AI Literacy Study Guide

1. What AI is (and isn't)

Narrow AI vs. general AI

• Narrow (weak) AI: built for specific tasks (translation, recommendations, face unlock). AI in use today is not general. General-purpose models like LLMs handle many tasks, but they are still not AGI.
• Artificial general intelligence (AGI): a hypothetical system able to learn and perform any intellectual task a human can. It does not currently exist.

Predictive vs. generative AI

• Predictive / discriminative: classifies or forecasts (Is this email spam? Will this loan default?).
• Generative: produces new output (write an essay, draw a picture).

2. How machines learn

The basic pipeline

1. Data: collect examples (text, images, numbers). Quality and representativeness matter more than almost anything else.
2. Training: an algorithm adjusts the model's internal numbers (parameters or weights) to reduce its errors on the examples.
3. Model: the trained result — a mathematical function that maps inputs to outputs.
4. Testing / evaluation: measure performance on data the model has not seen.
5. Inference: using the trained model on new inputs (e.g., you asking a chatbot a question).

Three main types of machine learning

Neural networks in one paragraph

Loosely inspired by the brain, a neural network is layers of simple connected units ("neurons"). Each connection has a weight. Input flows through the layers to produce an output; during training, the weights are nudged to make outputs closer to the correct answers. Modern models have millions to hundreds of billions or more of these weights.

Overfitting

When a model memorizes its training data instead of learning general patterns, it does well on training examples but poorly on new data. This is why models are tested on held-out data.

3. How large language models (LLMs) work

Chatbots such as ChatGPT, Claude, Gemini, and Copilot are built on LLMs.

1. Tokens: text is split into small chunks (words or pieces of words). Models read and write tokens, not letters or ideas.
2. Pretraining: the model reads a huge amount of text and learns to predict the next token. From this it picks up grammar, facts, styles, and patterns of reasoning — and also errors and biases in that text.
3. Fine-tuning and human feedback: the model is further trained to follow instructions and be helpful and safe, often using human ratings of responses (reinforcement learning from human feedback, RLHF).
4. Generating a response: the model produces one token at a time, each chosen based on probabilities given everything before it. Some randomness (temperature) means the same prompt can give different answers.

Terms to know

• Transformer: the neural network architecture behind modern LLMs (introduced 2017). Its "attention" mechanism lets the model weigh how words relate across a passage.
• Prompt: the input or instructions you give the model.
• Context window: how much text (in tokens) the model can consider at once. Content beyond it is not "remembered."
• Training cutoff: the model's knowledge stops at the date its training data ends, unless it is connected to search or other tools.
• Multimodal: models that handle more than text — images, audio, video.

Grounding, tools, and agents

• Grounding / retrieval-augmented generation (RAG): the model searches the web or a set of documents and uses the results in its answer. This reduces, but does not eliminate, hallucinations — it can still misread or misquote sources.
• Tool use and AI agents: models connected to tools can browse, run code, edit files, or take actions (send email, make purchases) over multiple steps. Mistakes can now have real effects, so human review matters more.

4. Limitations and risks of AI output

5. Ethics and societal impact

Bias and fairness

If training data is skewed, AI can discriminate — for example, in hiring, lending, policing, or healthcare. Facial analysis systems have been shown to be less accurate for some demographic groups, especially darker-skinned women (Gender Shades study, Buolamwini & Gebru, 2018). Fairness requires diverse data, testing across groups, and human oversight.

Privacy and data

• AI is trained on massive data sets, which may include personal information collected without clear consent.
• Prompts you enter may be stored, reviewed, or used to train future models, depending on the service's settings and policies.
• Don't enter passwords, financial details, health information, or other people's private data.

Misinformation and deepfakes

Generative AI makes it cheap to produce realistic fake text, images, audio, and video. Deepfakes can be used for scams, impersonation, harassment, and political manipulation.

Copyright and intellectual property

Many models were trained on copyrighted work without permission, leading to lawsuits from authors, artists, and news organizations. Who owns AI-generated content is also unsettled; in the U.S., the Copyright Office has said works need human authorship to be protected.

