Artificial intelligence (AI) is a broad name for computer systems that perform tasks such as recognizing patterns, making predictions, recommending options, or generating content. A spam filter, a route planner, and a writing assistant can all use AI, although their jobs and risks are different.
For a beginner, the useful question is not just “Is this AI?” It is “What does this system do, what information does it use, and how will I check the result?” This guide explains the basic terms, then gives you a small exercise with an answer checklist you can use without technical knowledge.
AI, machine learning, and generative AI: what changes?
AI is the broad category. Machine learning is an approach in which a model learns patterns from data. Generative AI produces content, such as a draft, image, or summary. These categories overlap; they are not three competing products. Google’s introduction to machine learning explains learning from data and the distinction between prediction and generation.
| Task | Output | Useful check |
|---|---|---|
| Classify an email | A label such as spam or not spam | Did it block a legitimate message? |
| Estimate delivery time | A prediction | How often does the estimate match the actual arrival? |
| Draft a customer reply | New text | Does it accurately reflect the order and policy? |
| Summarize meeting notes | A shorter account of supplied text | Are decisions, conditions, and unknowns preserved? |
These examples describe possible applications, not guarantees about any particular service. A product can combine several methods. A support assistant might classify a question, retrieve a policy, and generate a reply in the same workflow.
How does a model learn, and what happens when you use it?
In machine learning, training adjusts a model using examples. Inference is the stage when the trained model processes an input and produces an output. A simple distinction is “learning a pattern” versus “applying the learned pattern.” The exact training process depends on the model and task.
Consider a system trained to recognize spam. Its development examples may associate message features with labels. When a new email arrives, it uses the learned relationships to estimate a label. That estimate can still be wrong. A writing model has a different output task: it generates text rather than returning only a spam label.
For a deeper explanation of text generation, tokens, and model training, read How Does Generative AI Work? Here, the main practical lesson is that a polished response does not establish that the system checked the facts.
A five-minute exercise: summarize a fictional project note
This is an editorial teaching exercise, not a benchmark or a report of a model test. The project, people, and figures below are fictional. The reference answer was written for this guide; it is not presented as output from a named AI tool.
1. Copy the source note
Project note — fictional example The team discussed launching a newsletter on November 12. That date is tentative and depends on completing the review. Maya will prepare the draft by November 5. Leo will review accessibility after the draft is ready; no review deadline was set. The proposed budget is $300, but it has not been approved. The team has not selected an email platform.
2. Give the assistant a bounded task
Summarize the note below using only information in the note. Use these headings: Proposed plan, Assigned tasks, Unresolved items. Keep tentative plans separate from confirmed commitments. Do not invent deadlines, approvals, or a platform choice. If a detail is missing, say that it is not specified. [Paste the fictional project note here.]
You can use a text assistant you already have access to. Use the fictional note rather than private project information. The task is intentionally small enough that you can check every claim yourself.
3. Compare the answer with this reference
- Proposed plan: Launch a newsletter on November 12, subject to completing the review.
- Assigned tasks: Maya prepares the draft by November 5. Leo reviews accessibility after the draft is ready; his deadline is not specified.
- Unresolved items: The $300 budget is proposed but unapproved. An email platform has not been selected. The launch date remains tentative.
Your assistant does not need to use these exact words. It does need to preserve the facts and their status. “The newsletter will launch on November 12 with an approved $300 budget” fails this exercise because it turns two unresolved matters into commitments.
Use a claim-by-claim review checklist
| Check | Pass condition | Example of a failure |
|---|---|---|
| Launch status | November 12 remains tentative and conditional | Calls the launch confirmed |
| Draft owner and deadline | Maya; November 5 | Changes the owner or date |
| Review sequence | Leo reviews after the draft is ready | Says review is already complete |
| Missing deadline | No review deadline specified | Invents a November 8 deadline |
| Budget | $300 proposed; not approved | Calls the amount approved |
| Platform | No platform selected | Names a platform as the team’s choice |
Mark each check pass or fail. Count passes out of six if that helps you compare two drafts. This is a local exercise score, not an estimate of a model’s overall accuracy. Passing once does not establish reliability on other documents.
If a check fails, correct the draft against the source. You can ask the assistant to revise, but inspect the new answer yourself. Asking the same system “Are you sure?” does not provide independent evidence.
Why can an AI answer be fluent and still wrong?
A response can introduce unsupported information, omit a condition, or confuse similar facts. Sometimes the problem is an ambiguous request; sometimes the system lacks relevant evidence. A summary can fail even when every sentence sounds reasonable. Read Why Do AI Chatbots Hallucinate? for more about unsupported outputs.
The NIST AI Risk Management Framework provides a broader framework for managing AI risks. For this beginner exercise, the practical application is to define what a correct answer must preserve before deciding whether an output is useful.
When should you use AI, and when is a simpler method better?
| Your need | Reasonable starting point | Your responsibility |
|---|---|---|
| Reword your own short note | A text assistant | Check meaning and tone |
| Sort a sheet using fixed rules | Spreadsheet filters or formulas | Check the rule and input data |
| Check a current policy or deadline | The authoritative source | Confirm the date and applicable conditions |
| Send a message or change a record | A controlled workflow with review | Check permissions, recipient, and final action |
AI is useful when flexible language handling helps with the task. A deterministic tool can be easier to verify when the task follows an exact rule. For example, filtering rows where a value exceeds 100 does not necessarily need a language model.
Generating a draft also differs from taking an action. An assistant that suggests a reply has not necessarily sent it. If you are considering software that can act through connected tools, read AI Agents vs. Chatbots and decide where review should happen.
A repeatable beginner workflow
- Choose a bounded task. Start with a note short enough to verify in full.
- Provide the source. Separate source information from your instructions.
- Define the output. Specify headings and how to handle missing information.
- Check each claim. Compare dates, names, amounts, conditions, and unknowns with the source.
- Decide whether to use it. Correct errors before sharing or acting on the answer.
Keep the original source alongside the final version so another person can check it. If you need answers from a larger document collection, retrieval-augmented generation (RAG) explains one way applications provide relevant material to a model. Retrieval still needs evaluation; it does not make every answer correct.
Sources and example notes
- Google for Developers: What is Machine Learning? — background on learning from data, prediction, and generation.
- NIST: AI Risk Management Framework — background on managing AI risks.
The fictional project note, reference summary, review checklist, and beginner decision guide are editorial examples created for this article. They illustrate a checking method and do not establish performance claims about any AI product.
