What AI Can Do—and How to Recognize Its Limits

Learn to match the job to the right kind of help.

You ask an assistant to write an invitation, estimate attendance, and reserve a meeting room.

It replies with an attractive message, a confident number, and a sentence saying the room is booked.

Those three results deserve three different questions.

Does the invitation contain the right facts? Is there evidence behind the attendance estimate? Was a reservation actually made?

Learning those differences helps you get useful work from AI. It also helps you recognize when a good-looking answer has gone beyond what the system knows or can do.

By the end of this lesson, you will be able to sort common tasks by the kind of help they require and choose an appropriate way to check the result.

Four kinds of work you may encounter

AI is often discussed as if it were one tool doing one thing. In practice, your tasks may involve several kinds of work.

Creating material

You provide a goal and some information. The system generates something new: a draft, an image, an explanation, a list of ideas, or a piece of code.

For a fictional cleanup event, it might turn your notes into a friendly invitation.

Your main questions are whether the result serves the audience, follows your instructions, and stays accurate.

Predicting an outcome

A system estimates something that is not yet known, such as future demand or likely attendance.

An estimate needs relevant evidence and a method suited to the question. A plausible number written in a conversation is not automatically a sound forecast.

If you have records from past events, you can examine them. If you have no records, the assistant should help you identify assumptions and gather information rather than pretend uncertainty has disappeared.

Sorting or recognizing

A system assigns categories or identifies patterns. It might sort support messages by topic or help extract dates from supplied documents.

The categories must be clear, and you should check cases that could fit more than one category.

A message saying “I need to change my order because the delivery address is wrong” might involve both order changes and shipping. A useful process handles that ambiguity.

Taking action

An application may use connected tools to create a calendar event, update a record, or send a message.

This adds another question: Was the correct action authorized and successfully performed on the correct target?

Generating text that says “Your meeting is scheduled” does not establish that a calendar entry exists.

These categories are a practical way to organize work, not a claim that every system fits into only one box. Machine-learning applications can include both prediction and generation, and broader workflows may combine several methods. Google's introduction to machine learning.

Look for tasks where the value is easy to see

Imagine that you have three pages of notes from an introductory class. You could ask an assistant to group them by topic, identify unfamiliar terms, and create five questions you can answer from the notes.

That gives you a concrete result to inspect. You can compare the topic groups with the original notes. You can check whether the questions are answerable. You can attempt them without looking at the answers.

Other approachable tasks include:

Turning a rough paragraph into a clearer draft.
Comparing options against criteria you provide.
Explaining an unfamiliar term at your current level.
Reformatting information into a table.
Identifying missing details in a plan.

Current assistants may offer additional abilities through browsing, file analysis, voice, and other tools. The available combination depends on the product and your access. OpenAI's capability guide illustrates the range within one application; it should not be read as a promise that every feature is available in every session. ChatGPT capabilities overview.

Missing context can spoil an otherwise sensible answer

Suppose you ask for a plan to organize your garage. The answer recommends placing heavy containers on a high shelf.

Perhaps the assistant was never told that you cannot comfortably lift overhead. Or that the shelf is not designed for those loads.

More polished writing will not fix missing information. The task needs better context and appropriate judgment.

Try asking:

Before you suggest a layout, ask about the space, the items, and any access or lifting limits that would change your recommendation.

For a student, relevant context might include the course level and assignment rules. For an event planner, it might include the budget, expected group, or available facilities.

Use a different check for each kind of result

Result	Useful check
Draft message	Compare facts, audience, tone, and promises with your source notes.
Explanation	Compare it with reliable learning material and try a new example yourself.
Extracted table	Check a sample of rows and every important value against the original.
Calculation	Reproduce it with an appropriate calculator or spreadsheet.
Prediction	Inspect data, assumptions, method, and performance on relevant examples.
External action	Check the actual record, recipient, or confirmation in the destination system.

The amount of checking should fit the consequences. A slightly awkward birthday greeting is easy to revise. A wrong number in a payment or a mistaken instruction about equipment deserves more careful handling.

You will learn detailed verification methods later. For now, make the check part of the request rather than an afterthought.

Sometimes a simpler tool is enough

If you need to add five numbers, a calculator may be all you need. If you need today's library hours, the library's official page or a direct call may answer the question. If a familiar checklist already works, you may not need to rebuild it with AI.

AI can still help you frame a question, organize information, or understand a result. You decide where that help earns its place.

An excellent outcome might be discovering that one part of a task benefits from an assistant while another part belongs in a spreadsheet.

Try it: choose the method before asking

For each task below, choose a method and a check:

Draft an invitation from supplied notes.
Find today's opening hours for a museum.
Add the costs on a shopping list.
Group ten fictional customer messages by topic.
Estimate next month's demand using historical records.
Explain a paragraph from a textbook.
Create an illustration for a fictional story.
Send an approved message to a specific person.
Summarize a policy that includes exceptions.
Create practice questions for a permitted study session.

A reasonable answer might pair the invitation with a writing assistant and a fact check, the opening hours with an official current source, and the total with a calculator. The sending task requires an authorized messaging tool and confirmation of the actual send. The policy requires checking the original exceptions.

More than one method can be appropriate. You succeed when you can explain why your choice fits the job and how you will know whether it worked.

The next time you open an assistant, start with those two questions. They will serve you through the most advanced parts of this series.