Two days before my history submission, I asked an AI for sources. It gave me a gorgeous list. Author, title, publisher, year, even page numbers. I felt like a genius.

My teacher couldn't find one of them. Neither could I. Neither could the school library.

The book did not exist. The AI had invented the whole thing — down to the page number — without a single hint that it was making it up. Full confidence. Zero truth.

That was the day I stopped treating these tools like magic and actually learnt how they work. Every word below is going to follow you into classrooms and job interviews for the next ten years — might as well get it now, while it's still a head start.

---

What an LLM actually is (and why it lied to me)

ChatGPT, Claude, Gemini — all LLMs: Large Language Models. And here's what nobody says clearly enough: an LLM does not look anything up. It's not Google with a personality.

It read an enormous pile of text and learnt, statistically, which word usually comes next. That's the whole engine. When you ask it something, it's producing a very good guess at what a correct-sounding answer looks like.

It's predicting what comes next — not retrieving a fact from a drawer.

Which explains my fake book perfectly. Citations follow a pattern, so it generated something shaped exactly like a citation. It wasn't lying to me — lying needs intent. It was pattern-matching, and the pattern came out as fiction.

There's a word for this: hallucination. The model says something false with total confidence, no wobble in its voice. It's worst with the exact things you'd want to quote — names, dates, numbers, statistics, citations.

---

RAG: how you make AI read your notes

So if a model can't look anything up, how does "chat with your PDF" work? How does it answer questions about a textbook it never saw in training?

RAG — Retrieval-Augmented Generation. Scary name, simple idea: search first, then answer. That's it. That's the trick.

Find the relevant pages first, then let the model answer using only those.

Here's what actually happens the moment you upload your chemistry notes:

1. Your file gets chunked — sliced into bite-sized pieces, a few paragraphs each.
2. Each chunk becomes an embedding: a list of numbers that captures its meaning. Chunks about the same thing end up with similar numbers.
3. Those numbers live in a vector database — storage that's brilliant at "find me the bits closest in meaning to this question."
4. You ask something. It grabs the 3–5 most relevant chunks and quietly pastes them into the prompt alongside your question.
5. The model answers from those pages. Not from memory.

Step 5 is the whole point. The answer is anchored to a real source, so hallucination drops hard and you can flip to the page and check. The idea comes from a 2020 research paper, and today it runs quietly inside almost every serious AI product.

---

MCP: the USB-C moment (stay with me)

This is the newest one, and I promise almost nobody in your class has heard of it yet. MCP: Model Context Protocol. Anthropic put it out as an open standard in late 2024 and the rest of the industry jumped on it fast.

The problem it fixes: an AI that can only talk is kind of useless. You want it to check your calendar, open your files, pull from a database. But every one of those connections used to be hand-built for every single app. Fifty tools, fifty different plugs, nothing fitting anything.

Basically: the drawer of chargers your family refuses to throw away.

One standard plug, many tools. That's the whole idea.

MCP is the one standard plug. Build a connector once, and any AI app that speaks MCP can use it.

Why you should care: it's the reason assistants have started doing things instead of just describing them. That jump — from "here's how you'd do it" to "okay, done" — is the biggest shift happening in tech right now, and it's happening while you're still in school. Good timing, actually.

---

Agents: when AI stops talking and starts doing

Give a model tools (through MCP or otherwise), a goal, and a loop — and you've got an agent. It takes a step, looks at what happened, decides the next step, repeats until the job's done.

"Find every source on this topic, check which ones are real, put them in a table" — that's agent work. Not one answer. A sequence, with checking in between.

Which is exactly why "can you use AI?" is quietly becoming the wrong question. Everyone can type into a chat box, yaar. The actual skill is knowing what to hand over, and how to check what comes back.

---

Screenshot this bit

*You'll hear every single one of these within a year. Probably from someone who doesn't know what they mean.

---

Okay but what actually affects you

Vocabulary is the easy half. This is the half that changes your marks.

Check anything specific. Names, dates, stats, citations, formulas. If it's a fact someone could look up, look it up. My teacher caught mine in about thirty seconds.

Submitting AI work as your own is a bad trade. Not mainly because you'll get caught — though schools are getting scarily good at spotting it — but because the thing you skipped is the thing being graded. Use it to explain the concept you didn't follow in class. Then write the answer yourself, in your own words.

Cross-check before you submit. Thirty seconds saves a very awkward conversation.

Never paste personal stuff into a chatbot. Aadhaar numbers, your address, your friends' details, screenshots of family chats. Assume anything you type could be stored and read by a human someday. Costs you nothing to be careful, saves you a lot.

Assume voices and videos can be faked. Deepfakes are cheap now, and India has already had scams using cloned voices of people's relatives. If someone calls in a panic asking for money, hang up and ring them back on their actual number. Better: pick a family code word this week. Sounds paranoid, takes two minutes, works.