What is generative AI? A plain-English explainer for professionals

Generative AI is software that produces new content (text, images, code, audio, or video) by predicting what should come next, based on patterns it learned from enormous amounts of existing material. When you type a question into ChatGPT, Claude, or Gemini and get a paragraph back, that paragraph did not exist anywhere before. The model wrote it, one small piece at a time.

That single fact explains most of what these tools do well, most of what they get wrong, and how you should treat their output at work.

The one-sentence definition, and what it leaves out

"A model that generates new content by predicting likely continuations" is accurate but skips the three things a professional needs to understand:

1. It learned from examples, not rules. Nobody wrote instructions for how to draft a resignation letter. The model saw huge numbers of letters, articles, emails, and conversations during training and learned the patterns well enough to produce a convincing new one on demand.
2. It has no database of facts. It does not look up your answer; it composes one. Facts it saw many times during training tend to come out right. Facts it saw rarely, or never, come out as plausible guesses.
3. It is probabilistic. Ask the same question twice and you can get two different answers. That is not a bug; it is how the generation works, and it is why you can ask for "three versions" and get three genuinely different drafts.

The large language model is the engine behind text-based generative AI; the glossary entry and the deeper guide on how large language models work cover the mechanics. This guide stays at the level of what it means for your Monday.

How it differs from the software you already use

Most software you have used at work is deterministic. Give Excel the same numbers and formula and you get the same result every time. A database query returns exactly the records that match. A search engine returns pages that already exist. Generative AI breaks all three expectations.

The last row is the practical takeaway. Generative AI does not replace the tools in the first two columns. It sits beside them, and the best workflows use each for what it is good at: the model drafts the formula, Excel runs it; the search engine finds the source, the model summarizes it; you supply the facts, the model supplies the prose.

What it is good at

The strongest use cases share a pattern: you provide the substance and the model provides the form, or you provide the form and the model reshapes the substance.

• Drafting from raw material. Bullet points into an email. Meeting notes into minutes. A rough outline into a proposal. The more of your own material you give it, the better and safer the result.
• Transforming. Change the tone from blunt to diplomatic. Cut 800 words to 200. Turn a paragraph into a table. Translate for a colleague in another office. Rewrite a policy at an eighth-grade reading level.
• Summarizing and extracting. Pull the decisions, deadlines, and open questions out of a forty-message thread. List every obligation in a contract, then ask follow-up questions about the ones that matter.
• Explaining. Ask what a term means, how a process works, or why a formula does what it does, at whatever level of detail you want. Then ask it to quiz you.
• Brainstorming and critiquing. Twenty names for a program. The strongest objections to a plan. What a skeptical customer would say about your pricing page.
• Pattern-following work. Spreadsheet formulas, regular expressions, boilerplate code, document templates, and anything else where the shape is well known and the details vary.

What it is bad at

The weaknesses come from the same source as the strengths: it predicts rather than knows.

• Specific facts it rarely saw. Details about small companies, niche regulations, individual people, product specifications, and local rules are where it invents the most. The phenomenon has a name, hallucination, and a dedicated guide: AI hallucinations explained.
• Arithmetic and counting. It handles numbers as text patterns, not quantities. Simple sums are usually fine; anything with several steps or large figures should be done in a formula, a calculator, or by asking a tool that can run code.
• Recent events. Every model has a training cutoff. Without web search turned on, it does not know what happened after that date and may confidently describe an outdated state of the world.
• Your business. It knows nothing about your customers, your history, your org chart, or last quarter's numbers unless you tell it. Anything it says about your specific situation is a guess dressed as an answer.
• Precision under many constraints. Ask for a 150-word bio that mentions five things, avoids three words, and ends with a question, and expect to check every requirement. It will usually get most of them.
• Consistency. It can contradict itself between paragraphs or between conversations, because each answer is generated fresh.

None of these are reasons to avoid the technology. They are reasons to design your use of it around the review step.

The kinds of generative AI you will meet at work

Text is where most professionals start, but the same idea now spans every medium.

• Text and chat: ChatGPT, Claude, Gemini, and Microsoft Copilot. Drafting, summarizing, analysis, coding help.
• Images: Midjourney, Adobe Firefly, and the image tools built into Canva and the major chat assistants. Concept art, marketing visuals, quick mockups.
• Audio and voice: ElevenLabs for generated speech, Descript for editing recordings by editing the transcript.
• Video: Runway for generated and edited clips, Synthesia for presenter-style videos from a script.
• Code: GitHub Copilot, Cursor, and Claude Code write and edit software from plain-English instructions.

Many assistants are now multimodal: you can paste a screenshot, a chart, or a photo of a whiteboard and ask questions about it. That is often the fastest way to get a messy diagram into a clean list.

A worked example: one task, two ways

Take a task most managers face: a customer had a bad experience and you owe them a message today.

Without AI, you stare at the blank email for a while, write something too defensive, soften it, delete half, and send it forty minutes later, still not sure of the tone.

With generative AI, you spend two minutes typing the facts and the outcome you want, and the model does the composition. The prompt below is copy-ready; replace the bracketed parts.

You are a customer success lead known for honest, warm, non-defensive communication.

Task: Draft an email to a customer after a service failure.

Context:
- What happened, in my words: [DESCRIBE THE PROBLEM AND WHAT WE DID WRONG]
- What we have already done to fix it: [WHAT HAS BEEN FIXED OR REFUNDED]
- What I can offer now: [CREDIT, CALL, TIMELINE, OR NOTHING BEYOND AN APOLOGY]
- Relationship: [NEW CUSTOMER / LONG-TERM ACCOUNT / AT RISK OF LEAVING]

Format: Under 150 words. Plain, direct sentences. Acknowledge the problem in the first line, take responsibility without legal-sounding hedging, state the fix, offer one concrete next step. No exclamation marks.

Constraints: Do not promise anything I have not listed. Do not invent timelines. If you need a detail I have not given you, ask me before drafting.

What you get back is a solid draft in seconds. What you still do: read it as the customer would, check that every promise in it is one you actually made, and adjust the one sentence that does not sound like you. That review is the job, not overhead. The model wrote the form; you own the substance.

The mental model to keep

Think of a generative AI assistant as a fast, well-read colleague who joined this morning: broadly knowledgeable, tireless, eager to please, with no memory of yesterday's conversation, no access to your systems unless you connect them, and no accountability for being wrong. You would happily hand that colleague a first draft, a summary, or a brainstorm. You would not let them send the client the final numbers unread.

Three habits follow from that picture:

1. Give context every time. The colleague forgot everything overnight. Paste the background, name the audience, state the format.
2. Ask for what you can check. Drafts, options, structures, and explanations are easy to review. Bare facts and figures are not, so get those from a source and hand them to the model.
3. Own the output. The moment you send it, it is yours. Read it like your name is on it, because it is.

Next steps

• Go one level deeper with How large language models work and the glossary entries for generative AI and machine learning.
• Understand the main failure mode in AI hallucinations explained.
• Put it to work with the operating manual How to use AI at work and the best AI tools for work.
• Try the prompts in writing and email, or build your own with the prompt builder.
• See what it means for your role under AI for your job, or start the course.