What is AI, in plain words?

A working definition

"AI" is a broad label. The U.S. National Institute of Standards and Technology (NIST) describes an AI system as a machine-based system that, for a given set of goals, produces outputs such as predictions, recommendations or decisions [1]. That covers a lot: spam filters, fraud alerts, the route your map app suggests.

When business owners say "AI" today, they usually mean something narrower: generative AI. These are tools that create new content, such as text, images, audio or video, after learning from large amounts of existing content [2]. The chat assistant you type questions into is the most common example.

How a chat tool actually writes a reply

Under the hood, a chat tool runs a large language model. Researchers at Georgetown University's Center for Security and Emerging Technology explain the core idea simply: the model reads text and predicts which word is most likely to follow it, based on patterns it found in the data it was trained on [3]. It then adds that word and repeats, building a reply piece by piece.

Think of the way your phone suggests the next word in a text message, then scale that up enormously, with far more training text and far more careful tuning. That is a rough but useful picture. NIST describes the same thing in its guidance on generative AI: language models predict the next word or piece of a word in a sentence [2].

This explains both what these tools do well and where they slip.

What they are good at

Because they have absorbed patterns from a vast amount of writing, chat tools are good at work where the shape of the language matters:

• Drafting. First versions of emails, job posts, product descriptions, social captions and policies.
• Summarizing. Turning a long document or meeting notes into a short list of key points.
• Rewording. Making a message friendlier, shorter, more formal or easier to read.
• Organizing. Turning messy notes into a checklist, a table or a step-by-step process.
• Brainstorming. Giving you ten options so you can pick the two worth developing.

In each case, you stay the editor. The tool gives you a starting point, and you decide what is right.

What they are not

They are not a fact database. A model predicts likely text. Most of the time, likely text and true text line up. Sometimes they do not, and the tool states something false with full confidence. NIST calls this "confabulation," and notes it is a natural result of how these models are designed [2]. Lesson 3 covers how to catch it.

They do not know your business unless you tell them. The model has not seen your price list, your service area or your customer history. If you want an answer that fits your business, you have to give it the facts. Lesson 2 covers how.

They are not a person. A chat tool can sound warm and sure of itself, but it has no judgment, no accountability and no stake in the result. Decisions that affect customers, money or people belong with a person. Lesson 5 covers where to draw that line.

They can change. Providers update their tools regularly, so the same request can produce a different reply next month, or even on the next try. Part of the reason is built in: when a model picks each next piece of text, most tools add a controlled amount of randomness, which Google's developer guide describes as a setting called temperature [4].

Common mistakes

Treating the first answer as final. The first reply is a draft. Read it, correct it, and ask for changes.

Asking it for facts you do not check. Dates, prices, laws, quotes and statistics all need to be confirmed from a reliable source before you use them.

Expecting it to read your mind. A short, vague request gets a generic answer. The tool only knows what you put in front of it.

Assuming it is either magic or useless. It is a strong writing and thinking assistant with clear limits. Both halves of that sentence matter.

Worked example

A car dealership's service manager wants a friendly reminder text for customers who are due for an oil change.

She types: "Write a short, friendly text reminding a customer their oil change is due. Our service hours are Monday to Saturday, 7 AM to 6 PM. Include a link placeholder for booking. Keep it under 300 characters."

The tool returns three versions in seconds. She picks one, changes a word that does not sound like her store, and checks that the hours match what is posted on the website. The tool saved her the blank page. She still made every decision that mattered: the facts, the tone and the final wording.

Notice what she did not do. She did not ask the tool what her service hours were. It could not know, and it might have guessed.

Checklist

• I know that chat tools write by predicting likely text, not by looking up facts.
• I use them for drafting, summarizing, rewording, organizing and brainstorming.
• I give the tool the facts about my business instead of expecting it to know them.
• I check every fact, number and date before I use it.
• I keep decisions that affect customers, money or people with a person.

Try it

Open any AI chat tool you have access to. Ask it to rewrite one email you sent this week so it is shorter and friendlier. Then ask it a question about your own business, such as your opening hours. Compare the two answers. The first is something it does well. The second shows why it needs your facts.

Tips

• Think of the tool as a fast first-draft writer, not an expert.
• Give it the facts you want used, every time.
• Ask for two or three versions so you have options to choose from.
• Read every reply as a draft that still needs your judgment.

Sources

1. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework (AI RMF 1.0)," NIST AI 100-1, January 2023, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
2. National Institute of Standards and Technology, "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile," NIST AI 600-1, July 2024, https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf
3. Matthew Burtell and Helen Toner, "The Surprising Power of Next Word Prediction: Large Language Models Explained, Part 1," Center for Security and Emerging Technology, March 8, 2024, https://cset.georgetown.edu/article/the-surprising-power-of-next-word-prediction-large-language-models-explained-part-1/
4. Google, "Prompt design strategies," Gemini API documentation, https://ai.google.dev/gemini-api/docs/prompting-strategies