How AI Tools Actually Work (And Why They Still Think 2+2=5 Sometimes)
By Katherine McKean, Junior and President of my high school AI Exploration Club
If you’ve ever typed something into ChatGPT and gotten an answer that made you pause, reread, and ask your dog, “Is that… right?”, you’re not alone. One time, I asked it to help with my statistics homework, and it confidently explained that the average of 3, 5, and 9 was 8. It even bolded the answer, like it had no doubts. AI, my friends, is not a calculator—it’s a very convincing guesser.
what even is a large language model?
Let’s start here. A large language model (LLM) is like the kid in class who’s read every textbook, overheard every hallway conversation, and can mimic any teacher’s voice. It doesn’t know things in the way your brain does. It just recognizes patterns in language so well that it can predict the next word in a sentence with wild accuracy.
Think of it like supercharged autocomplete. You type “The mitochondria is the…” and it fills in “powerhouse of the cell.” Because, duh. But if you type “The mitochondria is a state in…” it might follow with “northern Canada.” Because it’s confused but still trying really hard to impress you.
why 2+2 sometimes equals 5 (according to ai)
Here’s the thing: LLMs aren’t built for math. Not really. They learn by reading text, not by running calculations. So if a bunch of the examples it’s seen have math mistakes (because the internet is full of those), it might repeat them. It’s like that one group project partner who always sounds confident but never checks their work.
I once gave ChatGPT a math worksheet and asked it to double-check my answers. It got 9 out of 10 right. On the last one, it told me the square root of 81 was 8. When I asked it to explain, it said: “Because 8 times 8 is 81.” The audacity. And yet… I’ve made that mistake at 2 a.m. too.
what ai is actually doing under the hood
Let’s break it down. The model is trained on billions of sentences from books, websites, Wikipedia pages, Reddit threads, and even those weird recipe blogs where someone tells their entire life story before getting to the ingredients. During training, it learns the patterns of how words show up next to each other.
So when you ask it a question, it’s not searching the internet or calculating an answer like a search engine or calculator would. It’s generating text, word by word, based on what seems most likely to come next. That’s how you end up with answers that sound right—but aren’t always correct.
example: history homework gone sideways
Let’s talk about that time I asked it for a quick summary of the causes of World War I. It gave me a solid paragraph, mentioned the assassination of Archduke Franz Ferdinand, alliances, militarism—all good. Then it added that the war ended in 1920 with the Treaty of Greenland.
I blinked. “Treaty of what now?”
There’s no such treaty. I looked it up. It just invented something that sounded historically plausible. That’s what AI people call a “hallucination.” Which is funnier when it happens in history class than when it happens in a biology paper due in ten minutes.
why it helps… even when it’s wrong
Despite all this, I still use ChatGPT and Claude (and even Gemini when I’m feeling chaotic). They help me brainstorm essay intros, outline presentations, and explain hard stuff in a less painful way. Just last week, Claude helped me turn a lab report into something my science teacher called “surprisingly readable.” High praise.
It’s like having a very enthusiastic intern. They don’t always know the answer, but they’ll give you five drafts and cheer you on anyway.
but wait—doesn’t it “learn” from me?
Short answer: not really, unless you’re using a version with memory turned on (like ChatGPT’s GPT-4o with history enabled). But even then, it doesn’t learn the way humans do. It remembers patterns, not meaning. I once tried to teach it that I’m left-handed and allergic to kiwis. Came back two days later, and it suggested a left-handed kiwi-picking technique. Thanks, buddy.
the limitations aren’t bugs—they’re features
AI models aren’t trying to be perfect. They’re trying to sound human. And humans? We mess up all the time. So a model that mimics us is going to mess up, too. The difference is, it does it with confidence and a polite tone. Kind of like if your dog learned how to write essays and really wanted to please you.
homework hacks and hazards
When I use AI for school, I’ve learned a few tricks:
- If I’m doing math, I double-check the answers with a real calculator.
- If I’m researching, I ask it to give me sources—and then actually read them.
- If I’m writing, I use its suggestions as starting points, not final drafts.
Also, I never ask it to write entire essays. Not just because that’s kind of cheating (and most teachers now check), but because it often misses the point. Like when I asked it to write a paragraph on The Great Gatsby’s green light symbol, and it told me it represented a traffic signal. Close, but no.
talking to ai is like texting a very formal ghost
Seriously. You type something, and it replies with perfect grammar and just enough enthusiasm to sound slightly unhinged. It’s polite, almost too polite. It remembers the topic for a while, then completely forgets what you were talking about—kind of like me after a three-hour chemistry lab.
I once had a 30-minute conversation with ChatGPT about the ethics of time travel, and it suddenly pivoted to a paragraph about climate change policy. I didn’t even ask. It just vibed into it like, “And another thing!”
how we explain it in ai club
In our AI Club meetings, we compare LLMs to those refrigerator poetry magnets. They can make sentences that sound deep, but they don’t know what any of it means. It’s up to us to provide context—and to know when to roll our eyes and say, “No, the Treaty of Greenland is not a real thing.”
We also test these models by throwing weird prompts at them. One time, we asked it to write a breakup text from a time-traveling robot to a toaster. Results were… poetic. Not helpful. But poetic.
why i still think it’s worth learning
Even with all the quirks and math fails, learning how AI works has made me a better writer, a more skeptical reader, and way better at troubleshooting tech. Plus, now when someone says “ChatGPT told me,” I get to raise an eyebrow like a detective in a noir film and ask, “Did it now?”
Understanding how LLMs work isn’t just about school—it’s about navigating a world where half the stuff online might’ve been written by a robot who once thought the capital of France was “Paris, Texas.”
If you’re curious, start small. Ask questions. Challenge the answers. Learn to spot when the AI is faking it. (Pro tip: if it starts confidently explaining a math problem with no steps and a suspicious smile, double-check.)
And if your school doesn’t have an AI club yet?
That’s where you come in.
Want to bring the power of AI to your school? Check out this step-by-step guide on How to Start a High School AI Club: 6 Easy Steps for Success.