Transparency, explainability, and accountability

• Black box: it is often hard to explain why a model gave a specific output.
• Accountability: when AI causes harm, responsibility lies with the people and organizations who build, deploy, and use it.
• Human in the loop: keeping people involved in high-stakes decisions.

Regulation and governance

• EU AI Act (adopted 2024): the first major comprehensive AI law. It sorts AI uses by risk level — banned (unacceptable risk), high-risk (strict requirements), limited-risk (transparency duties, e.g., disclosing chatbots and deepfakes), and minimal-risk.
• United States: no single comprehensive federal AI law; rules come from a mix of existing laws, state laws, and agency guidance.
• Institutions: schools and employers set their own AI-use policies.

Work, environment, and access

• Jobs: AI automates some tasks, changes many jobs, and creates new ones; effects are uneven across fields.
• Environment: training and running large models use significant electricity and water for data centers.
• Digital divide: unequal access to AI tools and skills can widen existing inequalities.
• Hidden labor: much AI depends on human workers who label data and review harmful content, often for low pay.

6. Using AI effectively and responsibly

Writing good prompts

• Be specific: state the task, audience, length, and format.
• Give context: include background or examples of what you want.
• Assign a role when helpful ("Act as a tutor quizzing me on…").
• Iterate: refine and follow up; treat it as a conversation.
• Ask for reasoning or sources — then check them.

Verifying AI output

1. Check claims against reliable sources (library databases, official sites, reputable news, textbooks).
2. Confirm every citation actually exists and says what the AI claims.
3. Use lateral reading: open new tabs to see what other trusted sources say.
4. Be extra careful with numbers, dates, quotes, medical, legal, and financial information.

Academic integrity

• Follow the AI policy for each course and assignment; policies differ.
• Disclose how you used AI when required, and cite it in the style your instructor requires (APA, MLA, and Chicago all publish guidance for citing generative AI).
• You are responsible for everything you submit, including AI errors.
• AI detectors are unreliable. They produce false positives (flagging human writing as AI), especially for non-native English writers, and can be fooled. Keep drafts and notes to show your process.
• Good uses usually include brainstorming, explaining concepts, quizzing yourself, and feedback on your own drafts. Submitting AI-written work as your own is typically not allowed.

7. Key milestones

8. Glossary

AGI

Artificial general intelligence: hypothetical AI with human-level ability across any task.

AI agent

An AI system that uses tools and takes multi-step actions toward a goal.

Algorithm

A set of step-by-step instructions for solving a problem or doing a task.

Bias (algorithmic)

Systematic, unfair outcomes from an AI system, often caused by its data or design.

Chatbot

A program that converses with users in natural language.

Context window

The maximum amount of text a model can take into account at once.

Deep learning

Machine learning using neural networks with many layers.

Deepfake

AI-generated or altered media that realistically imitates a real person.

Fine-tuning

Additional training of a pretrained model on specific data or tasks.

Generative AI

AI that creates new content such as text, images, audio, or code.

Hallucination

Confident AI output that is false or fabricated.

Inference

Using a trained model to produce outputs from new inputs.

LLM

Large language model: a model trained on vast text to predict and generate language.

Machine learning (ML)

AI that learns patterns from data instead of following hand-written rules.

Natural language processing (NLP)

The AI field focused on understanding and generating human language.

Neural network

Layers of connected units with adjustable weights that learn from data.

Overfitting

When a model memorizes training data and performs poorly on new data.

Parameters (weights)

The internal numbers a model adjusts during training.

Prompt

The input or instructions given to an AI model.

RAG

Retrieval-augmented generation: the model looks up documents or search results and uses them in its answer.

RLHF

Reinforcement learning from human feedback: training a model using human ratings of its responses.

Supervised learning

Learning from labeled examples that include the correct answer.

Temperature

A setting controlling how random a model's word choices are; higher means more varied output.

Token

A chunk of text (word or word piece) that a language model processes.

Training cutoff

The date after which a model's training data ends.

Training data

The examples a model learns from.

Transformer

The neural network architecture behind modern LLMs, based on "attention."

Unsupervised learning

Finding patterns in data that has no labels.

9. Self-test

Answer out loud or on paper first, then open each question to check.