AI Foundations

Understand what LLMs are, how they work, and why they sometimes "hallucinate." Learn the AI fact-checking workflow to verify information reliability.

What is a Large Language Model (LLM)?

A Large Language Model (LLM) is an artificial intelligence system trained on massive amounts of text data from books, websites, and other written sources. Think of it as a sophisticated pattern-matching engine that predicts the most likely next word based on the patterns it learned during training.

Key Concepts

Prediction, Not Understanding: LLMs don't "know" facts—they predict statistically likely responses based on training data patterns.
Training Data Matters: The quality and recency of training data directly affects output accuracy and potential biases.
Probabilistic Output: AI generates responses based on probability, not truth—it can sound confident even when wrong.
Context Window: LLMs can only "remember" a limited amount of conversation history within a single chat session.

Critical Insight: AI is a "probabilistic engine," not a "truth engine." It generates plausible-sounding text, but plausibility does not equal accuracy. Always verify important information.

The Hallucination Showcase: When AI Makes It Up

Learn healthy skepticism through real examples of AI errors

What is a Hallucination?

In AI, a "hallucination" occurs when the model provides information that is factually incorrect but presented with total confidence. The AI doesn't "know" it's wrong—it simply generated a plausible-sounding response based on patterns.

Example 1: The "Ghost" Scholarly Citation

Asking an AI to find a specific academic source

The AI Output:

"Smith, J. (2022). The Impact of Micro-plastics on Freshwater Algae. Journal of Environmental Science, 14(3), 201-215."

The Reality:

This article does not exist. The AI "hallucinated" a plausible title, author name, journal, volume, and page numbers—all completely fabricated.

The Lesson: Always verify citations in your college library database or Google Scholar. If you can't find it in two minutes, it's likely fake.

Example 2: The "Confident" Math Error

Solving a multi-step word problem

The AI Output:

A perfectly formatted solution with clear steps ending in: "The volume is 450 cubic centimeters."

The Reality:

The AI missed a unit conversion in Step 2, making the final answer completely wrong despite the professional formatting.

The Lesson: AI can't "see" logic errors. Always do a "sanity check"—does the answer make sense given the problem? Work through the steps yourself.

Example 3: The "Historical Revision"

Asking about a specific local historical event

The AI Output:

"Your college library was founded in 1954 by Dr. Sarah Jenkins."

The Reality:

The library was founded in 1968, and Dr. Sarah Jenkins never existed. The AI fabricated both the date and the person.

The Lesson: AI training data is weakest on local, niche, or recent topics. Always verify local history, campus-specific information, and current events with authoritative sources.

The AI Fact-Checking Workflow

Follow these 3 steps before using any AI output in your work

The Golden Rule of Verification: If you can't find it in a real book or a verified website, you cannot use it in your paper. You are 100% responsible for every "fact" you turn in, regardless of whether an AI suggested it.

Step 1: The "Source Search"

Verify facts, statistics, and citations in authoritative databases

If the AI provides a fact, statistic, or citation, search for it directly in your college library database or a reputable search engine (Google Scholar, .gov, .edu sites).

Red Flag: If you can't find the specific title or author after two minutes of searching, the AI likely made it up.

Step 2: The "Logic Stress Test"

Read the AI's output critically and check for logical consistency

Read the AI's output critically and ask yourself:

Does the math add up? Check calculations manually.
Does the conclusion follow the evidence provided?
If you ask: "Wait, can you double-check that third step? It seems off," does it immediately apologize and change its answer?

Pro Tip: If the AI quickly "apologizes" and changes its answer when challenged, that's a sign the original response was unreliable.

Step 3: The "Lateral Verification"

Cross-check information using multiple AI tools or sources

Open a new, separate chat window or use a different tool (like Perplexity or Copilot) and ask the same question.

Critical Rule: If two different AI models give you conflicting "facts," you must assume both are wrong until you verify with a human-authored source.

AI is NOT a Search Engine

Many students treat AI like "Super Google," but they function very differently.

Feature	Search Engine (Google)	Generative AI (ChatGPT/Claude)
Goal	Finds existing information on the web.	Generates new text based on patterns.
Accuracy	Directs you to a source; you judge the source.	Creates an answer; it may invent facts (hallucinations).
Context	Single-search queries.	"Remembers" the conversation flow.
Best For	Fact-checking, finding URLs, current events.	Brainstorming, summarizing, explaining concepts.

The Anatomy of an LLM: How it "Thinks"

AI doesn't have a "brain"; it has a Neural Network.

Training Data

AI has "read" billions of pages of internet text, books, and code. It learns the relationship between words.

Probability

When you ask a question, the AI is essentially playing a high-speed game of "Predict the Next Word."

The Token

AI breaks words into chunks called "tokens." It doesn't see "Apple"; it sees a mathematical representation of that concept.

Ethics & Bias: The Mirror Effect

AI models are mirrors of the data they were fed. Because much of the internet is Western-centric and contains human prejudices, AI can:

Stereotype: It may default to certain gender or racial roles in its stories or code.

Exclude: It may not have enough data on non-Western cultures, languages, or local institutional history.

Your Role

As a student, your job is to be the Critical Editor. Never take an AI's perspective as the "neutral" truth.

Digital Footprint: Privacy & Your Data

When you use a "Free" AI tool, your data is often the payment.

The Training Loop

In most free versions, the prompts you type (and the essays you paste) can be used to train the next version of the model.

Confidentiality

Never paste personal ID numbers, unpublished research, or private emails into a public AI.

Pro Tip

Use Microsoft Copilot with your school account whenever possible; it typically offers higher data protection than personal accounts.

AI Jargon Buster (The Mini-Glossary)

Generative AI

AI that can create new content (text, images, audio).

Prompt

The instruction or question you give the AI.

Hallucination

When AI confidently provides false information.

Parameter

The "adjustable knobs" in an AI model that determine its complexity.

Multimodal

AI that can process more than one type of data (e.g., text AND images).

The Environmental Cost of AI

AI is "virtual," but its physical footprint is massive.

Energy Consumption

Running a single AI query can use 10x more electricity than a standard Google search.

Water Usage

Data centers require millions of gallons of water to cool the servers that process AI requests.

The Takeaway: Being AI-literate also means being an intentional user. Only use AI when it adds real value to your learning process.

AI for Accessibility: Leveling the Playing Field

For many students, AI is a life-changing assistive technology.

Learning Differences

AI can summarize complex academic texts into "Plain English" for students with dyslexia or cognitive processing differences.

Visual/Hearing Aids

Multimodal AI can describe images for the visually impaired or provide high-accuracy live transcription for the hearing impaired.

The Goal

AI should be used to remove barriers to learning, not to bypass the learning itself.

Ready to Put This Into Practice?

Now that you understand how AI works and its limitations, explore our toolkit to find secure AI tools and learn effective prompting techniques.

Explore the Toolkit