Everyday AI: Best Practices for Using AI

Artificial intelligence (AI) is a broad term used to describe any technology that mimics aspects of human intelligence, such as machines or software that perform tasks like problem-solving, predicting, learning, and acting autonomously. Web search engines, face ID on your phones, spam filters, streaming service recommendations from platforms like Netflix and YouTube, voice assistants like Amazon Alexi or Apple Siri, and apps like Google Maps, Expedia, and Uber all fall under this AI umbrella.

Because of the wide range of technologies under this umbrella, lumping them together leads to confusion, misunderstandings, and misdirection. In fact, computer and software engineers cannot even agree upon one definition of AI, as you can see in Table 1.

How Does Generative AI Work?

On a very basic level, when you type a prompt into a generative AI like Copilot or ChatGPT, the AI breaks your text into small units called tokens. Based on the data and models it was trained on, it creates a response token by token based on the probabilities of what usually surrounds those tokens. In essence, generative AI tools analyze patterns in text, such as the order, frequency, and relationships between words, to mimic human speech. 

As convincing as this mimicry can be, this means that, unlike a human who understands the actual semantics of a sentence (in other words, what the sentence actually means), generative AI tools don't understand language the same way humans do. While many AI tools use words like "summarize" and "analyze" to describe the AI's output, these tools aren't analyzing or summarizing in the same way you would. The best way to describe what AI is doing is giving you its best guess of what it thinks an answer to your query should look like. 

Common AI Terminology

Large Language Models (LLM) – Self-contained computer programs that mimic the pattern and structure of human language by statistically analyzing vast quantities of language data. The process of analyzing such vast quantities of data is referred to as machine learning. Because LLMs are trained on fixed data sets, they must be updated periodically in order to incorporate information from recent events. Examples include GPT-5o, Claude, Gemini, Llama, and Mistral.

LLM Wrapper – Most people interact with LLMs through software wrappers like ChatGPT, which serve as intermediaries. These wrappers augment an LLM’s fixed information by adding external or user-provided information to the prompt in order to generate more current and relevant responses.

Prompt – Sentences given to an LLM to help it generate new content by predicting the words associated with those sentences. These prompts may be processed by the LLM Wrapper first before going to the LLM.

Tokens – LLMs don’t read words like we do. Instead, LLMs process language by converting text into tokens — small units of word parts, punctuation, or spaces — before analyzing the patterns between them. Unlike humans, who understand meaning, LLMs generate responses based on the statistical relationships between tokens.

Table 1. Definitions of AI