Free preview. Enroll for free to track progress and earn credit hours.
What generative AI is (and isn't)
In this lesson you'll
- Explain, in plain words, how a large language model produces text
- Name what generative AI tends to do well and where it tends to fail
- Describe why an AI tool can sound confident and still be wrong
- Describe the three ways of working with AI: Ask, Partner and Assign
Welcome to AI Fluency: Foundations from the Enterprise Technology Association (ETA). ETA starts with what these tools actually do, because everything else depends on it.
What generative AI is
Generative AI is software that creates new content, such as text, images, audio or computer code, when you give it an instruction. The instruction you type is called a prompt. ChatGPT, Claude, Microsoft Copilot and Gemini are well-known examples; your organization may provide one of these or another approved tool.
Most text tools are built on a large language model (LLM): a computer program trained on a very large amount of written text so that it can answer questions, write drafts, summarize and rewrite.
How an LLM produces text
- It works in tokens. A token is a small chunk of text: a word, part of a word, or even a single character.
- It learned patterns during training: which tokens tend to follow which.
- It predicts what comes next. Given your prompt, it picks a likely next token, adds it, and repeats until the response is built.
Many assistants also search the web or read files you upload. Even when a tool looks something up, the model still writes the answer, so it can still misread or misstate what it found.
The key point: an LLM generates likely-sounding language. It is not looking up a verified fact in a database.
Why AI can be confidently wrong
The U.S. National Institute of Standards and Technology (NIST) uses the term confabulation for "the production of confidently stated but erroneous or false content." You will also hear these called hallucinations. Common reasons:
- Gaps get filled with something plausible. If the model lacks the real answer, it may still produce an answer-shaped response, including names, numbers or citations that do not exist.
- Training has a cutoff date. Without a search feature, the model may not know recent events, prices or rules.
- It does not have your context. It does not know your organization's policies, customers or history unless you provide them.
- Training text contains errors and bias, and those patterns can show up in the output.
Most important: tone is not evidence. The model writes in the same fluent, confident style whether it is right or wrong.
Two real cases: in 2023 a U.S. federal judge fined lawyers who filed a brief citing court cases ChatGPT had made up (Mata v. Avianca), and in 2024 a Canadian tribunal ordered Air Canada to compensate a customer misled by its website chatbot. In both cases, the people and the organization were accountable, not the tool.
Strengths and limits
| Generally strong at | Watch out for |
|---|---|
| Drafting and rewriting emails, memos, posts | Exact facts, figures and quotes with no source |
| Summarizing long text that you provide | Citations and links (they may not exist) |
| Changing tone or format | Your organization's facts, unless you supply them |
| Brainstorming options and angles | Multi-step math, and judgment calls about people |
The same prompt can give a different answer each time, so no single answer is "the" answer. And what you type may be stored; you will learn what not to paste in Module 2.
Workplace example: healthcare administration
A benefits coordinator at a regional hospital asks, "Summarize our paid time off policy for new hires." The tool has never seen her hospital's policy, so it writes a smooth, generic policy with made-up accrual rates. She tries again in the approved tool, pasting the actual policy text, and still checks each number before sending, because summaries can drop or change details.
Three ways of working with AI
ETA describes three ways people work with AI. They differ in how much of the work the AI does.
| Way | What it looks like | Example |
|---|---|---|
| Ask | You make one request; the AI answers or drafts | "Draft a polite reminder email about Friday's timesheet deadline." |
| Partner | A back-and-forth conversation to think or create together | Planning a training session: you share goals, it suggests an outline, you push back, it revises |
| Assign | You hand AI a multi-step task and it carries out the steps | An AI agent gathers information from several sources and fills in a report template |
An agent is an AI system that can take actions, not just write text: it might search, open files, fill in forms or use connected apps. Whether you may use agents at work depends on your organization. This course focuses on Ask and Partner; the AI Agents at Work courses cover Assign in depth.
A good rule for beginners: start with Ask, move to Partner for thinking work, and use Assign only when your organization allows it, the task is low-risk, and you can check the work.
Where to be especially careful
Across industries, the most common pattern is the same: AI drafts, people decide. Take extra care, or get clear approval first, with decisions about people (hiring, discipline, benefits), advice others will rely on (legal, medical, tax, financial), anything published or sent without review, and tasks that need sensitive data. The riskier the task, the more a person must stay in charge.
Try it
In your approved AI tool (or any general AI chat tool), ask a question about something you know well, such as your job or a hobby. Write down one thing it got right and one thing that was vague, outdated or wrong. Then ask: "Which parts of your last answer should I double-check, and why?" Compare its list with yours.
Key takeaways
- An LLM builds responses by predicting likely next tokens based on patterns from training.
- Likely-sounding is not the same as true. Fluent, confident tone is not evidence.
- AI is strong at drafting, summarizing, brainstorming and reformatting. It is weak on unsourced facts, citations and your private context.
- Ask is one request, Partner is a back-and-forth collaboration, and Assign means AI carries out multi-step work. Start with Ask; use Assign only when allowed, low-risk and checkable.
- You, not the tool, are accountable for what you use.
Sources