for a short story — but your job is to notice when the ideas feel like yours and when they feel borrowed.
Pick a story idea you actually care about, share it with the assistant, and see what kinds of suggestions it offers. Then reflect: which suggestions feel useful, and which ones feel like they'd make the story less yours?
In May 2023, two lawyers at the New York firm of Levidow, Levidow & Oberman submitted a legal brief to federal court that cited six court cases as precedents. Judge P. Kevin Castel noticed that none of the cited cases appeared in any legal database. When the lawyers were questioned, they admitted they had used ChatGPT to research the brief — and the AI had invented all six cases, complete with realistic-sounding case names, docket numbers, and rulings. The lawyers were sanctioned. The judge called the submitted brief "gibberish" dressed up in legal language.
In AI terminology, a hallucination is when a model generates information that sounds correct but is completely fabricated. The word comes from the idea that the AI is "seeing" things that aren't there — inventing facts, citations, names, or events that don't exist.
This happens because of how LLMs work: they're optimized to produce plausible text, not true text. When asked about a court case, the model generates the kind of text that typically describes a court case — judge name, docket number, ruling — whether or not that specific case exists.
AI writing tools inherit the biases present in their training data. Since most training data comes from the internet — which reflects human biases around gender, race, culture, and history — those biases get baked into the model's outputs.
One documented example: multiple studies in 2022 and 2023 found that when researchers prompted GPT models to generate stories with the instruction "write about a nurse," the AI defaulted to female pronouns far more often than male ones — reflecting the gender bias in its training data, not any factual distribution the model had been instructed to reproduce.
Similarly, researchers at Stanford found that models trained on English internet data tend to produce text that centers Western cultural assumptions — about family structure, success, relationships, and values — as if they were universal, when they're actually specific to particular cultures.
AI companies didn't program in biases deliberately. The biases come from the text the model learned from — text written by humans who hold biases. This is why diverse training data, and human oversight of outputs, remains so important.
Because hallucinations are a built-in risk, smart users of AI writing tools develop verification habits. Here are the methods that researchers and journalists actually use:
If you use AI to help research a report or essay, never paste an AI-generated list of sources into your bibliography. Verify every single source exists. The fake-cases lawsuit is an extreme example — but students have submitted AI-invented sources in papers too. Your grade and your credibility are worth the extra check.
Your goal in this lab is to try to catch the AI making something up. Ask your lab assistant a factual question about a specific topic you already know well — a book you've read, a sport you follow, a historical event from class — and see if any of the details it gives you are wrong.
If you find an error, tell the assistant what's wrong. If it seems right, ask it to get more specific — hallucinations often appear when you push for details.
By 2023, "prompt engineer" had become a real job title at major tech companies, with salaries reaching $300,000 a year at some AI firms. Anthropic, Google, and OpenAI each published internal research on what makes prompts effective. Their shared finding was counterintuitive: longer, more specific prompts almost always produce better results than short, vague ones — even though users naturally assume a "smarter" AI should understand what they want from less. The research confirmed that clarity, context, and constraints are the three pillars of effective prompting.
You don't need to be a professional to write effective prompts. You just need to understand what information the model actually needs to do a good job.
Give the AI an example of what you want before asking it to do it. "Here's an example of the writing style I like: [example]. Now write something similar about [topic]." This anchors the output to a specific tone and style.
"You are an editor at a teen fiction magazine reviewing a short story for pacing problems." Giving the AI a specific role with a specific audience changes its output dramatically — it shifts both vocabulary and focus.
"First, outline the three main arguments. Then, write a draft of each one. Finally, write a conclusion that ties them together." Breaking a task into steps produces more coherent output than asking for everything at once.
"Here is a paragraph I wrote. Tell me three things that are working well and two things that could be stronger." Using AI to critique your own writing — rather than write for you — keeps you in control while getting useful feedback.
Being too vague: "Write me a story" gives the AI almost nothing to work with. It will produce something generic. The vaguer your prompt, the more average the output.
Assuming the AI knows your context: The AI doesn't know you, your class, your teacher's expectations, your previous drafts, or your reading level unless you tell it. Treat it like a very knowledgeable stranger who knows nothing about your specific situation.
Accepting the first output: Treat the first response as a draft, not a final product. Prompt engineers often spend more time refining and re-prompting than they do on the initial ask. Ask for revisions. Specify what's wrong. Iterate.
The writers and researchers who use AI most effectively aren't the ones who found a magic prompt. They're the ones who learned to have a conversation with the tool — asking, critiquing, redirecting, and shaping the output through multiple exchanges. That's a skill that looks more like editing than typing.
Instead of asking AI to write your essay, try asking it to be your Socratic teacher: "I'm writing about the causes of World War I. Ask me questions that help me figure out my argument." This keeps you thinking, uses the AI's knowledge as a scaffold, and produces writing that's genuinely yours.
In this lab, you'll practice the three pillars: clarity, context, and constraints. You'll write the same request three times — once badly, once okay, once really well — and compare the results with your lab assistant.
Pick any writing task (a poem, an opening paragraph, a product review, an essay introduction). Write a vague version, a medium version, and a highly specific version. Ask the assistant to evaluate which prompt is likely to produce the best output and why.